DTM 2019 (Diabetes Technology Meeting)

November 14-16, 2019; Rockville, MD; Full Report - Draft

Executive Highlights

  • DTM 2019 is now in the books, and we were nearly blown into the Atlantic Ocean by stellar sessions and commentary.

  • Highly anticipated data from the ten-month (!) Project Nightlight trial headlined in automated insulin delivery, showing an impressive Time in Range benefit of +2.6 hours/day with Control-IQ vs. sensor-augmented pump (SAP). The unique study (n=80) tested study participants across four eight-week phases: sensor-augmented pump, evening/night-only closed loop (twice), and 24/7 closed loop. Time-in-range was 70% during 24/7 closed loop phases compared to 59% during SAP phases; time below 70 mg/dl was ~2% during 24/7 closed loop compared to 4% with SAP. The study also showed significant glycemic and patient satisfaction improvements with the pump-embedded Control-IQ algorithm with Dexcom G6, compared to the research platform with the algorithm running on a smartphone with Dexcom G4. These data are a big win for Tandem and Dexcom, aligning nicely with the pivotal data (ADA 2019, NEJM).

  • We heard from six (!) FDA speakers on day #2, touching on the fast-moving regulatory environment. Dr. Courtney Lias called for the development of CGM-specific decision support tools (e.g., bolus calculators) and discussed needed research in the time-in-range movement. Ms. Naomi Schwartz discussed the FDA’s stance on cybersecurity issues (“we’d prefer all devices … have the ability to be updated.”). We also heard Mr. Bakul Patel’s update on PreCert and Dr. Yiduo Wu on using real-world evidence as part of a regulatory submission.

  • CGM and Time in Range dominated much of discussion, headlined by a powerhouse panel on expanding sensor uptake with Drs. Roy Beck (Jaeb Center), Bruce Buckingham (Stanford), David Price (Dexcom), Francine Kaufman (Senseonics), Andreas Stuhr (Ascensia), and Robert Vigersky (Medtronic). Dr. Lutz Heinemann gave a more critical look at Time in Range, asking whether its growing acceptance was being driven by “evidence or eminence?” Dr. Boris Kovatchev presented a promising new mathematical model linking Time in Range with A1c (correlations >0.9), while Dr. Richard Bergenstal added to the lexicon with his “Texas two-step” approach to CGM analysis (“slow, slow, quick, quick”): both (i) long-term, retrospective review; and (ii) short-term, corrective action.

  • Connected insulin pens and decision support were discussed throughout, headlined by Novo Nordisk announcing a 2Q20 launch start for its NovoPen 6 and Echo Plus pens. Presumably the launch will begin in Europe, where the devices already have CE Mark, before coming to the US. A reusable connected pen attachment for Novo’s disposable pens is now expected to launch in 2021, behind the initially anticipated launch and a welcome respite – the company should absolutely take its time to get this right (there appears to be more nuance to the disposable attachment,). Additionally, a roadmap slide for Novo showed three grayed-out “digital therapeutic” projects, something the company is working on “very actively.” Sansum’s Dr. David Kerr also gave an interesting and optimistic talk on this emerging category.

  • There were CGM updates aplenty, as Dexcom, Medtronic, and Senseonics all shared R&D efforts to improve CGM performance further – wow the field is ambitious and has come such a long way. Dexcom’s Mr. Peter Simpson presented on a modified G6 algorithm and early field data from a new G6 adhesive; there were unfortunately no new updates on G7. Medtronic’s Dr. Robert Mucic gave a look at the company’s adhesive selection process for its next-gen CGM (Synergy?), including some 14-day tests. Senseonics shared a look at its next sensor design (same size, more hardware), its plans for increasing Eversense wear time to 180 and then 365 days, and real-world data from 945 European Eversense users across four sensor cycles (TIR: 63%-65%, GMI: 7%). This was recently published in Diabetes Technology and Therapeutics – “Longitudinal Analysis of Real-World Performance of an Implantable Continuous Glucose Sensor Over Multiple Sensor Insertion and Removal Cycles.”

  • In decision support, a field where we expect to see increasing investment and further progress, OHSU’s Dr. Jessica Castle shared some details and screenshots for the Helmsley-funded “DailyDose” app for CGM+MDI users. The app will include a “smart” bolus calculator, recommendations for insulin:carb ratios, correction factor, and basal dose, hypoglycemia predictions, and more.

  • One Drop and Roche both shared information on their machine learning-based efforts to make glucose predictions. One Drop’s CGM-based 30-minute predictions were quite accurate, with 91% of predictions falling within 20 mg/dl of the actual value; as expected, this dropped to 63% for a 60-minute forecast and 41% for the very ambitious 120-minute forecast. One Drop ambitiously hopes to make the CGM predictions available to app users in mid-2020, using real-time CGM data (regulatory class TBD). Roche is also working with IBM to use large data sets – 70 variables! – to predict glucose with accuracy up to three hours out. It is still in the research phase.

  • Audience members were noticeably excited to hear ambitious updates from Thermalin Chief Innovation Officer, Dr. Michael Weiss, who divulged details on his company’s early stage “StampPump” and (still) pre-clinical ultra-concentrated, ultra-rapid insulins. Profil’s Dr. Tim Heise shared insights on the feasibility of oral insulin and potential barriers to once-weekly options. In addition, MannKind’s Dr. David Kendall weighed in on the benefits and challenges of inhalable Technosphere insulin.

Greetings from Rockville, MD, where the 2019 iteration Diabetes Technology Meeting took place. Below, you’ll find our top highlights from (i) Automated Insulin Delivery (AID); (ii) CGM and Time in Range; (iii) Regulatory; (iv) Smart Pen and Insulin; and (v) AI and Decision Support.

Table of Contents 

Top Automated Insulin Delivery (AID) Highlights

Project Nightlight Trial: +2.6 Hours/Day TIR with Control-IQ vs. SAP, Time <70 mg/dl Halved; 93% Time in Closed Loop with Embedded Control-IQ Algorithm vs. 87% with mobile inControl Algorithm

On the heels of the NEJM publication of Tandem’s Control-IQ pivotal trial, Dr. Sue Brown (University of Virginia) presented highly anticipated results from the unique, ten-month Project Nightlight study (n=80). Participants alternated across 8-week phases of: (i) sensor-augmented pump (SAP); (ii) evening-night-only closed loop (start between dinner and sleep, stop at wake-up); and (iii) 24/7 closed loop. There were two-week washout periods in between. A1c and psychometrics were performed at baseline and during each wash-out period. Of the 80 participants, 35 used TypeZero’s inControl running on a smartphone with Dexcom G4 and Roche Spirit Combo Pump. In 2018, that was replaced with the Control-IQ algorithm embedded in Tandem’s t:slim X2 pump and Dexcom G6 (i.e., the commercial product); 41 participants used Control-IQ and the remaining four used both systems. Participants’ mean age was 42 (range: 18-69 years), with A1c of 7.4% at baseline. 80% of participants were CGM users prior to beginning the study. Time in Range improved from 59% with sensor-augmented pump (SAP) to ~68%-70% with closed loop; time <70 mg/dl was halved from 4% to ~2%. These outcomes are almost identical to the Control-IQ pivotal, where time-in-range improved from 59% to 71% – nicely confirmatory data from another long term study of the under-FDA-review product.




Evening-Night CL

24/7 CL





Time <70 mg/dl




Time >180 mg/dl




  • Time-in-range during 24/7 closed loop phases was 70% vs. 59% in SAP phases (p<0.001) – an 11% difference or 2.6 hours/day. With the vast majority of time-in-range improvement coming at night, time-in-range was only slightly higher in the 24/7 closed loop phases compared to the evening-night closed loop phases. (Time <54 or >250 was not reported, though presumably will be when this is published.)

  • Time below 70 mg/dl was cut in half during the 24/7 closed loop phases (1.8%) compared to SAP phases (4%; p<0.001). Evening-night closed loop showed a similar, but slightly lower, improvement in time <70 mg/dl.

  • A1c was around 7%-7.1% for both evening-night closed loop and 24/7 closed loop phases. Depending on order of study phases, one group recorded a mean A1c of 7.6% at baseline and 7.3% during SAP, while the other group recorded a mean A1c of 7.2% at baseline and 7.1% during SAP. In other words, A1c appears to have improved between 0.1%-0.3%, ending up around 7% with closed loop – similar to the Control-IQ pivotal, and not surprising considering the quite-low starting A1c at baseline. It’s unclear whether these differences were statistically significant.

  • Greater improvements in time-in-range were seen for the Control-IQ subgroup (+13%; 3.1 hours/day) on the embedded Tandem/Dexcom G6 system compared to the mobile inControl subgroup (+7.8%; 1.9 hours/day), p<0.01. This was an exploratory analysis. Similarly, time >180 mg/dl was reduced significantly more in the Control-IQ group (-10.9%) vs. the inControl group (-5.4%), p<0.01. Lastly, there were no significant differences in time <70 mg/dl improvement.

  • Closed loop use for the Control-IQ (t:slim X2) subgroup was strikingly high at 93%, compared to 87% for the inControl (smartphone) subgroup. Median time in closed loop with Control-IQ was consistently high across all participants, ranging from 86% to 98%. Time in closed loop ranged from 59% to 97% for the inControl participants – far more variable. It’s not surprising that participants spent more time in closed loop with the embedded Control-IQ algorithm – the G6/t:slim X2 commercial product is more reliable than the precursor research platform (smartphone with inControl algorithm, Roche pump, Dexcom G4/G5). It will be interesting to compare outcomes once more smartphone-focused systems come out – e.g., algorithm on smartphone vs. algorithm embedded in pump. Both approaches have pros and cons.


Time <70 mg/dl


Time >180 mg/dl

inControl group, SAP




inControl group, 24/7




inControl difference

-2.4%; -35 min/day

+8%; +112 min/day

-6%; -78 min/day

Control-IQ group, SAP




Control-IQ group, 24/7




Control-IQ difference

-2%; -29 min/day

+13%; +186 min/day

-10%; -157 min/day





  • Preliminary data from a technology acceptance survey showed strong patient preference for the embedded Control-IQ system. “Ease of use” was rated 4.6/5 for the embedded system vs. 3.8/5 for the mobile-based inControl system. The embedded system was rated 4.4 for “usefulness” and 4.1 for “trust,” compared to 3.9 for both “usefulness” and “trust” for the mobile-based system. Lastly, based on structured interviews with participants, Control-IQ users rated probability of long-term use 5/5 (“Extremely Likely”), compared to 3.3 for the inControl system users. Structured interviews also revealed that participants favored 24/7 use of closed loop over evening-night-only use. We believe this trial was envisioned back when overnight-only closed-loop was more seriously considered as a first commercial product; the field has since settled on 24/7 closed loop as the default.

