DTM 2017 (Diabetes Technology Meeting)

November 2-4, 2017; Bethesda, MD; Days #2-3 Highlights – Draft

Hello from Bethesda and a very eventful final two days of DTM 2017! We have a record 21 highlights from days 2-3 in four areas: (i) CGM; (ii) Automated Insulin Delivery; (iii) Smart Pens, Insulin Dose Titration, Clinical Decision Support, and BGM; and (iv) Insulin, Insulin Delivery, and Insulin Monitoring. Get the detailed highlights below! Boy, this field has never been busier or more exciting.

We congratulate Harvard’s Dr. Frank Doyle for winning DTS’ Artificial Pancreas Award, as well as Sansum’s Dr. David Kerr for winning the Diabetes Technology Leadership Award.

In case you missed it, you can also find our day #1 report here.

Table of Contents 

CGM Highlights

1. Dexcom G6 No-Cal Feasibility (N=165): 9.3% MARD vs. YSI; Decision Support Screenshots; Verily Gen 1 Updated Design; Diabetes education “Programs”; Smart Pens!

Following last year’s initial no-calibration G6 feasibility accuracy data, Dexcom’s Peter Simpson shared brand new results (9.3% MARD!) from four aggregated feasibility studies with the same platform. He also discussed forays into connected pens, AID, the Verily-partnered sensor pipeline (modified design for gen one), decision support and CGM-enabled diabetes management programs (first screenshots), smartwatch partnerships (finally a mention of direct-to-Apple-Watch transmission). With Abbott’s factory-calibrated FreeStyle Libre sensor coming to the US market by end of year, the pressure is on Dexcom to eliminate fingersticks – fortunately, the company shared on its 3Q17 call that G6 was submitted to FDA with one fingerstick calibration per day in 3Q17, but there is also now a pathway to launch “no calibration” G6 sometime before the end of 2018. Highlights and screenshots below!

  • Accuracy for the next generation Dexcom CGM without calibrations from four aggregated feasibility studies (n=165 adults + pediatrics) was highly impressive and in line with DTM 2016 data: an MARD of 9.3% vs. YSI at five in-clinic visits (n=4,555 paired points) – with no calibrations. Outliers were uncommon, with a remarkable 91% of readings within 20%/20 mg/dl. Even on day one, MARD came in at 10%, with 89% of points within 20%/20 mg/dl – this is certainly strong enough for dosing insulin. The company also broke out impressive data on accuracy in hypoglycemia (conferred by a membrane that blocks out background signal): an MAD of only 8 mg/dl vs. YSI and 94% of readings within 20 mg/dl. Mr. Simpson reminded the audience of the new sensor’s removal of the acetaminophen contraindication (lots of people will be happy about this). He later noted that the company is “very excited” about the product and confirmed the prospect of launching in 2018. As a reminder, G6 was filed with FDA for use with one calibration/day in 3Q17, but there is also an FDA pathway that could enable a G6 no-calibration launch in 2018, either after the one-calibration version or instead of it. (This will be determined by FDA discussion over the coming months.) The slide (see below) positioned the next gen as “no calibration,” perhaps lending more confidence that G6 will indeed launch from the get-go with no cal.
  • We haven’t yet seen accuracy data from the large G6 pivotal trial testing one calibration per day (n=300+ and 30,000+ matched CGM-YSI pairs). In Q&A, Mr. Simpson shared that MARD is important, but the “next layer of goals for accuracy” right now is to eliminate outliers, noting that patients trust the system more and have less false alarms when there aren’t outliers.

  • Accuracy was strikingly comparable to the recently-FDA-approved factory-calibrated, 10-day wear, 12-hour warmup Abbott FreeStyle Libre (MARD 9.7%, 91% of values within 20%/20 mg/dl; see Libre’s US label here, “Performance” section). Of course, there are meaningful product differences (real-time vs. flash, alarms vs. no alarms, form factor, cost), and Abbott may have a 6-12 months head-start on no-calibration in the US, depending on when G6 launches in 2018. Next year is going to be quite the year for CGM in the US!
  • Mr. Simpson didn’t actually refer to the “next generation” sensor as G6, and neither did the slide deck. We’re not sure how deeply to read into the omission, because all other aspects of the system are equivalent (form factor, applicator, 10-day use, 2-hour warmup). It is possible that the no calibration version will be called something different, particularly in the scenario that Dexcom needs to differentiate it from the one calibration/day version.
  • A comprehensive forward-looking slide showed specific decision support screenshots, implying a CGM-based bolus calculator that includes trend information. The slide also showed “CGM enabled diabetes management programs” focused on exercise, diet, sleep, and medication optimization. Great news!

  • The same slide also showed a slightly flatter gen one Verily on-body profile, at least relative to what we saw at DTM 2016. We think the updated gen one will be more competitive with Abbott’s FreeStyle Libre.  This gen one will move Dexcom to a smaller wearable and a fully disposable design. Per Dexcom 3Q17, timing on the Verily gen-one product depends on the G6 no calibration regulatory strategy.

  • Also in line with 3Q17, Dexcom’s slide showed a connected pen and automated insulin delivery pump partners (Tandem, Insulet, Beta Bionics). The 3Q17 update suggested there are “irons in the fire” on connected pens, but did not provide a concrete update. The picture in the slide was a generic illustrated pen, which makes it hard to know who Dexcom is working with at this stage... All three insulin players are working on smart pens – Lilly is a Companion Medical investor, Sanofi is a Common Sensing investor/partner, Novo Nordisk has its own smart pen piloting in Sweden – so we’d guess many partnerships are possible here.