    • The technology acceptance data mirror the results from Control-IQ’s pivotal trial (ADA 2019). Users rated Control-IQ ease of use 4.7/5, usefulness 4.6/5, trust 4.5/5, and desire to continue using was an impressive 4.8/5.

  • In total, only one serious adverse glycemic event was recorded during closed loop use: one participant experienced a severe hypoglycemic event (requiring third-party assistance) “following physical activity and delay of hypo treatment.” (This is a good reminder for the field – AID makes insulin therapy safer, but not zero-risk. Patients may also feel safer generally, which could make them less likely to move quickly on problems that deserve fast fixes, like severe hypo brought on by exercise.) Four severe hypoglycemic events were reported during open-loop use. Control-IQ was suspended for up to four weeks in eight participants in March (Tandem patched a bug through the Device Updater), though there was no impact on outcomes.

Dr. Boris Kovatchev Compares iDCL1 and iDCL3: Form Factor and No-Calibrations Translate to Improved Glycemic Outcomes; Calls for Looser Inclusion Criteria in Closed Loop Studies

In an overview of recent closed loop studies, Dr. Boris Kovatchev (University of Virginia) compared results from iDCL protocol 1 (smartphone-embedded inControl algorithm, Roche Spirit Combo pump, Dexcom G4 or G5) and iDCL protocol 3 (Control-IQ algorithm embedded in Tandem t:slim X2, Dexcom G6; i.e., the pivotal trial we saw at ADA). As stated by Dr. Kovatchev, direct comparisons between the two trials have validity, because both studies’ control groups received the same treatment (sensor-augmented pump). Control-IQ/G6 delivered an 11% absolute increase in Time in Range compared to SAP vs. a ~5% improvement seen with inControl running on the phone. Similarly, time >180 mg/dl was reduced by 10% (2.4 hours) with Control-IQ vs. 3% (~0.7 hours) with inControl. Differences in time below 70 mg/dl were not significantly different. Given that Control-IQ and inControl are the same algorithm, the results suggest embedding the control algorithm into the pump drives higher time in closed loop (i.e., less communication failure points, less stuff to carry) and using the no-cal Dexcom G6 produced significant improvements in glycemic outcomes.


inControl (n=125)

Mobile System

Control-IQ (n=168)

Embedded in Pump


Time <70 mg/dl




Time in Range




Time >180 mg/dl




Mean glucose

-2 mg/dl

-13 mg/dl


  • Dr. Kovatchev also called for looser inclusion criteria for closed loop studies, particularly when using time-in-range as an outcome. Dr. Kovatchev referred to the iDCL 3 (Control-IQ pivotal) trial, which had no entry restrictions on A1c and showed Control-IQ delivered improved time-in-range across all baseline A1cs. Dr. Kovatchev noted how Time in Range varied in the intervention group by baseline A1c from 85% for baseline A1c ≤6.5% to 60% for baseline A1c ≥8.1%. By setting tight inclusion criteria (e.g., baseline A1c ≤8%), a closed-loop study can show very high time-in-range. He recommended that all studies have a chart like the one below, to demonstrate how baseline A1c affected Time in Range outcomes. We love the idea that AID can be a great equalizer – helping people improve Time in Range regardless of the starting baseline.

  • A feasibility study for Control-IQ in 2-6 years olds has completed. The study was conducted in 15 participants at UVA, Barbara Davis, and Stanford with a 48-hour supervised hotel stay and training with 72-hours of home use. The French FreeLife Kid AP study with 120 participants ages 6-13 reported partial results at ADA 2019. Lastly, a pivotal for children ages 6-14 for Control-IQ has finished recruiting (ClinicalTrials.gov) and is on track for FDA submission in 1Q20.

Thermalin’s Dr. Michael Weiss Details Early-Stage Pre-Filled, Shelf-Stable, Patch Pump, “StampPump,”; Positive Pre-Clinical Ultra-Rapid U-500 Insulin Results

Audience members were undeniably intrigued by next-generation insulin company Thermalin, as session Q&A was dominated by questions directed at Founder and Chief Innovation Officer Dr. Michael Weiss following his presentation. Dr. Weiss outlined the company’s current paths of innovation: (i) ultra-stable and ultra-concentrated (U-500 to U-1000) rapid-acting insulin analogs for miniaturized pumps; (ii) ultra-heat-stable basal analogs for global distribution in the developing world; (iii) ultra-potent insulin analogs to reduce the cost of insulin products; and (iv) “smart” insulin analogs as safer glucose-responsive hormones. On the first front, Dr. Weiss introduced the audience to his company’s early-stage “StampPump” – a proposed pre-filled, shelf-stable, stamp-sized patch pump for closed loop usage (check out the prototype photos below). The pump’s tiny size would obviously require the highly concentrated U500 or U1000 insulin. Presumably this would also require ultra accurate delivery, since errors of even 5% could be highly dangerous. It’s currently unclear how far along in development the pump is, since minimal updates on timeline were shared, however, an NIH proof of principle grant is set to complete in March 2020, and an Indiana University grant was awarded in June 2019. Of note, Thermalin’s collection of insulins have remained at the pre-clinical stage since the company’s start in 2010. To the best of our knowledge, no formal timeline has been shared, and we’re curious to see when a candidate will progress into human trials.

  • As a stand-alone product and for the purpose of the StampPump, Thermalin is currently developing both novel single-chain insulins (SCIs) and more traditional two-chain insulin structures. While the SCIs have shown striking thermo-stability in pre-clinical studies for up to one year, Dr. Weiss noted that the company has not yet been able to achieve ultra-rapid action with subcutaneous administration. For two-chain insulin, the challenge is more so creating a stable, concentrated product without hexamers and minimized dimer formation. So far, positive euglycemic clamp data in Yucatan swine has demonstrated similar ultra-rapid action profiles between Thermalin’s two-chain U-500 insulin and Fiasp.

  • Dr. Weiss shared that his broader vision for the company is to create insulins that can be used in developing countries that struggle with limited storage capabilities – an aspiration that developed during his rotations in East Africa as a medical student. Dr. Weiss also divulged that in a meeting with Bill Gates, he was told that the most feasible way to achieve his goal was to build a business model in “affluent societies,” whose product could then be utilized in the developing world. Once approved by the FDA, Dr. Weiss hopes that the Gates foundation will provide Thermalin’s technologies to the global community in the same manner as AIDS/HIV drugs.

Q: Have you given up on single chain insulin? It seems to me like there are still a lot of options. Or is there something about SCI that prevents rapid action?

A: We’ve achieved rapid action, Humalog or Novolog levels with subcutaneous SCI, but not ultra-rapid action like Fiasp. We’re excited for intraperitoneal pumps or basal or biphasic SCIs, however.

Mr. Gary Scheiner (Integrated Diabetes Services): We know that preservatives and other things in insulin interfere with micropump function. Will your insulin require the same preservatives of current insulin?

A: That’s a wonderful question. There was a serendipitous discovery at Lilly that certain ligands enhance its physical and chemical activity. We don’t need that. We’re open to a much wider chemical space of preservatives, as the pre-filled StampPumps would not have to operate for more than a week. Long-term concerns might apply to a peritoneal permanently implanted pump.

Q: Why was speed achieved with double-chain insulin but not SCI? 

A: I don’t know, but I have a guess. Based on unpublished X-ray crystallography data, we see that SCI opens up, and there’s a domain of the receptor that threads through the hole – like thread in a needle. In two-chain, there is no hole to thread. I think the threading is keeping a longer profile.

Q: So, it’s more of a chemical rather than physiological effect?

A: That’s what I think.

Q: I have daughter with type 1, and I’ve worked with type 1 diabetes. They’re on pump, and I sometimes find it very difficult to manage post-prandial insulin with any of the rapid-acting insulins, especially with mixed foods like pizza. Are there any dual-action insulins planned in the future? One part faster, the other part later for patients not on a pump.

A: That’s a great question, and one of the mysteriously favorable properties of SCI is that we can tune the ratio of immediate action and delayed action. We don’t know what the molecular rules are and if the tunable quality that is unique to SCI is unique to biphasic action.

Q: You have a wonderful U-500 insulin, and you implied that it can specially be used by young adults and maybe children. That made me wonder, do you have ideas on smaller insulin needs? The tiny amounts that need to be delivered might be a challenge for mechanistic pumps.

A: Over the last decade, there have been so many advances in microfluidics – more advanced than diabetes formulation technology. I want to emphasize that there’s an important subsegment in type 2 patient, which Dr. Lane focused on, that require more than 200 units a day. They’re now using Lilly U-500 in pumps. You can see how there’s much more benefit in ultra-rapid acting U-500 in a pump, even a conventional pump.

Top CGM and Time in Range Highlights

Quotable Quotes from a Powerhouse Panel on Increasing CGM Use: Drs. Roy Beck, Bruce Buckingham, Francine Kaufman, David Price, Andreas Stuhr, Robert Vigersky

On CGM in Type 2s

  • “I think the use of CGM in type 2 is probably the greatest unmet need there is right now. A patient who was in [a study] came to see me about 2 years later, and I had never seen him personally. He said, ‘Dr. Vigersky, I was in your trial.’ I asked, ‘How’d it go?’ He said, ‘This [CGM] was an amazing device. I’m Italian, and I love bread. I saw what bread did to my blood sugar, so I cut out bread. I lost 30 pounds, my A1c went down from 9% to 7%, and I’ve stayed there ever since.’ To me, that was an example of the power of what this technology can do. I think it has that power in type 1, but to me, I think we can do an awful lot of good for our type 2 population.” – Dr. Vigersky

  • “At the Jaeb Center, we had the DIAMOND study which was not only a type 1 study, but also a parallel type 2 diabetes study with adults on MDI. Going into it, I didn't really have any idea whether we were going to show benefit or not with this group or whether they’d use CGM or not. It was a pretty low-touch protocol. There were a pretty minimal number of visits between 3-6 months. Like [Dr. Vigersky] said, there was not only a substantial reduction in A1c compared to the control arm, but patient satisfaction was incredibly high. The median amount of CGM use at six months was about 90%. Patients really adhered to it and demonstrated high satisfaction and benefit. I think we’ve just kind of touched the surface in type 2.” – Dr. Beck

  • “We don’t know the exact frequency or how long to do intermittent monitoring. It may be just a couple of weeks every 3 or 4 months, and you might get the desired effect. I call on someone in the audience to do this study so that we can get what the frequency of this is and the cost-effectiveness.” – Dr. Vigersky

On Expanding CGM Uptake

  • “It’s not any longer about proving that the technology works, but about how we can get more people to adopt this technology we know that works. That’s why I’m so excited about it. There are probably many roadblocks, but the one I like to focus on is the value people see this technology provides  … We did market research and asked people before and after what they expected to achieve when adopting this technology. Most people originally said better A1c and better time-in-range. After half a year, half of the people said their expectations were not met – was this because the technology didn’t work, or maybe because they don’t know how to use it in the most effective way. I believe that education on how to use it and what to do with the information would be one area to address. Clearly, access is a major roadblock, but even in countries where there is access and limitations are not a problem, there is still not a majority of people with type 1 using it...We’ve heard in the past about the annoyance of false alarms or being overwhelmed by the data, but people who have managed diabetes without CGM need to unlearn things and need to learn new things in order to get the benefit out of CGM.” – Dr. Stuhr

  • “Earlier generation CGMs expectations’ were not at the right level. Providers didn’t give [the right expectations] to patients, and there were false expectations. In general, now, a big issue is awareness among adults. In this country, the majority of adults with type 1 diabetes are being cared for in primary care settings and not endocrinologists. Either awareness of CGM is not there in primary care, or they're so busy with so many things that they don’t want to deal with it. It would perhaps be a surprise to this audience, but there are a lot of adults who may have heard of CGM but don’t know anything about it. We did a feasibility study of a direct-to-patient approach to starting CGM. In Wisconsin, we partnered with a primary care research network throughout the state that integrated primary care practices. We decided that we would try to identify adults with type 1 diabetes who had not seen an endocrinologist and see if they wanted to try CGM and do it all remotely. We sent out a letter to patients, and beyond that, we had a website they could go to with information on CGM. If they were interested, they elected to consent and gave them information. We used the Abbott FreeStyle Libre and Dexcom G6, and they select which one they wanted. We sent it to them, had a CDE do a virtual visit with them to give them some training, and then they were on their own for a bit. There was some other contact, after they used it for a little, they got some advice on how to use it in their day-to-day management. We started with 34, mostly type 1, but there were about seven type 2s. It was remarkable. Just putting it on in the first week, time-in-range went up by 2.5 hours, and that’s really without instructions about how to use it ... At the end of three months, all 34 were still using it, and actually, CGM use was about 95% over the whole time. A1c dropped from where they started at 8.3% to 7.2%. We didn’t have a control group, and this was a group naive to CGM and didn’t know much about it from the start, but it just shows the simplicity of the devices now and the lack of burden that used to be there. We’re about to embark on a much larger national study to do this.” – Dr. Beck; we might see the full data on this Jaeb/Cecilia Health pilot study (“Geek Squad”) at ATTD 2020

  • “I think that the Abbott Libre direct-to-patient advertising has really helped all the CGM companies in terms of awareness. I know people with diabetes, and they’ve asked me about what I think. They would have never heard about it before, so I actually think that what Abbott did going down that road as far as awareness of CGM was a public service. I think though that there is more to go for awareness.” – Dr. Beck

  • “[Dexcom is] doing some [direct-to-consumer advertising], and as we are able to ramp our production, we’ll do a lot more. Another real barrier is access. People who are poor can’t access CGM, and that’s really a problem. To overcome access issues, people need to meet with payers and present clinical and health economics. There needs to be work with advocacy because we need to be able to get it and particularly poor people.” – Dr. Price

On Remote Monitoring

  • If Rayhan Lal is here, he had a new-onset on CGM for a few months and got a call at ~1 AM. The mother said the CGM stopped working, and that she didn’t know what [her son’s] glucose was. She asked, “what should I do?” He suggested, ‘take a fingerstick,’ and she said, ‘How barbaric!’ It’s totally changed the way that people manage their diabetes.” – Dr. Buckingham; FFL 2019 was the first time we heard this story!

  • “We asked who people wanted to share the data with, and the pediatric population was pretty clear. [Dr. Buckingham] is correct about the adolescents as there’s not much accounting for what an adolescent may or may not want to do. However, the adults really surprised me. They didn’t want to share, for the most part, with anybody other than either their HCP, which I thought was quite generous, or an emergency medical responder. The person they didn’t want to share it with the most was their spouse.” – Dr. Kaufman

  •  “We're putting people recently diagnosed on CGM from the beginning, and people are just hooked on it. It’s how they get sleep at night, having alarms when something goes low. It’s not been a problem. With adolescents, they sometimes need a little bit of space, so when we start them up, we always ask for terms of engagement like when someone wants to be notified and allowing them to set the limits.” – Dr. Buckingham

On CGMs in Inpatient Care

  • “We need the right kind of clinical trial data to show that it performs to whatever claim it is that we are trying to make, whether it’s in the ICU or general wards, so it’s going to be a little while. Hopefully, it will happen soon because the need is there.” – Dr. Price; see a review of Dexcom’s hospital studies here

  • Wouldn’t it be nice if the patient coming into the hospital with a CGM could keep it? I think one of the bigger problems is right in the practice guidelines, getting those with diabetes who we know need CGM are allowed to continue it and maybe extend that to closed loop. I think we need to get there first before we can start putting it on people who are not using it.” – Dr. Kaufman; read Adam’s experience Looping and using CGM in the hospital here

  • “This is a little bit cynical, but I think that it’ll be ubiquitous when hospital administrators recognize it improves outcomes and reduces length of stay … We have to show that this device would at least be offsetting that additional cost, but hopefully, reducing overall costs.” – Dr. Vigersky

Dexcom G6 New Adhesive Patch and Modified Algorithm Launched; No New G7 Updates (late 2020 launch)

Dexcom’s Mr. Peter Simpson kicked off DTM 2019 with a few updates on G6:Mr. Simpson shared some initial field data (n=~2,500) showing a 43% relative improvement in 10-day patch survival rates using a new patch. At Keystone, Dexcom EVP Andy Balo shared that Dexcom had filed a new adhesive, called “MA19,” with the FDA. Based on the graph shown, the new patch, which has apparently already launched, shows a 10-day survival rate of ~95%, compared to ~90% with the original patch.

  • A modified G6 algorithm has shown a 24% relative improvement in data availability (i.e., reduced sensor errors where data is not displayed). Preliminary analysis from 310 subjects showed data was available 98.3% of the time with the original algorithm and 98.7% of the time with the modified algorithm. Data availability on the 10th day of wear was improved from 96.6% to 98.5% (a 56% relative reduction in data unavailable errors). Based on Mr. Simpson’s comments, this modified algorithm has also been launched – not acknowledged on any recent calls (Dexcom is getting quite stealth about product improvement!). The pictures below demonstrate the algorithm’s improvements: periods where data is unavailable are highlighted in blue.

  • An unpublished study (n=78 type 1s) showed today suggested that G6 has strong accuracy performance with one-hour warmup and out to 14 days. G6 currently requires a two-hour warmup and automatically shuts off after 10 days. Between hours 1-2 of G6 wear (n=162 matched pairs), overall MARD was an impressive 9.9% and 93.2% of values were within 20%/±20 mg/dl of reference (likely YSI, but not specified). At day 10, MARD was 9.2% with 93.7% within 20%/±20 mg/dl. Accuracy was not significantly diminished at day 14, which showed a MARD of 9.2% and 91.9% within 20%/±20 mg/dl. For reference, Dexcom advertises a MARD of 9.0% in its G6 User Guide. We do not believe Dexcom plans to go for 14-day wear for G6 in advance of G7.

    • G7, however, is planned to have longer 14-15 (or even 16?) day wear and faster warm-up. The results of this unpublished study suggest achieving these goals shouldn’t be a huge issue, even using the same chemistry and algorithm as the G6. G7 will obviously need to meet iCGM benchmarks out to 14 days, and the adhesive will reliably need to reach that length of wear. Many Dexcom G5 users wear their sensors for longer than 10 days by restarting the sensor; the G6 has a mandatory shutoff as part of iCGM special controls.

  • Mr. Simpson reiterated the planned focus for Dexcom G7 and the targeted launch of “late 2020.” He outlined three key areas for G7: (i) making the device more convenient to wear (i.e., smaller); (ii) simplicity and ease of use (“just a few steps from opening package to getting CGM data”); and (iii) reducing cost. Last week, we heard Dexcom and Verily were making “final tweaks” to the design. The slide below is taken from previous Dexcom presentations; we should hopefully see an updated look at the hardware at JPM in January.

Dr. Lutz Heinemann’s Critical Appraisal of TIR: Is Its Growing Acceptance Being Driven by “Evidence or Eminence?”; TIR is a Meaningful Supplement to A1c, but Not a Replacement in Clinical Practice

Dr. Lutz Heinemann (Science and Co.) presented a critical look at time-in-range (TIR), arguing that while TIR is a valuable addition to A1c in clinical care settings, it is not a replacement for it. Noting the quick, ~six-month turnaround for consensus guidelines around CGM metrics to be developed and then incorporated in ADA Standards of Care, Dr. Heinemann asked the rhetorical question, “Was this driven by evidence or eminence [of the people on the consensus committee]?” Presenting an argument outlined in Dr. Heinemann’s recent publication in JDST, Dr. Heinemann emphasized the lack of a reference standard or quality control for TIR. As TIR is a parameter calculated from data from CGMs, biases in CGM values can significantly affect TIR values. Underscoring this point, Dr. Heinemann showed data presented on a poster at EASD 2019, finding that FreeStyle Libre reported twice as much time below 70 mg/dl as the Dexcom G5 when worn on the same 24 subjects. Additionally, Dr. Heinemann pointed at the relatively weak evidence linking TIR with long-term complications and expressed skepticism that a large-scale study like DCCT could ever be done again. Still, he conceded that smaller, less-convincing studies have already been done linking TIR with long-term complications (e.g., Beck et al., Diabetes Care 2018).