  • For the first time, a Dexcom slide showed the direct-to-Apple Watch, cloud-connected potential with Series 3. Notably, Dexcom has not actually commented on this since Apple announced the potential in June. We see enormous potential for this on a convenience front, especially for kids (i.e., no need for a phone to be carried around). The slide also showed the newly launching G5 touchscreen receiver and the just-signed Fitbit Ionic integration partnership (launching in 2018). Mr. Simpson stressed the priority to make the CGM experience seamless for patients. Data on the wrist without a phone nearby will be a big step forward in our view. Now, we’d just love to see the experience grow more seamless for ALL customers, namely Medicare beneficiaries who are ludicrously not allowed to share data with caregivers and providers.

  • The public API platform that launched in September has >300 registered developers and seven data partners including One Drop, Tidepool, Rimidi, and Nutrino. While Apple Health data sharing is good for many iOS users, this API integration will enable an even bigger ecosystem of useful CGM data apps (e.g., Android and beyond). We’re very excited to see Dexcom pushing this ahead and hope to see far more partners and users get on board.
  • Mr. Simpson made some enticing comments during Q&A on wear time and aspirations: “The biggest challenge in extended sensor wear is reliably the adhesive patch. There’s a lot of science behind that – we’re working on some custom adhesives, 90% work up to 14 days, possibly longer. Our sensors will last for months on the benchtop – in one publication, a sensor lasted for three months. The foreign body response limits it for some people. 14 days is fairly straightforward, but then you have to deal with biology past that. Membranes and manufacturing consistency comes up, as we scale to tens of millions of devices per year, that’s our focus.” We wonder if a 21-day sensor is possible with the right adhesive design, or perhaps even a 30-day design.

2. Medtronic’s Sugar.IQ Improves TIR by 33 Mins/Day, Good Engagement; ~10.4% MARD for GS3 in Real World; 670G Peds Trial Complete (7-13 Yrs); Now ~10,000 on 670G

Medtronic’s Dr. Robert Vigersky shared new Sugar.IQ outcomes data and reported encouraging accuracy data on the Guardian Sensor 3 (~10.4% MARD vs. fingersticks in real-world 670G use). In an email conversation with the company related to this talk, we’ve subsequently learned there are now “about 10,000” patients on the MiniMed 670G, a big increase since the last update (~1,000) and nearly one-third of the ~35,000 in the priority access program. The IBM Watson-partnered Sugar.IQ pattern recognition app showed a small benefit on time-in-range in the continued limited launch, but did show very good engagement, positive responses to the “insights,” and a decrease in prolonged hypoglycemia and hyperglycemia events >120 minutes (see below). The app is expected to launch in tandem with the Guardian Connect Mobile CGM later this year (per Medtronic 2Q17), though the latter is still under FDA review (over a year now). This app will be critical for Medtronic as a CGM differentiator, since Guardian Connect as a standalone CGM lags behind Dexcom and Abbott on other important device metrics (daily calibrations, phone/Watch device compatibility, on-body form factor). On the plus side, Guardian Sensor 3’s accuracy in both the real-world setting in adults and the now-complete pediatric (ages 7-13 years) 670G trial were confidence-inspiring – a continued MARD of ~10% (~3 calibrations per day). There was no update on the constrained sensor manufacturing due to global sensor demand and storm-related global manufacturing disruptions; we continue to assume Guardian Sensor 3 sales won’t really pick up until early next year.

  • In Sugar.IQ’s continued ongoing “limited learning launch” phase, 256 users showed encouraging engagement metrics. In an n=136 subset of users (with 2+ weeks of data), Sugar.IQ’s glycemic insights helped increase time in range by a modest 33 minutes per day. Users spent an additional 2% of the day in range (baseline not given, not statistically significant), as well as 14% fewer lows longer than 120 minutes per year (p<0.001) and 3.9% fewer highs longer than 120 minutes per year (p<0.001) – the latter were very strange metrics, and we’d speculate it the more traditional time <70 mg/dl and >180 mg/dl were not significantly changed (given the small bump in overall time-in-range). Of course, reductions in prolonged events more than two hours are very positive and presumably of strong interest to payers. Within a week after a low glucose “insight,” 55% of users had fewer lows, with an overall 2.4% less time spent below target; within a week after a high glucose insight, 54% of users had fewer highs, with an overall 1.7% less time spent above target – we assume the latter were absolute reductions, meaning ~30-minute daily improvements in time spent low/high after insights. These are fairly modest, but presumably the group on the app is doing quite well at baseline. On average, participants in the study used the app 2.1 times per day – indicating it is engaging and could drive good ROI from the pattern recognition. Users “liked” 86% of the nearly 4,700 insights generated in the period, and 68% were logging food (2.8 items/day, on average). This analysis was part of the limited learning launch, from which we saw preliminary data at ADA (smaller n of 81). FDA review of the Guardian Connect Mobile CGM system remains the gating factor for launch. As a reminder, Sugar.IQ has been very delayed from the initial guidance to launch before the end of 2016.
    • Dr. Vigersky emphasized that “simple, judgement-free nudges can lead to sustained behavior improvement.” Aka…“your Jewish mother in an app.” We think Sugar.IQ has done some nice things to pull more insight out of CGM data, particularly the app’s “motivational” insights. How much value will it add to Guardian Connect? Will a non-early adopter population use the app? Will it emerge as a revenue source for Medtronic over time?