  • Dr. Heinemann admitted that time-in-range was still an incredibly valuable metric, particularly for use as an education tool. As TIR is much more reflective of day-to-day glucose variations, it is a very powerful tool to help patients visualize and understand their glucose patterns and decision making. Additionally, Dr. Heinemann noted several physiological factors that can affect A1c, which TIR is not affected by. For this reason, he highlighted a particularly strong use case for TIR in pregnant women.


Table taken from Dr. Heinemann’s publication in JDST.

Dr. Boris Kovatchev: >0.9 Correlation Between A1c and TIR Using a Differential Equation Model

University of Virginia’s Dr. Boris Kovatchev presented a new differential equation model correlating A1c with time-in-range (TIR) that could accurately estimate A1c with a correlation coefficient of 0.93 between estimated A1c and actual lab A1c. Notably, the equation estimates A1c daily, using TIR derived from CGM daily profile – very cool. The model is based on the premise that both TIR and A1c reflect the same underlying phenomenon (i.e., blood glucose), but that interpersonal variability in glycation rates drive the discrepancy seen between two people with the same TIR and different A1cs. The differential equation he proposes contains three fixed population-based parameters (i.e., biological constants for all humans) and a fourth variable, representing individual glycation rates, that is different for each person. Using the dataset from iDCL 1 (mobile-enabled inControl, Dexcom G4/G5, Roche Spirit Combo; n=120), the three fixed population constants were calculated. Then, individuals’ A1c values from month three of iDCL 3 (Control-IQ embedded in t:slim X2, Dexcom G6; n=168) were used to calibrate the individual glycation rate variable for each study participant. The equation was used to estimate participants’ A1c daily and then compared to reference lab A1c at months 6 and 9 of iDCL 3 with impressive accuracy.

  • Estimated A1c and lab A1c values were strongly correlated (r=0.93) 3-months and 6-months after “calibration” to individuals’ glycation rates. After 3 months, 98% of estimated A1c readings were within 10% of the lab measured A1c; after 6 months, this percentage deteriorated only slightly, to 96%. For context, the traditionally used linear models correlating TIR with A1c from Drs. Roy Beck (2019) and Robert Vigersky (2019) have correlation coefficients around 0.6-0.7. Of course, Dr. Kovatchev’s differential model has a trade-off in that it requires one individualized calibration point.

    • The picture below shows the difference between Dr. Kovatchev’s model and a linear model. The blue dots represent estimated A1c and lab A1c scatter using a linear correlation, while the red dots represent the differential model. Note that the red dots are bunched much more closely around the perfectly correlated eA1c = lab A1c line.

In the differential equation above, the three population constants are represented by tau, a, and b. The individualized glycation rate parameter is represented by gamma. TIR and eA1c represent daily time-in-range and estimated A1c, respectively.

Medtronic Shows 14-Day Patch Survival Study for CGM “Prototype”; Could Synergy Have 10 or 14-Day Wear-Time?

Increasing wear-time was a major theme throughout Thursday morning’s CGM session, and Medtronic’s Dr. Robert Mucic shared data from the patch design selection process for Medtronic’s next-gen CGM with studies out to 14-days. The studies were done with multiple adhesives and four different adhesive patch sizes, using a “representative payload” with the same size and profile as the planned CGM. Medtronic gathered user input during an internal study on level of skin discomfort, adhesive “edge lift,” appearance acceptability, and “device awareness during wear,” on eight combinations of adhesive and patch size. While skin discomfort, appearance acceptability, and device awareness during wear were all similar, there was significant differentiation in edge lift. Following patch removal, there were also significant differences in the level of adhesive residue and clean-up required. This helped Medtronic choose three materials for an external study with 28 people in a “Northern State” (mid- to high- 70s temperature, upper 80s humidity) and 31 people in a “Southern State” (high-80s to low-90s temperature, low 90s humidity). The hotter, more-humid weather showed one material clearly performed worse than the others over the 14-day wear period. Finally, a larger, multi-site study (n=310) helped differentiate the final two chosen adhesives; the 14-day survival rate for the best-performing adhesive appeared to be around 85%. Great to see Medtronic thinking so deliberately about adhesive!

  • Notably, the picture of the “one-piece sensor and transmitter prototype” was identical to the picture we’ve seen for Medtronic’s fully disposable, 7-day, Synergy iCGM. Dr. Mucic expanded a bit on the form factor of the device, emphasizing the radial symmetry as making the device easier to insert (users won’t have to worry about orientation). Given the studies Dr. Mucic described tested adhesive patches out to 14-days, could Medtronic potentially launch Synergy with a longer wear time of 10-days or 14-days? Synergy is targeted for FDA submission in November 2020-April 2021 and launch by ADA 2021.

“About Half” of Eversense Sensors Removed at 180-Days Would Support 365-Days of Performance; Real-World Data from Europe Shows ~64% TIR, ~12% MARD vs. BGM Reference

Senseonics’ Mr. Alex Ghesquiere shared a quick look ahead at Senseonics’ plans for the implantable Eversense CGM, noting that “current R&D clinical tests” are already ongoing with Eversense up to 365-day wear. Additionally, he shared that “~50%” of current Eversense XL sensors that are removed at 180-days could actually support 365-days of wear. Currently, Eversense is approved for 180-days in Europe and 90-days in the US; enrollment for the 180-day US PROMISE trial just completed, with study completion expected in 1Q20 and launch of the 180-day sensor by “the end of 2020.” Extending sensor wear out to a full-year will require some chemistry and hardware modifications, primarily to prevent oxidation from degrading its glucose indicator monomer, TFM. Currently, the Eversense sensor has a dexamethasone coating, an anti-inflammatory steroid, to reduce oxidative species around the sensor. Mr. Ghesquiere shared that Senseonics is also considering providing alternate oxidation sites, among other methods, to further limit degradation of the TFM. Lastly, Mr. Ghesquiere presented a new sensor design, which will begin “clinical testing” starting this month. The new sensor will maintain the same size, but pack significantly more inside. The new design includes two integrated chips with four excitation LEDs each, compared to a single excitation LED in the current sensor. The improved sensor will include two hydrogels with indicator chemistry (rather than one). According to Mr. Ghesquiere, the new sensor design will enable a reduction in calibrations (Eversense currently requires two fingersticks/day), longer sensor life, and improved system accuracy.

  • Mr. Ghesquiere also presented data from the recently published real-world data from the first 945 European users to complete 4 cycles of sensor wear. The sensor MARD was between 11.5% and 12% for the four sensor cycles, compared to BGM calibration (n=705,462 paired points!). Across the four cycles, user time-in-range was between 63% and 65% with ~5% time spent below 70 mg/dl. Mean GMI was right around 7% and median transmitter wear time increased slightly from 83% in the first sensor cycle to 86% in the fourth sensor cycle.


Sensor Cycle 1

Sensor Cycle 2

Sensor Cycle 3

Sensor Cycle 4

Mean sensor glucose

157 mg/dl

158 mg/dl

157 mg/dl

157 mg/dl











Time <54 mg/dl





Time <70 mg/dl










Time >180 mg/dl





Time >250 mg/dl





Median wear time





Dr. Richard Bergenstal Shows Off His “Texas Two Step” of CGM: Quick, Corrective Actions + Slow, Retrospective Review

The engaging Dr. Richard Bergenstal (International Diabetes Center) added another catchy aphorism to the time-in-range lexicon, dubbing the need for both (i) long-term, retrospective review of CGM data and (ii) short-term, corrective actions the “Texas Two Step” of diabetes management (i.e., “slow, slow, quick quick”). Self-admittedly, Dr. Bergenstal’s new phrase is a rework of his “slow” and “fast” CGM thinking concept, based off of Nobel Laureate economist Dr. Daniel Kahneman’s book Thinking Fast and Slow, shared at ENDO 2018 and ATTD 2019. The “quick step” of the dance refers to making in-the-moment therapeutic or behavioral changes based on real-time CGM numbers or trends – for example, adjusting postprandial insulin based on a CGM trend arrow. For the “slow step,” Dr. Bergenstal used deliberative analysis of an AGP as an example. Using AGP, clinicians can work with patients to look for patterns of hypoglycemia or hyperglycemia, in conjunction with principles from the Nine Steps to Interpreting an AGP. Overall, Dr. Bergenstal called for greater use and education surrounding both “slow” and “quick” data.

  • Dr. Bergenstal also brought back two of his most popular AGP acronyms: “MGLR” and “FNIR.” Patients should strive for an AGP that is “Flat Narrow and In-Range” and a TIR chart that has “More Green, Less Red.”

  • GMI (Glucose Management Indicator), one of the ten standardized CGM metrics for clinical care, was highlighted as a measure that is becoming “more and more important.” As a reminder, GMI is the rebranded “estimated A1c,” converting mean glucose from CGM or fingerstick readings to an A1c-like %. For clinicians, any mismatch between GMI and lab A1c can provide insights into how a patient is doing more recently – e.g., Dr. Anne Peters gave an example at her talk on moving beyond A1c at CMHC 2019, stating, “I use GMI to say to patients, ‘Your measured A1c is 7.6%, but your GMI is 6.8%, so it shows that your last two weeks were better and you’re trending in the right direction.’” A difference between GMI and A1c would also individual variation in glycation – i.e., a lab-measured A1c that is over- or under-reporting actual measured glucose. 

  • Dr. Bergenstal countered the critique he often hears that ‘patients will never [look at an AGP] over and over again’ or use trend arrows consistently with an interesting NYT article entitled “The Unexpected Joy of Repeat Experiences.” We loved this read! As behavioral psychologist Dr. Ed O’Brien states in the piece, “when an experience has many layers of information to unveil, it’s probably a good bet to repeat it” – we definitely think AGP qualifies here!  

Selected Questions and Answers

Q: On using trend arrows to adjustment insulin doses, I know a lot of patients with type 2 diabetes are now using CGM. Can people with type 2 diabetes use the recommended insulin adjustments? Or, are the amounts just recommended for type 1s?

A: I think the first honest answer is that we’re not sure. We need a ton more CGM data in people with type 2, which is just starting to be collected now. We need to run those analyses. My guess is that the current tables that are published are actually going to be pretty close, but to be honest, we need a lot more data in type 2.

Q: You’ve looked at time <70 mg/dl, but how would you define a serious hypoglycemic event?

A: We’re realizing that while we’re doing studies now, we need to collect a series of hypoglycemic datapoints, so we can figure out what best correlates with things like emergency room visits, etc. Not only do we need to study the time <54 mg/dl, but further than that – is it the duration of time in hypoglycemia? Does it have to be a prolonged period? Do you need to be <54 mg/dL for, say, two hours to have a seizure? We’re collecting all those metrics: two hours <54 mg/dL, 30 minutes <54 mg/dL, two hours <70 mg/dL, 30 minutes <70 mg/dL.

Addressing the “Diet Challenge” with Machine Learning: Early-Stage “DietHabit” Project Attempts to Predict Meals Eaten Based on CGM Data

PATHS-UP’s Dr. Ashutosh Sabharwal (Rice University) presented on fascinating new AI-based models to tackle the “diet challenge” – the difficult task demanded of patients with diabetes to estimate the nutritional content of food (i.e., estimating carbs, fats, and proteins) and accurately dose for it. As Dr. Sabharwal explained, the challenge is so difficult because people have to replace how they once thought about food (simple labels like “pizza”) with an entirely new way (“how many carbs does this pizza have? How much protein or fat?”). Currently, Dr. Sabharwal’s own lab is attempting to create a model, entitled “DietHabit,” that predicts what exact food has been eaten based off of CGM data. Quickly realizing this was too ambitious, the group decided to look integrate real-world food log data and generate predictions from a personalized “food bank.” As we are creatures of habit, Dr. Sabharwal’s early work suggests that people have a limited set of meals that they commonly eat. They developed a clustering model to unify inconsistent food diary wording (“pasta with chicken” grouped with “chicken pasta”) and found that most participants had fairly reliable breakfast meals. However, lunch and dinner consistency tended to depend on the person while snacks were inconsistent for almost everyone. We think this would pair really nicely with an app like UndermyFork, which does a good job of tagging meals based on photos and pairing it with CGM traces.

  • Further along, Dr. Ricardo Gutierrez-Osuna’s team at Texas A&M University recently published a pilot study in IEEE, aimed at predicting macronutrient content if given a CGM trace, creating what Dr. Sabharwal refers to as a  “Continuous Diet Monitor” using data from CGM. In the study, 15 patients with a “healthy BMI,” between 60-85 years old were provided meals with known carbohydrate, fat, and protein content after eight hours of fasting. CGM traces for nine meals with varying macronutrient ratios were measured for each patient. Although the general glucose response for carbs (larger and longer) and fats/proteins (dulled amplitude, extended duration of recovery to baseline) could be easily predicted, subject-to-subject variability was extremely high, reducing the generalizability of results.

    • More recently, the researchers then developed a machine learning model to better counteract variability. Using input from fasting glucose levels, subject body weight, and “important features regarding the glucose response shape and duration,” the neural network was able to predict carbs, fats, and protein content with correlation coefficients of 0.76, 0.57, and 0.37, respectively. Estimating macronutrient amounts in general categories (low, medium, or high) further improved accuracy of 95%, 65%, and 65%, respectively. Results from that study are expected to be published soon.

  • Moving forward, Dr. Sabharwal has a new partnership with Sansum Diabetes Research Institute in the works. The collaboration will focus on integrating CGM readings, food photography, and smartwatch outputs into a single data collection platform. The project seeks to develop strategies for personalized dietary understanding and recommendations.

  • For context, PATHS-UP (The Precise Advanced Technologies and Health Systems for Underserved Populations) is an NSF-funded Engineering Research Center started in 2017. The 10-year, $14 million center’s mission is to “develop next-generation devices for people with chronic disease and underserved populations.” Collaborators include Texas A&M, UCLA, Rice University, and Florida International University.

Top Regulatory Highlights

FDA’s Dr. Courtney Lias Calls for CGM-Specific Decision Support Tools; Says 5% Time-in-Range Improvement Still Needs to be Demonstrated to be Meaningful

Closing out a session on Time in Range, FDA’s Dr. Courtney Lias expressed her concerns about using CGM point values (ignoring trend arrow information), especially when using dose calculators that are designed for BGM users. While the value of CGM is undeniable, Dr. Lias noted that many existing tools are designed for use with BGMs and may or may not be appropriate for direct use with CGM. In particular, “emerging data” suggest that using traditional insulin bolus calculators with CGM values may put patients at risk, due to slightly less accuracy in CGM point values. In order to manage these limitations, Dr. Lias called for the development of CGM-based tools that are able to utilize the powerful data provided by CGMs, including trend arrows. This is of course a tricky regulatory issue, since: (i) many BGMs that people routinely use to dose bolus insulin are less accurate than CGMs (JDST 2016); (ii) real-world accuracy of BGMs can be much lower than studied accuracy, as many don’t wash hands (real-world CGM accuracy is also lower than studied accuracy); and (iii) many CGM users in real-world use already take their CGM value and use it for bolus calculation (though would hopefully consider the trend and other info, rather than just populating the calculator with a single point value). Between Abbott, Dexcom (TypeZero), Lilly, Novo Nordisk, Companion Medical, Bigfoot, DreaMed, and others, there should be quite a few players willing to work with the FDA to develop these tools. We’re not sure if clinical trials will be required, though we’ve been waiting a long time for CGM-based bolus calculators that integrate the new trend arrow adjustment guidelines (FreeStyle Libre, Dexcom). Regulation of CGM-based decision support tools is also still an open question; class II seems possible for non-adjunctive iCGMs, as that is where bolus calculators within pumps and apps fall (to our knowledge).

  • During her discussion on Time in Range (TIR) as a regulatory endpoint, Dr. Lias listed a number of open questions, beginning with definitions of what glucose ranges and differences are meaningful. In the new consensus targets for CGM metrics, a 5% TIR improvement is considered clinically meaningful, but Dr. Lias stated that must be demonstrated in a study. Additionally, Dr. Lias noted that different CGMs have different performance levels at various glucose concentrations. She did not specify a specific brand/model, but this is a good point for the field to remember. (One comparison that comes to mind is FreeStyle Libre vs. Dexcom – FreeStyle Libre does seem to report more time in hypoglycemia than G5; see poster at EASD 2019.)

  • To begin her presentation, Dr. Lias provided a brief update on a recent reorganization of CDRH. The changes are designed, in part, to centralize the review of diabetes devices, but will not change the regulatory review process. Most notably, the change will bring decision support and insulin pumps into the same division as that of glucose monitors, the Division of Chemistry and Toxicology. We view that as very good news! Insulin pens will still be led in the hospital group, though the diabetes group will be consulted.

Ms. Naomi Schwartz: FDA Would Prefer All Devices Coming in Now Have the Option to be Updated

FDA In Vitro Diagnostics and Radiological Health scientific reviewer Ms. Naomi Schwartz expressed the Agency’s preference that cybersecurity issues be disclosed in coordination with the FDA and that “all devices that are coming in now have the ability to be updated.” The Agency has put considerable thought and effort into thinking cybersecurity issues, which has traditionally not been part of the “safety and effectiveness” mandate. During a Patient Engagement Advisory Committee meeting in September, it was clear that automated insulin delivery is one of the biggest areas the FDA is considering with regards to cybersecurity. Highlighting the difficulty in cybersecurity regulations, Ms. Schwartz noted that estimating the likelihood of a cyber-attack is nearly impossible and that with increasing connectivity and interoperability, a single attack may be able to affect large swaths of users. Additionally, she pointed out that while a device’s features are “positive,” i.e., things the device must be able to do, device security is “negative” and “infinite,” i.e., things the device must not do. Lastly, manufacturers have traditionally had a “passive” approach to cybersecurity, not patching security vulnerabilities until the next-generation product (the example that comes to mind is Medtronic pumps). This 4-5 year updating process was no longer safe or practical in an era where security vulnerabilities are discovered all the time – just last month, the FDA put out a safety communication regarding the “URGENT/11” security vulnerabilities. Manufacturers need to make devices that are updateable, and quickly updateable, e.g., “we can’t take six months for updates to be pushed to hundreds of thousands of devices.” This will soon be true with all insulin pumps, once Medtronic gets the MiniMed 780G out (Tandem’s t:slim X2 is the leader here, and we’d assume Insulet’s Omnipod Dash could be updated over Bluetooth or WiFi).

  • Ms. Schwartz highlighted the disclosure of the Animas Ping vulnerability (2016) as a good example of coordinated efforts by J&J and the FDA. The vulnerability was discussed with the FDA “early and often,” allowing the groups to assess risk and share mitigation strategies early on. Renowned diabetes hacker Mr. Jay Radcliffe, who actually discovered the Animas Ping vulnerability, has also credited that effort as an example of how manufacturers have improved at dealing with cybersecurity communications. Ms. Schwartz also brought up Abbott/St. Jude’s pacemaker security vulnerability communications (2017) as an example of poor coordination that resulted in delayed assessment and confusion.

    • In the FDA’s September meeting on cybersecurity, the issue of disclosure timing was a point of disagreement among Committee members. While some found it “paternalistic” to withhold information from patients, others maintained that patients shouldn’t be unduly burdened, especially when the scale of the risk is unknown or when mitigation steps are not available. The Medtronic pump cybersecurity issue is a case in point – it was not communicated by FDA and Medtronic until eight years after the issue was discovered. Ultimately, it’s clear that this is a multi-dimensional issue, but any communication should always be carefully considered with patient perspectives first – including numbers on the probable risk of harm and consideration of the risks of discontinuing a therapy.

  • Ms. Schwartz emphasized a multi-stakeholder approach to creation of regulatory guidelines, involving the FDA (and other international regulatory agencies), manufacturers, clinicians, and patients. We think patients could play a role in the regulatory process. While some aspects of manufacturers’ devices and algorithms may need to be hidden for commercial reasons, the paradigm of “security by obscurity” is no longer practical. By making the pre-market review of a device or algorithm more open, allowing the public to examine a product, any future issues and confusion could be reduced. We’ve heard Stanford’s Dr. Rayhan Lal express a similar sentiment towards the safety of DIY systems, which are open-source and public for anyone to examine.