  • Dr. Vigersky highlighted data from the first ~3,000 people on the 670G (see Dr. Sherr’s talk for more detail), and we’ve since learned that “about 10,000” people are now on the MiniMed 670G (a 10-fold improvement since the last update at Keystone). After Hurricane Maria hit, Medtronic conveyed via an email that patients in the 630G to 670G Priority Access Program can expect to receive pumps and transmitters in the second half of November, with Guardian 3 sensors to follow “soon” after, based on patients’ sensor re-order date. As of the August financial update, there were close to 35,000 patients enrolled in the Priority Access Program, meaning close to one-third have a system at this point. Considering FDA approval came ~13 months ago, it’s surprising how long this product has taken to trickle out. Of course, Medtronic is also the first to launch hybrid closed loop, so the caution is understandable. Medtronic has also historically not been a “CGM” company, but is now required to climb the learning curve and really scale up manufacturing quickly. Per the October email update, new 630G/670G pumps were expected to ship by the end of October, with sensor and transmitter orders expected to ship by the end of 2017/early 2018.
  • Guardian Sensor 3 accuracy has remained remarkably consistent through all phases of 670G deployment, with a MARD between 10.3% and 10.6%. The figure below shows the progression – from 10.55% MARD vs. YSI in the pivotal trial (3 calibrations/day), 10.3% MARD vs. SMBG in the continued access phase, 10.4% MARD vs. SMBG in the customer training phase, and 10.4% MARD vs. SMBG in the formal launch (the latter three are on ~3.2 calibrations/day). Dr. Vigersky called the data “very reassuring,” to know that performance in the real-world matches that seen in the clinical trial. The real Achilles heel of Guardian Sensor 3 is the calibration burden (3-4 recommended per day), which will be critical for Medtronic to manage as patients compare to Abbott’s no-calibration FreeStyle Libre and Dexcom’s upcoming G6 (no-cal or 1 cal/day). We wonder what the hold-up is at FDA for Guardian Sensor 3 – is it about human factors, manufacturing, a non-adjunctive claim, or something else?

  • Over on the pediatric side (data from the 7-13 year old 670G study), the Guardian Sensor actually performed with a slightly lower MARD of 9.9% on two calibrations per day (999 paired points). Nice! Further, 89.5% of points fell within 20%/20 mg/dl, very strong data for Medtronic. We look forward to seeing the glycemic outcomes from this study – Dr. Vigersky previously indicated in April that the data would be submitted to FDA by the end of the year. We’ll hope to hear on update on Medtronic’s upcoming 3Q17 call.

  • Patient-reported outcomes on the Guardian Sensor 3 are quite positive. In a survey of 2,231 670G users: (i) 92% trust the system to help manage their glucose levels; (ii) 76% are extremely or very satisfied with MiniMed 670G; and (iii) 75% are extremely or very satisfied with Guardian Sensor 3. As Medtronic speakers keep pushing, this news sensor line really is something of a new dawn and a door-opener for the company. Anecdotally, we’ve heard this too, including from former Dexcom users – the new Medtronic sensor really delivers relative to prior versions. Near-term, Medtronic just has to scale manufacturing! There are more than enough patients for multiple companies to convert …

3. Real-World Libre Data (>235k Readers; 2.3B glucose Readings): Scan frequency Tied to Reduced A1c & Hypo

Marking the first major US conference in which Abbott’s FreeStyle Libre (real time) could be discussed as an FDA-approved product, Dr. Tim Dunn presented new real-world data collected from September 2014-2017 (n= 237,747 readers) demonstrating strong correlations between the number of daily scans and reductions in estimated A1c and time spent in hypoglycemia (<55 mg/dl; see below). With over 237,000 readers as of September, this data set is now more than quadruple the size relative to the first glimpse in February (55,000+ readers). Notably, the curve of scan frequency vs. eA1c looked pretty much the same. Dr. Dunn also showed data recently presented at EASD supporting an inverse relationship between number of scans and time spent in hyperglycemia (>180 mg/dl), as well as interesting regional comparisons between Germany and France (see below). For time spent in hyperglycemia, the two countries exhibited minor differences: in Germany, time spent in hyperglycemia dropped by nearly 4 hours/day from a baseline of 9.9 hours/day, while in France, time spent in hyperglycemia decreased by 3.6 hours/day from a high baseline of 10.5 hours/day – either way, more scans drive less hyperglycemia. However, for time spent in hypoglycemia, time dropped 9.3 minutes/day from a low baseline of 35.9 minutes/day in Germany, while in France, time decreased by 18.6 minutes/day from a much higher baseline of 59.1 minutes/day – notably, the benefits from scanning appear to level off at <20 scans/day, which makes sense (that would be more than once per waking hour). We wonder what cultural, healthcare, or other factors could underlie these differences – this data set seems like a treasure trove and only the surface is being scratched right now. Future directions for these real-world studies include generating observations over time, observing populations with recent approvals or reimbursement (possibly Belgium or France), or most intriguingly, trying to identify strategies of Libre usage (around hypoglycemia, pre/post-meal, insulin dosing, and exercise) that beget glycemic success.

  • Dr. Dunn did not comment on the impending US launch, which is expected in retail pharmacies by the end of the year. Read more in out Abbott 3Q17 report.

  • Dr. Dunn showed accuracy data for the 10-day, 12-hour warmup US version of FreeStyle Libre (based on 5,772 paired CGM-YSI points): MARD was 9.7%, and 91% of values came within 20%/20 mg/dl. More stringently, 82% of values were within 15%/15 mg/dl. No data were presented in hypoglycemia, though we’ve discovered the label is now posted here on FDA’s website (key performance data below) – FreeStyle Libre reports a strong 81% of points within 20 mg/dl for glucose levels 51-80 mg/dl, and 58% of points for levels 40-50 mg/dl. Unlike the Pro version, the real-time US version of FreeStyle Libre (with the longer 12-hour warmup and shorter wear time) does not have a contraindication for hypoglycemia. Read the full performance data here.

  • FreeStyle Libre, which obtained FDA approval in September, is now available in 42 countries, with full or partial reimbursement in 19 of them. Here is a non-exhaustive list of markets with coverage – non-exhaustive because there have been two additions since Abbott last shared specifics! The clip at which countries are deciding to reimburse the sensor is impressive, and presumably stems from the two RCTs (in type 1 and type 2), the growing real-world data, and the appealing price point relative to fingersticks and traditional CGM.
  • During the panel discussion, Dr. Dunn mentioned that he was surprised by the high number of daily scans apparent in the real-world data – 16 times/day on average. This is indeed quite notable, considering this is nearly half of FreeStyle Libre global installed base (assuming 1 reader = 1 person, and there are now 400,000 users). As Dr. John Pickup noted, some individuals averaged 40-50 scans/day, raising the question of the psychosocial impact of CGM. Might CGM actually make some people more anxious or obsessed? Psychosocial expert Prof. Katharine Barnard stepped up to the mic to reassure that people appreciate the reassurance associated with CGM and that published data exists demonstrating specific quality of life benefits due to CGM. As she pointed out, scanning 40-50 times/day might be no different than the high frequency at which we check our phones (or may be less for some). Dr. Dunn cited Abbott’s type 1 study published in The Lancet showing improvement in quality of life associated with the Libre, but acknowledged that it is an area requiring further investigation. We suspect scan frequency will change depending on the population (early adopters vs. late majority), and the benefit of no fingersticks will be a huge quality of life driver for everyone.

4. Reimbursement for interpretation of CGM Data May Decline Slightly in 2018 – New Fee Schedule Proposes $36.72 for Code 95251

In a talk on CGM reimbursement, Dr. Allan Glass (The Endocrine Society) said that the reimbursement rate for physician interpretation of CGM data with report (95251) may decline in 2018 – and it’s already a low $44.50 through Medicare! Since the talk on Thursday, the new fee schedule has actually come out, and reimbursement for this code has indeed declined 17% to $36.72. We are extremely frustrated to hear this, since stronger reimbursement for CGM data interpretation could encourage more sensor prescribing and HCP engagement with the results. Until the AMA prints the final fee schedule, who can bill, and the limits (i.e., how many times an HCP can use a code within a time frame), it will not be final. More details are expected in the coming months, probably closer to January 2018. There are important nuances to this – e.g., when and where codes can be used – and we seriously hope the final AMA publication at least broadens the criteria of who can submit this for reimbursement. Interestingly, this reduction is based on time/motion surveys of physicians, which could mean they are getting faster at reviewing CGM data with the next-gen reports that are now available (Abbott’s Libre View, Dexcom Clarity, Medtronic CareLink). On the other hand, we hope it does not further dissuade providers to use these codes – e.g., will they bother, given the low reimbursement? If physicians are monitoring hundreds of patients remotely, could they bill each month for each patient? On a positive note, existing code 95250 for HCP-owned professional CGM (sensor placement/hookup/training) is proposed at $156.58, almost identical to the current $159.70. We’re concerned about the poor reimbursement providers continue to receive for CGM. Will this improve? Medicare definitely deserves kudos for finally reimbursing Dexcom’s G5 CGM in type 1/2 insulin users, but payment reform needs to make CGM prescribing/utilization sustainable for providers too. 

  • A new code, 95249, now exists to reimburse physicians for patient education and setup of personal real-time CGM (i.e., used by private payers for all CGMs; currently only Dexcom’s G5 under Medicare). The proposed amount (that came out Thursday night) is $56.15 as a national average – low, but obviously better than nothing. This can only take place if the patient comes into the office (i.e., no telehealth), and it can only be reported once during the entire time someone owns a CGM (i.e., no reimbursement for retraining or refreshers).
  • We also learned that CMS’ reimbursement rate for Dexcom’s G5 (~$250 per month for supplies, ~$250 per receiver) is actually based on the home glucose monitoring that CGM replaces, not on the cost of the actual CGM device itself! This did not at first glance make sense since BGM reimbursement is much lower. According to Dr. Glass, this may mean future CGMs covered by Medicare will get the same pricing (though obviously things can change). This could be a positive for the field if longer wear time or lower cost systems (e.g., Abbott’s FreeStyle Libre, Dexcom G6/Verily products) are reimbursed at the same rate. Of course, it would be even better if Medicare understood that SMBG is not equivalent to CGM.

5. Senseonics Aiming for Gen 2 Sensor that Lasts One Year & “Relies Less on User Interaction, Such as Calibration”

Senseonics is hard at work developing a second-gen sensor, “looking for use up to a year” and “relying less on user interaction, such as calibration,” according to R&D Director Dr. Andrew DeHennis. He optimistically pointed out that there is just one factor limiting the extension of duration past the recently CE-marked 180-day indication – signal degradation of the indicator that cleaves boronic acid groups via reactive oxygen species – but the team is working on mitigation solutions. If Senseonics were to succeed in developing a 365-day sensor, requiring just one clinic visit per year, its sensor would last nearly 26-times longer than the next competitor in the EU (Abbott’s 14-day FreeStyle Libre), and 36-times longer than that FreeStyle Libre in the US (as of now). Of course, this was also significantly enhance the value proposition of Senseonics’ implantable CGM, which still does require a small on-body transmitter. On the other gen two proposition of eliminating fingersticks, we were glad to hear this project is still high on the list of priorities; we haven’t heard about it in a couple quarters. Last we heard in May, this sensor was in human feasibility trials. The sensor is expected to use redundant glucose-sensing capabilities, facilitating improved accuracy, longevity, and functionality. The original 90-day Eversense with the first-gen on-body transmitter is currently under FDA review, with an Ad Comm expected in 1Q18. It is currently available in 12 EU countries and South Africa. Read more in our Senseonics 3Q17 report.