  • Following the FDA’s 2018 premarket draft guidance, the Agency is working with MITRE to develop a common vulnerability scoring system (CVSS) to help assess future vulnerabilities. The International Medical Device Regulators Forum has also issued a new cybersecurity work item, which is being co-led by the FDA and Health Canada. The FDA is also working with the Department of Commerce and Patient Engagement Advisory Committee.

FDA’s Pre-Certification Working Model for Digital Health is Feasible; “Time Will Tell” Whether it Needs to be Adapted to be Used Prospectively

Mr. Bakul Patel (FDA) once again presented his vision for the FDA’s Pre-Certification model for modernizing digital health software regulation. Early on, Mr. Patel highlighted the need for a new approach to regulation: the FDA currently receives ~4,000 510(k) submissions per year, but as new digital health and software products continue to develop, the number of submissions will grow exponentially. Additionally, given growing concerns about cybersecurity (see above), the need for a more iterative and continuous approach to regulation is obvious – i.e., the hallmark of great software is not “perfection” at launch, but continuous improvement and iteration based on real-world data. The FDA’s Precertification program (“PreCert”) was developed in 2017 as a new organization-based approach to regulating software as a medical device, sort of like “TSA Pre-Check” – once pre-certified, companies will get a streamlined digital health software review. Nine pilot participants are in the program, seven of whom work in diabetes in some way. A working model for PreCert was created in 2018 (see our coverage of the meeting) and the feasibility of the model has been tested, applying the model in parallel and retrospectively with submissions that have gone through the traditional regulatory process. According to a mid-year update from the FDA, the working model has performed well, demonstrating that the company “excellence appraisals” (i.e., are you a great software company? How good are your processes?) and “streamlined review” were sufficient to make regulatory decisions. Mr. Patel stated today that feasibility testing will continue through “this year and a little next year” with alpha-testing of the program starting “soon,” to test the PreCert approach with submissions from the program’s nine pilot participants – this could have the most immediate diabetes implications for Tidepool Loop, the automated insulin delivery iPhone app. The alpha-testing will help the FDA determine whether the PreCert program can work prospectively (“time will tell”) and whether the program is scalable.

  • “You’ve seen many updates, and that’s on purpose. The message is, ‘We don’t know all the answers for this vision to come true, but we want to get help from people and want to share how we’re progressing, so that when we actually put this program out to test, we can actually know it can actually work.’ … What we’re talking about is moving the oversight mechanisms and the information we get to different parts of the life cycle of product delivery. How can we provide benefits for the FDA and for industry and for users?” – Mr. Patel

Real-World Evidence Can Serve as Control Arms in Trials, Support Indication for Expansion; Evidence Must be “Fit for Purpose”; FDA Developing Standards for Real-World Data

Dr. Yiduo Wu (FDA) kicked off his presentation on the role of real-world evidence in regulatory decision making by defining the difference between real-world data and evidence: real-world evidence (RWE) is clinical evidence supporting potential benefits/risks of using a product derived from analysis of real-world data. Dr. Wu outlined many potential uses for RWE in a product’s life cycle, from product development to post-market surveillance. RWE could help inform hypothesis generation, inform trial design, serve as the control arm in a clinical trial, or support indication expansion for products. Dr. Wu also noted that sometimes real-world evidence collection is a mandatory condition of approval for products. However, to take advantage of real-world evidence, the data collected must be “fit for purpose,” which Dr. Wu defined as meaning the data must be “complete, consistent, accurate, and contain all critical data elements needed to evaluate a medical device and its claims.” To improve reliability of real-world data collection, Dr. Wu suggested manufacturers pre-specify the data elements to be collected, unambiguous definitions of those data elements, methods for aggregating and documenting data, and a timeframe for the data to be collected. To help organizations with this process, Dr. Wu shared that the FDA has been working with the CDC, NIH, CMS, device manufacturers, EHR vendors, standards developers, and others to develop standards for data and data interoperability.

  • As an example of RWE use cases, Dr. Wu noted that RWE was considered as part of the non-adjunctive indication expansion for Dexcom G5. The real-world evidence considered included comments and documentation from patients, caregivers, and providers who used the G5 off-label non-adjunctively, as well as data collected during real-world routine use. An upcoming test of real-world data could be Tidepool Loop’s FDA submission – it is using an observational, real-world, virtual study (Jaeb) to show safety. See Howard Look’s update at Diabetes Mine last week.

  • Dr. Wu provided a list of questions for manufacturers to ask when submitting RWE to the FDA: (i) were appropriate variables collected?; (ii) was the endpoint meaningful and defined in a consistent way?; (iii) was the period of data collection appropriate for the chosen endpoint(s)?; (iv) was the chosen population appropriate and representative?; and (v) was the analysis of the real-world data appropriate for the question?

Top Smart Pen and Insulin Highlights

Novo Nordisk’s Connected NovoPen 6 and Echo Plus to Launch in 2Q20, Removable Pen Attachment in 2021; Decision Support in Pipeline

In a brief update on Novo Nordisk’s connected pen strategy, we learned that the already CE-Marked NovoPen 6 and NovoPen Echo Plus pens will launch in 2Q20, presumably starting in Europe where they already have regulatory approval. The timing is in line with the “2020” timeline given in Novo Nordisk’s most recent update, and about a year back from the “early 2019” goal given when the devices were first announced – it’s important to get the product and manufacturing right here, so we appreciate the cautious approach. As expected, the reusable NovoPen 6 and Echo Plus will display the last insulin dose and time on their ends, have an 800-dose memory, and five-year battery life. Data from the pens can be downloaded with NFC and the pens are compatible with both basal (Levemir, Tresiba) and bolus (NovoLog, Fiasp) insulin cartridges. The NovoPen 6 will allow adjustments down to 1 U with a maximum 60 U dose, while the pediatrics-targeted Echo Plus will allow adjustments down to 0.5 U with a maximum 30 U dose. We also got a look at the reusable pen attachment designed for Novo Nordisk’s disposable pens, referred to as “Dialogue” at EASD 2019. That device is now targeted for launch in 2021, and while behind the initially anticipated launch, we’re happy to see decisions against rushing this to market and we’re glad Novo Nordisk is taking the time that it needs to get to market. It appears Novo Nordisk will launch its own app for that product, whereas the NFC pens will be more focus on in-clinic download. Presumably there is much more complexity with the Bluetooth-enabled add-on attachment. There’s so much room for fantatic apps we hope it is investing in this area and taking the time required to get this right. The first insulin pen was launched in 1985 – we’ve had over 30 years of pens that are not connected and we believe the broad diabetes ecosystem prefers organizations to take the time needed to make sure this is the best possible development and launch rather than rushing it. We also appreciate that this is a new arena and that timelines are challenging to make since there is much ‘unknown’ that can emerge.

  • Conspicuously grayed out in the roadmap above are three “digital therapeutic” projects. Dr. Anders Toft, Novo Nordisk’s Corporate Vice President of Commercial Innovation briefly referenced the projects as something the company is working on “very actively”: “That is where we see a major future promise with the data that comes passively from smart pens and connected CGM/BGM – leveraging that with algorithm-powered aps that can guide a patient safely and effectively to [better management].” Novo Nordisk already has an impressive list of data integration partners for its smart pen devices: Roche/mySugr, Dexcom, Glooko, Abbott, and Medtronic. Dr. Toft emphasized that these partnerships were “non-exclusive” and that the “data will belong to the patient.”

  • Dr. Toft presented results from the 12-site Swedish pilot study (n=94) with CGM (Dexcom G4 or Abbott FreeStyle Libre) and NovoPen 6 (results were first shown at two posters at ADA 2019). Two-weeks of CGM metrics at follow-up (14 days after the final clinic visit) were compared to two-week baseline metrics (14 days following obtaining the pen). Compared to baseline, mean time-in-range improved from 38% to 46% (+1.9 hours/day), driven by time >180 mg/dl decreasing from 49% to 42% (-1.8 hours/day). 81 adults were included in the adherence analysis, which showed a very impressive 43% fewer missed meal boluses (baseline: 0.74/day). We are also struck by how much better patients can do than 46% time in range – there must be multiple interventions that can go along with therapeutic and technological ones (apps like Under My Fork, etc. – though it’s easiest to use these with CGM, connected blood glucose monitors work also).

Sansum’s Dr. David Kerr: Smart Pens Have the Biggest Potential to Improve Outcomes (Outside of CGMs), But May Affect Clinician/Patient Dynamics

Sansum Diabetes’ Dr. David Kerr presented an optimistic overview of where he thinks smart pens will best fit into the current diabetes ecosystem, calling the device a “potential technology that [he] think[s] will be disruptive.” Other than with CGM, Dr. Kerr believes that insulin pens specifically have the most potential to improve outcomes by addressing insulin omission, which he categorized as (i) not giving any insulin or (ii) giving a dose of insulin that is not necessarily safe in amount or timing. Indeed, a study from March found that 36% of basal doses and 24% of fast-acting doses were missed or forgotten by people with diabetes. He noted that for clinicians in particular, not knowing how much insulin a patient has been taking makes it difficult to have meaningful dialogues about treatment, and this problem is especially prevalent in patients using MDI instead of pump therapy. Uniquely, smart pens can now allow patients and practitioners to “get a sense” of insulin omission. 

  • Interestingly, Dr. Kerr highlighted strain on the clinician/patient relationship as one of the potential downsides to using smart pens. Considering that diabetes can be a “difficult and miserable” condition to live with on a day-to-day basis, Dr. Kerr explained that enabling doctors to see how well a patient is taking insulin, or omitting it, can potentially complicate provider dynamics, especially for doctors in the primary care setting who may have less experience with patients with diabetes.

  • Dr. Kerr also introduced a new acronym T2iD or “Transitory Technology for Insulin Delivery.” He noted that with the influx of smart pens, technology might start to be used more intermittently, and patients might move between multiple devices based on their day-to-day needs. Dr. Kerr further surmised that smart pens might seriously impact the pump market. We believe the evolution of smart pens is so early that this is hard to call. On one hand, the rate at which pump-based AID is improving is radically faster than the smart pen field is currently moving; on the other hand, a compelling smart pen, CGM, and titration app could theoretically offer high time-in-range at a much lower cost.