6. OSU’s Academic Work on Contact Lenses to Measure Glucose – Early Stage, But Promising Use of Display Industry Technology

Dr. Greg Herman presented early-stage work on Oregon State University’s efforts on contact lenses to measure glucose. He immediately clarified that the team’s approach is different from Verily/Novartis, who are using a more typical amperometric (electrochemical) sensor with a copper antenna (and on which there has not been an update in over a year; the original ~5-year time frame for launch (~2019) is quickly approaching!). Notably, Dr. Herman’s team is leveraging core technology from the display industry, which enables up to 3,000 sensors per square millimeter – this allows for incredibly low costs and highly automated, large-scale manufacturing. OSU’s approach uses flexible transparent, “amorphous indium gallium zinc oxide (IGZO) field effect transistors.” Glucose oxidase is immobilized on IGZO – read how it works to measure glucose (via a chemical reaction and pH changes) in the bullet below. Notably, measurements of other analytes and environmental conditions are also possible with this technology. The sensor does not have acetaminophen and ascorbic acid interference, and can reportedly measure down to much less than the tiny concentrations of glucose found in tears. The sensors are fully transparent – an advantage for contact lenses – and the flexibility allows for application to other surfaces, including catheters. (Presumably this makes a combined CGM/insulin infusion set possible.) Two big questions, of course, are whether (i) tear glucose can be reliably and accurately measured; and (ii) correlates strongly with blood glucose. Dr. Herman noted concentration as a challenge: tear glucose levels are ~10x lower than blood glucose concentrations, meaning any sensor needs to be quite accurate. In one recent publication in people without diabetes (n=10), the correlation between tear and blood glucose was notably high (r2= 0.96) – of course, doing so in an ambulatory setting in people with diabetes will be important. The OSU contact lens work is still early, and audience questions highlighted many nuances that will be key to address – e.g., what about temperature, stress, environmental conditions, changes in tears, etc.? The use of consumer electronics manufacturing seems quite promising as a general approach to drive sensor costs down. Near-term, we’ll be fascinated to see if Verily/Novartis’ efforts come to fruition.

  • How does it work? Glucose oxidase is immobilized on IGZO using self-assembled monolayers, which have aminosilane groups and glutaraldehyde. Gluconic acid is formed through the biocatalytic oxidation of glucose by glucose oxidase, and this locally decreases the pH near the IGZO surface. This results in positively charged aminosilane groups, which are electron acceptors located at the IGZO surface. Electrons tunnel into these acceptors and reduce the conductivity of the IGZO film. The change in the IGZO conductivity can be directly correlated with glucose concentrations.
  • “With most technology, there are going to be struggles. There needs to be a lot more work on how environmental conditions affect glucose concentrations (in the eyes). One of the nice things with this technology is we can incorporate other sensing technologies and measure a range of analytes. We can measure temperature and concentrations of other biomarkers. There are going to be potentially 10-100 different sensors that we can be integrating. With AI, we can start correlating that information. If you only had one sensor monitoring glucose in the eye, that would be difficult with all the environmental and physiological factors.”  

 

7. Pacific Diabetes All-In-One CGM/Infusion Set to Enter Feasibility Study (N=10); Looking for Commercial Partners “soon”

Dr. Robert Cargill shared that Pacific Diabetes Technologies’ (PDT) all-in-one CGM/infusion set is progressing nicely, with the sensor showing promising results in swine and the first generation combo device scheduled to take part in a feasibility artificial pancreas study at OHSU “within the next six months or so.” The first-generation design, which is in the “later stages” of IDE submission prep, is comprised of a flexible sensor surrounding a rigid cannula with a laminated 25-gauge needle (“a little large, but what we thought we could get into trials quickly”). Dr. Cargill expects to conduct an eight-hour in-patient feasibility study (n=10 T1 adults) in the next six months, driven by OHSU AP system and with Pacific Diabetes’ first-generation device participating as a “bystander” – presumably meaning it will not feed into the algorithm and results will be compared to a standard commercial CGM. Dr. Cargill’s team is currently focused on designing the second-generation product, which is basically turned inside-out from gen-1 – the cannula is embedded inside the sensor so that the fluid path runs down the center of a thermoplastic core. The result is a much more flexible device. The development team is working on accelerating the warmup period, as well as improving the stability and lifetime of the sensing enzyme, aiming for at least seven days. Infusion set lifetime is obviously a concern as well – Dr. Cargill shared that an infusion set with redundant ports (“sprinkler” design) is currently in development to address this issue, and he is also closely following Capillary Biomedical’s work in the area. Now, the group is moving towards developing the entire CGM-infusion set and will begin looking for commercial partners to help scale manufacturing “soon.” We like the vision of an all-in-one CGM/insulin infusion set for pumpers, though sensors are already at 10-14 days, continue to get smaller, and continue to push into MDI – can Pacific Diabetes keep pushing ahead? Would an infusion set player (e.g., Unomedical, BD) or a CGM company (Abbott, Dexcom, Medtronic) consider a partnership?

  • One of the first challenges Pacific faced was designing a sensor unbiased by the presence of insulin. When exposed to insulin, conventional, platinum-based sensors exhibit a rapid spike resembling hyperglycemia followed by eventual poisoning and loss of sensitivity. By turning to redox-mediated chemistry, Pacific Diabetes developed a sensor capable of functioning normally, showing no inappropriate glucose spike when studied in swine.

Automated Insulin Delivery

Yale’s Dr. Jennifer Sherr presented real-world MiniMed 670G data (n=3,031) from five months of voluntary uploads to CareLink: In the entire cohort, mean blood glucose dropped by ~8 mg/dl (baseline: 160 mg/dl) and time-in-range (71-180 mg/dl) increased by nearly two hours/day (from 65% to 73%). These real-world results show a consistent time-in-range improvement relative to the pivotal trial (when time in range went from 67% to 72%). Notably, there was not a meaningful change in hypoglycemia (low at baseline), though time in hyperglycemia came down by more than an hour per day. This is excellent news for Medtronic and provides further confidence in the three-month pivotal data. In the <15 year-old sub-analysis, on average, mean blood glucose decreased by a whopping ~17 mg/dl (baseline: 178 mg/dl) and time-in-range increased by two hours/day (from 55% to 66%). Adolescents spent 77% of the time in auto mode (18.5 hours/day), notably lower than the entire cohort’s 85% of the time in auto mode (20.4 hours/day). Of course, seeing this kind of utilization in 14-15 year-olds indicates the system is providing clear benefits, even in the tough teenage population. For the entire cohort, sensor MARD was 10.4% with 3.6 calibrations/day, though for adolescents, MARD was slightly higher at 11.8% with 3.5 calibrations/day. Last time we saw real-world CareLink data at Keystone, patients (n=730) had jumped from 63% in range in manual mode to 76% in range in auto mode – no matter which way the data is cut, early 670G users continue to see nice time-in-range benefits from this first-generation hybrid closed loop. We look forward to seeing data in even younger 7-13 year old users. As an aside, we also hope to see Medtronic move to the now-industry standard cutoffs: <70 and 70-180 mg/dl.

 

Entire Cohort (n=3,031)

Pediatric (<15 years old) Cohort (n=557)

 

Initial Manual Mode

After Auto Mode

Initial Manual Mode

After Auto Mode

Time below range (%, ≤70 mg/dl)

3%

2%

2%

2%

Time in Range (%, 71-180 mg/dl)

65%

73%

55%

66%

Time above range (%, >180 mg/dl)

32%

24%

43%

31%

Mean sensor glucose

160 mg/dl

152 mg/dl

178 mg/dl

161 mg/dl

  • Dr. Sherr also presented subgroup analyses of the three-month pivotal trial, examining outcomes just in 14-17 years-olds (n=20) and 18-21 years-0lds (n=10). Among children 14-17 years old, time-in-range increased by nearly two hours/day (baseline 14 hours), and in those 18-21 years old, time-in range increased by approximately one hour/day (baseline 16 hours) – not surprising, since 18-21 is a tougher age group, given college and transition to adult care. Increases in time-in-range were even more marked overnight, in line with the adult data. During the overnight period, insulin rates were dynamic: For the younger and older adolescents cohorts respectively, insulin was not delivered at all for 16% and 19% of the time, and insulin was delivered at a maximum rate 33% and 38% of the time. The transferal of variability from glucose to insulin is closed loop’s great promise, of course. See below for glucose tracings (grey represents the two-week run-in phase, pink represents the three-month study phase; top = overnight; bottom = 24 hours). The overnight glycemic variability for the older pediatric cohort is not as tight as would be expected – Dr. Sherr cited the smaller sample size as a possible explanation. The following day, Medtronic’s Dr. Vigersky mentioned that the 670G trial for kids ages 7-13 has wrapped up, and he has previously indicated data would be filed with FDA for a lower age indication before the end of the year.

 

2. Bigfoot MPC Algorithm Adjusts Parameters, Including Glucose Target, on Hour-to-Hour Basis; 2018 Pivotal with FreeStyle Libre

Bigfoot Director of Clinical Innovation Ms. Jen Block shared new details on the  automated insulin delivery system studied in this trial: the Bigfoot algorithm starts with a single basal rate, carb ratio, ​and ​insulin sensitivity factor, and then proceeds to adjust basal rates, carb ratios, insulin sensitivity factors, and glucose targets on an hour-to-hour basis. We weren’t aware the system was actually changing the glucose target by hour, which is quite impressive and presumably could enable tighter control at times where there is less risk of hypoglycemia (our speculation). The 48-hour feasibility trial (n=20) was displayed in the poster hall on Day #1, but Ms. Block reviewed it in more detail. The study was in 10 adults (no one over 57 years old), five children ages 7-12, and five adolescents ages 12-18. The study was designed to accommodate MDI users, a demonstration of the algorithm’s ability to quickly adjust therapy parameters in non-pumpers, and a continued signal from Bigfoot that it really wants to expand the market. The 65% time-in-range, 0.9% of time <70 mg/dl, and no hypoglycemia requiring assistance was roughly in line with other AID trials, especially since the study included (i) day one breakfast with ~30% excess insulin delivery; (ii) day two dinner with complete omission of insulin bolus; and (iii) 45 minutes of unannounced exercise. Ms. Block also hearkened back to the vClinic that Mr. Lane Desborough, Bryan Mazlish, and team developed to run over 100 million subject-days of simulation prior to the human trial. The vClinic could run the equivalent of a three-month pivotal trial (200 subjects, 100 days) in “less than a minute.” As noted in our Day #1 report, based on the extensive simulation and modeling, the team almost exactly predicted the results of the in-clinic evaluation. Ms. Block, who has many years of experience in diabetes clinical research (including lots of AID studies with Dr. Buckingham), conveyed how impressed she was by this work multiple times through the talk. Bigfoot certainly has incredible algorithmic brainpower, a cool design vision, and an exciting monthly subscription approach, though lots of execution will be needed to hit the ambitious timelines – a 2018 pivotal trial with the factory-calibrated next-gen FreeStyle Libre and a possible 2020 launch (pending FDA approval of the system, which includes a smartphone app serving as the complete user interface).