  • Overall, Dr. Kerr seemed bullish on the future of smart pens, calling it a “very hopeful time for this technology.” Looking at his Roadmap to Smart Insulin Pens, Dr. Kerr pronounced that products in the “blue area” – referring to late-stage smart pens with dose calculators or advanced decision support – are starting to appear on the  market. Interestingly, we’re seeing a wide variety of players entering into the smart pen arena, from first-to-market Companion Medical’s InPen to big name contenders like Novo Nordisk (see above), Lilly, and Sanofi.

  • Dr. Kerr prefaced his talk with a firm reprimand to the overall diabetes community regarding lack of high-quality guidance, particularly when it comes to insulin. Despite insulin being a “very tricky and dangerous” treatment, the community often asks patients and clinicians to make important decisions in the absence of basic information: how much insulin to give and when. Citing an August 2019 paper in JAMA entitled “Evolution of the Cascade of Diabetes Care in the United States, 2005-2016,” Dr. Kerr noted that meaningful outcomes for patients with diabetes are not improving, especially for young, female, and non-white adults. (And, actually, especially for everyone, according to the 2019 Diabetes Atlas, just out today – see our update here.)

Profil’s Dr. Tim Heise on Insulin’s Path Forward: Questions Up-Titration Required of Once-Weekly Basal, Current Oral Options; Future of Prandial Insulin “Not Further Improvement in Time Action Profile”

Co-Founder and Lead Scientist at Profil Dr. Tim Heise “gazed into his crystal ball” to give a nuanced perspective on the future of insulin. To start, Dr. Heise pointed out that despite recent advances in insulins, including improved basal pharmacodynamics and the invention of ultra-rapid acting compounds, patients with both type 1 and type 2 diabetes continue to experience worse outcomes. Dr. Heise credited this incongruity in part to the fact that insulin is often initiated far too late in the treatment regimen, noting that in patients with type 2 diabetes, “it’s very clear that if you don’t use these wonderful insulins or use them very late, you can’t show improvements.” (A counterpoint would be that if type 2 diabetes is characterized by hyperinsulinemia and insulin resistance feeding each other in a vicious cycle, treating the glucose with more insulin does not address the underlying pathophysiology.) Therefore, a clear unmet need for insulins that are easier to initiate and adhere to exists – in particular, a 2012 global survey published in Diabetic Medicine showed that reducing the frequency of injections is a top priority amongst both patients and HCPs. Although he acknowledged that the speed of meal-time insulins has vastly improved, Dr. Heise also intriguingly added, “it’s easy to predict for an insulin pharmacologist that the future of [prandial] insulin probably is not further improvement in the time action profile.” Gazing ahead, Dr. Heise gave (i) glucose-reactive “smart” insulins; (ii) hepato-preferential insulins; (iii) oral/inhalable insulins; and (iv) once-weekly basal insulins as possible solutions that are currently in development.  

  • In terms of oral insulin, Dr. Heise emphasized that the primary current challenge is increasing bioavailability. While previous candidates struggled with issues of cost and stability, Novo Nordisk’s now-discontinued phase 2 candidate OI338GT was the first to demonstrate the aforementioned attributes and promising efficacy and safety, but required a commercially infeasible dosage. Specifically, for prandial insulins, Dr. Heise asserted that oral insulin’s “substantial food effect” (i.e., decreased efficacy if taken right before a meal) “makes it very difficult to believe that prandial oral insulins will have a future.” Dr. Heise also casted doubt on Oramed’s recent “successful” phase 2b study of oral insulin in type 2 diabetes, highlighting (i) the study’s 22% drop out rate and (ii) lack of statistical significance in the thrice-daily arm, despite the highest dose of oral insulin.

  • Using his own simulation data, Dr. Heise also provided interesting insights on the extensive waiting period that would be needed for a once-weekly insulin to reach steady state, as well as the potential real-world challenges of initiation. With a 200-hour half-life, a once-weekly insulin would require a substantial 40+ days to reach its full effect (see graph below). One way to speed up the up-titration period would be to take an additional “loading dose” at the first injection, but Dr. Heise commented that this might not be a good solution “psychologically,” as a patient on 40 units of insulin/day would have to be convinced to take a whopping 560 units of insulin at drug initiation. Currently, Lilly and Novo Nordisk both have candidates in phase 2, while Sanofi and Hanmi have partnered around a preclinical candidate.

Q: Insulin is not associated with a decrease in mortality benefit compared to oral drugs, but our goal with increasing glucose control is decreasing complications. Can you give commentary on why no change is seen on insulin?

A: That’s certainly an excellent question. I wish I had a good answer to that. It’s probably the shortcomings of insulin therapy that prevent using it in a good way. We could probably achieve even better glycemic control and CV benefit. When you look at the DEVOTE study results, there is at least a small trend to improved CV outcomes. The hope is that with better insulins, you can titrate better and achieve better glycemic control, but so far, we haven’t been there yet.

Technosphere Insulin Reduces “Late Post-Meal Hypoglycemia”; No Updates on BluHale Dosing or CDE Impression Studies

Chief Medical Officer and Executive VP of MannKind Dr. David Kendall gave a bullish outlook on the future of Afrezza (Technosphere insulin). There were minimal updates on the company’s Bluetooth-enabled inhalation monitor BluHale. Although the company stated that abstracts from the (i) BlueHale V2.0 with Technosphere insulin study, showing “improved dosing with BluHale” and (ii) BluHale Trainer study, documenting CDE impressions on the utility of Bluhale, had been submitted to DTM during its 3Q19 update, we did not hear briefings on either of the findings in this talk or see any posters listed on the program. Dr. Kendall did however show 2018 proof-of-concept data from two participants using BluHale’s dose detection feature, with dose readings overlaid onto a Dexcom CGM trace. Dr. Kendall noted that with the CGM data, one participant (see photo below) realized they were virtually always taking Afrezza while in hyperglycemia and might need to take the drug more “pre-emptively,” hinting at potential synergy between Afrezza, BluHale, and CGM.

  • As expected, Dr. Kendall focused most of his presentation on Afrezza’s favorable post-meal glucose control compared to injected rapid-acting analogs. He began by defining what characteristics “truly” make up an ultra-rapid acting insulin: (i) rapid onset of glucose lowering effects in <15 minutes; (ii) peak effect in 30-60 minutes, to rapidly suppress EGP; and (iii) duration of effect of ~two to three hours, as in healthy individuals, as well as “not just be faster than the last fast-acting insulin.” Afrezza fulfills all these requirements and Dr. Kendall heavily emphasized pulmonary absorption as the most “natural” means of biomolecular exchange. Afrezza is an interesting case study of a product that meets a clear need – insulin is too slow around meals – and yet hasn’t seen strong uptake in the market. Dr. Kendall addressed this exact point in Q&A:

  • Dr. Kendall highlighted the difference between pulmonary and subcutaneous insulin units throughout his presentation. It seems that this conversion factor of ~1.5-2x pulmonary units per subcutaneous unit has been a roadblock in communicating Technosphere insulin’s efficacy to regulatory bodies and HCPs. All of the data presented used this conversion factor when comparing Technosphere insulin to other treatments.

Q: Can you speak to the adoption of Technosphere insulin in clinics? My perceptions are that it’s quite low. Can you give any reasoning as to why it’s so low?

A: Thanks for your question. Yes, you’re right – uptake relative to injectable insulin is low. We understand injectable insulin quite well. There’s a learning curve to the Technosphere device and delivery. In addition, the understanding of the clinical profile as a pulmonary delivery option, that we often didn’t see before CGM was utilized, is something we’ve come to understand. Second, is the understanding of dosing. The FDA understands a unit as a unit, not the ‘less potent’ unit as a unit. Third, because uptake is limited, it’s difficult to get at the pharmacy, and coverage is low; fourth is the lack of familiarity.

Dr. David Harlan (University of Massachusetts): I will say that I have a few select patients on it, and they love it. The very active individuals love it.

Q: What is the patient preference? Has that been analyzed?

A: There is a preference for not taking injections. Again, once the learning is there, the individuals who utilize it, when they experience the in-meal dosing, the quick on and quick off, it has kept this their preference. I can give an n=1 example, who is my father in law. He dropped his pump after many years because of lipohypertrophy, and he’s not had numbers this flat. Acceptance is good once the learning happens.

Dr. Lutz Heinemann Advocates for More Attention on Waste from Diabetes Products, Commends Novo Nordisk’s Limited Plastic Use in Insulin Pens, and Encourages Patients to Push for Change

Science & Co.’s Dr. Lutz Heinemann called on manufacturers, regulatory agencies, and patients to pay more attention to the issue of environmental waste created by diabetes devices. Abbott’s FreeStyle Libre and Dexcom’s G5 CGM inserters contain 71 and 80 grams of plastic, respectively, which instead of being thrown in the garbage, could be reused. With “almost 1.6 million” FreeStyle Libre users, this would translate to a remarkable 6 million pounds of plastic (or approximately the weight of a Saturn V rocket from the Apollo program). While progress has been slow over the last decade, stakeholders including MedTech Europe, Novo Nordisk, and Biocon are beginning to act. For example, at a June 2019 MedTech Europe meeting, an Environmental & Sustainability Committee Working Group was held for the first time. Additionally, Novo Nordisk has a strict zero environmental impact policy, and the design of its FlexPen and FlexTouch insulin pens are environmentally-friendly. These pens contain only two types of materials which enables easy separation of the plastic waste into components during production and sale or reuse in other products. Biocon also may replace disposable plastic insulin pens with reusable pens.

  • During Q&A, Novo Nordisk’s Dr. Anders Toft said that the company hopes to be able to reuse half of its plastic by 2030, but was struggling to get pre-filled pens back from patients once empty and wanted to identify a patient engagement strategy. This is not surprising – unless the return approach is dead easy (the prestamped enveloped is included in the package), it won’t happen. Mr. Ed Krisiunas from WNWN International, a consulting firm focused on healthcare waste management, said that the Novo Nordisk needs to directly speak with end users and use those conversations to create an incentive. Any engagement opportunity requires companies to bring patients to the table and collaboratively work with them to identify gaps in the system.

  • One audience member noted during Q&A that the diabetes community actively thinks about the environmental impact of technology and believes that industry needs additional work on incentivizing patients. To make her point, she told a story about how she was unaware that there were three safe-needle disposal spots within a fifteen-mile radius of where she lived. She encouraged companies and regulatory agencies to think more about the burden of environmental waste and to include the diabetes community in conversations promoting policy change.