  • Ms. Block displayed a timeline of Bigfoot achievements so far, showcasing the recent Timesulin acquisition, Abbott FreeStyle Libre partnership, planned 2018 pivotal trial timing, and planned 2020 commercial launch. A separate slide titled “A vision – Automated Insulin Delivery for Everyone…” (below) depicted “Bigfoot Inject” (MDI auto-titration system) and “Bigfoot Loop” (automated insulin delivery system) side-by-side, both with a Bigfoot-branded next-gen Libre, the Bigfoot app running on a phone, and pens or the Asante pump, respectively. On the former, Ms. Block pointed out how young the company is – as a matter of fact, the team will be celebrating its third birthday this week. She also indicated that the company as a whole has almost 140,000 collective days of experience living with type 1 diabetes (i.e., 384 years(!), presumably including family members of employees).

3. iDCL Trial Update & Summary Table: Protocols 1-4 Beginning in 2017-2018

UVA’s Dr. Stacey Anderson provided a welcome overview on the long-in-development, multi-arm iDCL trial, detailing plans to complete protocols 1-3 of the NIH-funded project in the upcoming year, as well as a future protocol 4 to test the incremental effect of adding a next-gen adaptive control algorithm from Harvard. We appreciated the clearly-laid out cadence of trials coming down the line – boy are these groups, particularly at UVA and Jaeb, going to be busy! All of the details are included in the table below – many were clarifying or new to us. In line with Tandem’s 3Q17 call, the US pivotal for the treat-to-range system (with G6 and TypeZero) is set to start in 1H18 – this has been delayed, but it seems like the hardware pieces are now worked out and we’re glad to see CGM outcomes as the key endpoints. Protocol 2 – the EU pivotal for the long-term Roche-Senseonics-TypeZero system – will require a training protocol to test the new system configuration; we’re not sure if the training protocol is set to start in 1Q18, or the main study. The language on the slide for this Roche/Senseonics system is also notable, which says “subsequent FDA presentation” –  the focus is currently on the EU market, though this is the first definitive indication that the companies are interested in US commercialization as well. Who will actually undertake the commercialization remains up in the air (as far as we know), but we’d speculate it would be Roche as the pump partner. Lastly, the fourth protocol, beginning in 4Q18 and layering on a next-gen adaptive control algorithm (Harvard Zone MPC algorithm) is exciting and presumably offers potential for tighter time-in-range. We look forward to seeing this cohort of trials progress and hopefully bring at least two commercial systems to market. As a reminder, iDCL received a remarkable ~$12.7 million in grant funding from NIH.

Protocol #

Components

Timeline

Study Details

Notes

Protocol 1

(“To establish that a mobile closed-loop control system is a viable treatment for T1D”)

Roche Spirit Combo pump;

Dexcom G5 CGM;

TypeZero inControl AP software on smartphone and inControl Cloud;

Ascensia Contour Next One BGM

October 2017 – April 2018

3-month RCT

N=126 randomized 1:1 closed loop vs. SAP

Primary outcomes: Superiority in time <70 mg/dl, non-inferiority in time >180 mg/dl

Seven US sites, coordinated by UVA & Jaeb

10-site training study completed

FDA + IRB approvals complete

 

Protocol 2

(Pivotal trial for EU “and subsequent FDA presentation”)

 

Roche Accu-Chek Insight pump;

Eversense XL implantable 180-day CGM (Note: the slide said 90-day, but the companies have previously indicated 180);

TypeZero inControl AP algorithm

Start 1Q18

Similar procedure to protocol 1

N=72 subjects randomized 2:1 closed loop vs. SAP

Sites: Academic Medical Center (Netherlands); University of Montpellier; University of Padova; & Jaeb

Training protocol first to test new system configuration

Presumably a 3-month study, though the Eversense XL sensor can last 6 months.

Protocol 3

(US pivotal trial for Tandem/Dexcom embedded system – launch expected in 1H19)

 

Tandem t:slim X2 pump;

Dexcom G6 CGM;

inControl IQ AP algorithm embedded in pump;

Ascensia Contour Next One BGM

Start 1H18

Similar procedure to protocol 1

N=147 subjects randomized 2:1 closed loop vs. SAP

Primary outcomes: Superiority in time <70 mg/dl, non-inferiority in time >180 mg/dl

Secondary outcomes: A1c & technology acceptance

Seven US sites, coordinated by Jaeb & UVA

 

Intended as 3rd major software update for t:slim X2 pumpers (following G5 integration and PLGS)

Training protocol first to test new technology (Dexcom G6, X2 pump with inControl)

Protocol 4
(“Testing the incremental effect of a next-gen adaptive control algorithm that builds on the Zone MPC paradigm developed at Harvard”)

??

4Q18 (planned)

Coordination: Jaeb & Harvard

 

  • Included in Dr. Anderson’s slides were images of the Eversense CGM with the inControl AP and inControl IQ embedded in the t:slim X2 pump (with a G6 CGM). See below.

  • Dr. Anderson also reviewed a litany of previous studies from the UVA group, including data from home use (Diabetes Care 2016; DTT 2017), showing ~77% time in range, 1.3% time below 70 mg/dl, and 0.1% time below 50 mg/dl. She also showed closed loop studies in extreme conditions (adolescents skiing – elevation, activity, low temperature – and those with extreme hypoglycemia).