Top AI and Decision Support Highlights

One Drop CGM Predictions 30-120 Minutes in Advance; Aim to Launch in Mid-2020; Strong Accuracy for 30-Minute Prediction

One Drop shared a poster and press release describing a new machine learning effort to predict CGM values 30-120 minutes in advance, building on previous work with fingerstick data for people with type 2 diabetes (ADA 2019, October 2018). Data from ~3,000 One Drop app users – including 10+ million hours of CGM data, gender, year of diagnosis, and other self-reported data (food consumed, medications, and activity) – was used in the analysis. A random sub-sample of users was used to train a machine learning model (“seen users”) to predict CGM values into the future (30, 60, and 120 minutes). The model was then used to predict CGM values for people in the larger sample whose data were not used to train the model (“unseen” users). The 30-minute CGM forecast was quite accurate – falling within 20 mg/dl a strong 91% of the time. As expected, this dropped to 63% for a 60-minute forecast and 41% for the very ambitious 120-minute forecast. The Clarke A-Zone stats looked better, with 97% of predictions in Zone A for a 30-minute forecast, 77% in Zone A for a 60-minute forecast, and 57% for a 120-minute forecast. See the complete table below for all the stats (we’ve focused on “unseen” users, since that is more reflective of the model’s predictive capabilities). Download the poster here.

  • One Drop ambitiously hopes to make these forward-looking CGM predictions available to app users in mid-2020, using real-time CGM data. The regulatory classification is “TBD,” so this timeline could change quite a bit. It’s unclear to us where FDA would draw the line on this – is the 60- and 120-minute accuracy good enough? The planned implementation would also require immediate real-time CGM data flow, a change for One Drop – currently it pulls retrospective Dexcom CGM data through Apple Health and the Dexcom API. One Drop has done a nice job of already envisioning what this feature would look like – see the picture below, which resembles the existing feature for fingersticks/type 2 diabetes.

  • Importantly, One Drop used real-world data here, which means CGM, food, medication, and activity data were incomplete. Missing CGM data is also guaranteed, given expected periods of interruption due to sensor replacement, data transmission errors, or other reasons.

  • How useful are 30-,60-, and 120-minute CGM forecasts? We imagine this feature could useful for some users (i.e., real-time pattern recognition), but potentially drive alarm fatigue in others (i.e., another notification on top of high/low CGM alerts). This approach will obviously require maintaining and engaging with two apps – the CGM app and then the One Drop app. Dexcom has always been cautious about ensuring alarms are actionable (i.e., not too many false alarms), so we’ll be interested to see how it approaches this idea. 

    • For context, Medtronic has been working with IBM Watson for several years on similar machine learning/CGM prediction efforts, launching a hypoglycemia prediction feature earlier this year (IQcast) for Guardian Connect (up to four hours in advance). We have not heard any community feedback on the value of IQcast, and it does not appear to have driven major Guardian Connect CGM uptake in the US (to our knowledge). As of Medtronic’s last pipeline update on Sugar.IQ (ADA 2019), it hoped to launch a “smart guide” system by April 2020 (FY20), adding “basic” CGM-based insulin dosing guidance (presumably trend-based bolus calculator), a predictive trace, and some advising.

Dr. Jessica Castle Shares Helmsley-Funded Early-Stage Decision Support App “DailyDose” for CGM + MDI; Plans for n=20 Pilot Study

Dr. Jessica Castle (Oregon Health Sciences University) shared promising details regarding a new decision support app “DailyDose,” under development by OHSU and University Health Network – check out the user interface below. Using input from Dexcom G6 and Companion Medical’s connected pen, InPen, the DailyDose app integrates data to generate a comprehensive collection of outputs including:

  • A “smart” bolus calculator (presumably a CGM-based bolus calculator – something no company has gotten cleared yet);

  • A machine-learning recommender engine that gives weekly recommendations for insulin:carb ratios, correction factor, and basal insulin dose (this is so needed!);

  • An “Insight” page that explains to users why a recommendation was made;

  • Hypoglycemia prediction alerts (30-minute prediction horizon); and

  • Carbohydrate intake and insulin dose adjustment for exercise (based on JDRF’s PEAK guidelines).

We first heard about DailyDose at ADA 2018 as a decision support app co-developed by University of Toronto’s Dr. Joe Cafazzo and funded by Helmsley Charitable Trust. Recommendations are specifically programmed to prioritize improving time-in-range (TIR) and reduce hypoglycemia, in conjunction with a post-prandial hyperglycemia prevention algorithm (ALPHA). In terms of quality, a healthcare provider first sets the “range” of insulin settings that the recommendation engine can toggle between (i.e., “if my patients starts at a ratio of 1 U/10g carbs, recommendations can only go up to 1 U/5g carbs”), and “expert opinions” are used to prune out potentially erroneous recommendations. We’re curious to see how in-depth these “expert opinions” are and to what extent they regulate recommendations. Of note, the app’s final quality control measure is that patients have the option to accept or decline all recommendations, which Dr. Castle highlighted as an important safety feature.

  • Preliminary computer-simulated results (n=29) demonstrated promising improvements in TIR (70-180 mg/dl) and time <70 mg/dl – from 64% time-in-range at baseline to 83% by week 12! Wow! For the most part, the algorithm-generated recommendations matched physician recommendations, with a 56% overall agreement rate and a 3% overall disagreement rate (we’re not sure why they don’t sum to 100% - ostensibly the other 41% is a mix of agreement and disagreement). Although in-silico results must be interpreted cautiously and development is still early, reaching 70%+ time-in-range (to say nothing of over 80%, along with less than 1% below 70 mg/dL) with a smart pen decision support app is very promising! We’re eager to see how the app will compare to pump-based closed loop systems, as well as what the commercialization plan is – would Dexcom buy this, adding to its TypeZero portfolio? Dr. Castle emphasized that DailyDose is meant to take the place of an existing CGM app, so that users are not obligated to use multiple apps; we’re curious to see how comfortable users will feel relying on AI-generated dosing recommendations.



Week 4

Week 12



74% (p<0.001)

83% (p<0.001)

Time <70 mg/dl


0.8% (p<0.001)

0.8% (p<0.001)

  • Moving forward, Dr. Castle shared plans for an upcoming DailyDose pilot study in adults with type 1 diabetes. A total of 20 adults on MDI with an A1c ≥7.0% will be studied over the course of 9.5 weeks. Study participants will be given a Dexcom G6 and InPen for ten days of “usual care,” followed by 8-weeks using DailyDose. Time in Range, the primary outcome, from the ten-day run-in period will be compared to the last ten days of use.

  • Elsewhere on the decision support front, DreaMed’s provider-facing Advisor Pro is FDA cleared to provide pump setting recommendations (basal rate, insulin:carb ratio, correction factor) based on BGM or CGM data. According to our interview with DreaMed CEO Eran Atlas, a pilot for Advisor Pro in MDI has likely already commenced and support for type 1 MDI users is targeted for 2020. Dexcom also owns TypeZero’s MDI decision support software, which has done a study at UVA; we’re not sure when final outcomes will be presented, though preliminary outcomes from ATTD 2019 were mixed.

Roche and IBM Study Uses Machine Learning and Input from More than 70 Data Sources to Predict Blood Glucose up to ~3 Hours In Advance

Roche’s Dr. Rolf Hinzmann spoke about a study Roche and IBM conducted to use machine learning to accurately predict blood glucose values using more than 70 factors unique to a patient. The study collected numerous pieces of data from 221 type 1 patients: (i) 70 static parameters (e.g., demographics, vitals, diabetes therapy); (ii) 24/7 results from BGM/CGM, insulin, nutrition, and GPS; and (iii) five variables from Empatica wearables: activity, heart rate, heart rate variability, body temperature, and transpiration. Wow! Using machine learning techniques, researchers were able to differentiate the study group into subpopulation “precision cohorts” based on the similarity of glucose profiles, and through iterations, enabled the model to better predict glucose values for each cohort. Dr. Hinzmann mentioned that the final goal is to create an app that personalizes blood sugar predictions to the patient based on data from insulin pens, CGMs, BGMs, fitness trackers, food, and exercise (i.e., like One Drop’s poster above or Medtronic’s Sugar.IQ). The work still appears to be in the research phase, though perhaps we will eventually see predictions like these become part of mySugr, Roche Accu-Chek BGM app, or partner Senseonics’ Eversense CGM app. We did not catch information on the timeline, location of the study, or results, he mentioned during Q&A that the glucose prediction values were “accurate” for up to three hours.

  • Interestingly, this study appears to be very similar to a different one Roche and IBM also published in Nature Medicine which created a new algorithm that could accurately predict if a patient was at high risk of developing chronic kidney disease. In the study, EHR data from type 1 and 2 patients (a whopping n=417,912!) including age, body mass index, A1c levels, and concentrations of creatinine and albumin were used to train the algorithm. After the training phase, an independent sample set of data from 104,504 additional individuals in the same database was used for validation. The algorithm was then applied to data from type 2 patients (n=82,912) included in the Indiana Network for Patient Care (INPC) database. Notably, these extracted datasets represent the largest real-world data that has been used to assess CKD as a complication of diabetes. The algorithm received an area under the receiver operating characteristic curve (AUC-ROC), a measure used measure the quality of machine-learning algorithms, score of 0.79 and outperformed the algorithms derived from other major clinical trials like the ONTARGERT, ORIGIN, RENAAL, and ADVANCE studies. For context, an AUC-ROC of 0.79 means the model has a 79% likelihood of properly distinguishing between positive and negative CKD cases – quite impressive.

  • Medtronic has also been working with IBM Watson on similar CGM prediction efforts, including the IQcast hypoglycemia prediction feature launched in January. A “smart guide” system for Sugar.IQ was previously expected to launch by April 2020 (FY20), adding “basic” CGM-based insulin dosing guidance (presumably trend-based bolus calculator), a predictive trace, and some advising. By April 2021 (FY21), Medtronic aims to add a “virtual optimizer” bringing more personalized dosing (meals, activity) with the Synergy disposable sensor (see Medtronic’s ADA 2019 update).


--by Ani Gururaj, Rhea Teng, Adam Brown, Albert Cai, and Kelly Close