4. Outpatient Data on Insulet’s Omnipod Horizon: 74% TIR, 1.8% time <70 (n=11), mean Glucose of 150 mg/dl

Insulet’ Dr. Trang Ly shared interim outcomes from the ongoing five-day hotel study of the Omnipod Horizon Automated Glucose Control system. She summarized data from 11 patients so far, with a strong 74% time-in-range (70-180 mg/dl), 1.8% <70 mg/dl, and 25% >180 mg/dl. (As a reminder, this is the first study of the system in more free-living conditions.) Mean glucose was a solid 150 mg/dl. See the comparison below to the previous IDE trials. As she did at EASD, Dr. Ly showed the same n=1 example from a pediatric user, who entered with a mean A1c of 9.8%; on day #4, mean glucose had dropped to ~144 mg/dl, predicting an estimated A1c drop to ~6.7% (82% time-in-range). Dr. Ly did not give any launch or pivotal timing on Horizon, though an unmarked timeline slide (see below) confirmed that after this hotel study is complete, a pre-pivotal study will follow, with a pivotal as the last clinical development step. Insulet’s 3Q17 call expected a lunch of Horizon by “end of 2019” or “early 2020,” which feels quite reasonable given Insulet’s impressive pace of study progress so far – n=118 participants tested, n=7,584 hours of closed loop.

  • It will be fascinating to watch Insulet and Bigfoot progress side-by-side, as both have some similarities and some key differences. Both are taking a compelling hardware approach to automated insulin delivery (i.e., their on-body components do not have a user interface, and their systems will remain in closed loop even if the PDM (Insulet) / smartphone (Bigfoot) are out of range). Both are also really prioritizing an outstanding, consumer-grade user experience. Thus far, Bigfoot has focused its pre-pivotal efforts on modeling and simulation (100 million subject-days), with only one completed clinical feasibility study to date (see highlight above). Insulet is doing more rounds of clinical testing, in addition to simulation/modeling with the UVA/Padova simulator and its own data. Both algorithms are MPC-based, though with different flavors – Insulet’s algorithm was licensed from UCSB after years of academic development (but has since had substantial changes to it), while Bigfoot’s algorithm is developed in-house and leverages machine learning (see above). We expect both to be very strong products and both to come to market around roughly the same time, assuming current launch projections are hit.

Smart Pens, Insulin Dose Titration, Clinical Decision Support, BGM

1. Novo Nordisk Supplying NFC-Enabled NovoPen Echo & Tresiba Pens for NIH Study of TypeZero MDI Advisor

We learned that Novo Nordisk is supplying the NFC-enabled NovoPen Echo and a connected Tresiba pen for the ongoing three-month NIH-funded study (estimated n=132 type 1s ≥15 years old) of TypeZero’s inControl MDI Advisor at UVA, Stanford, and Mt. Sinai (Drs. Marc Breton, Stacey Anderson, Boris Kovatchev, Carol Levy, and Bruce Buckingham). Dexcom is also supplying CGM. We were aware that the NovoLog pen was piloting in limited fashion in Sweden, but this was the first that we can recall hearing (i) it is being used in the US; and (ii) there is also a connected basal insulin pen. We wonder if both pens are being used in other studies/pilots around the world, and whether there are near-term plans to add Bluetooth or submit to regulators in Europe or the US. Novo Nordisk is clearly deeply committed to connectivity and digital health, a point confirmed in our recent interview with CEO Mr. Lars Jørgensen and in a futuristic video at the HITLAB symposium a couple weeks ago. The control arm of the NIH study, expected to conclude in late May 2018, is CGM alone, and the primary outcome is time-in-range (70-180 during the day; 80-140 at night). As a reminder, the smartphone-based MDI Advisor (interface below) provides “smart” bolus advice, as well as exercise advice, sleep advice, hypoglycemia prediction, and eA1c (we’re not sure if it adds basal insulin titration). Dr. Buckingham displayed results from a small inControl pilot study in pumpers and MDI on Day #1.

  • Dr. Boris Kovatchev also shared impressive data on his group’s dynamic, CGM-based hypoglycemia prediction. In one clinical evaluation, the algorithm classified future hypoglycemia risk in stoplight fashion (red=high, yellow=mid, green=low) – in the following hour, those with a green light at baseline had no glucose values <70 mg/dl, those with yellow had very little (<1% of time) <70 mg/dl, and those who had received a red alert spent nearly 25% of the time <70 mg/dl in the next hour, as well as ~13% at <60 mg/dl and 5%+ at <50 mg/dl. The power of this model to predict and allow the prevention of hypoglycemia is evident, and we love the simple green, red, yellow light interface – we’re not sure if that’s just for the trial/data presentation, or if it is also the way risk is conveyed to the user. Notably, the algorithm can also take into account other data, such as activity, presumably improving predictive power. We’re not clear on the commercialization pathway here – perhaps Dexcom would launch this as a decision support app for MDI users (see talk above from Peter Simpson).
    • The UVA team has also been working on SMBG-based decision support system, which relies on a model (“knowledge”) to fill in data gaps, for a number of years. We wonder how this will stack up to stronger predictive value from a CGM-based system.

  • Dr. Kovatchev and team’s >20-year diabetes portfolio at UVA has amounted to 58 issued patents, 109 invention disclosures, and 309 applications of registered copyrights. Wow! Can TypeZero (either alone or with partners) execute and successfully commercialize some of this IP? So far, TypeZero has announced several partnerships and longer-term studies, but launches are still pending.

2. Dreamed Clinician Advisor Pivotal Beginning This Month; Pilot #2 Data

Schneider Children’s Prof. Moshe Phillip announced that the NextDREAM Consortium’s  six-month, multi-site pivotal trial for Dreamed’s clinician advisor (decision support) has begun as of this month, a slight delay from previous plans to begin in September. Prof. Phillip also shared data from the MD-Logic Advise4U Pilot 2 study (n=13) demonstrating that the Advisor once again results in comparable glucose control relative to expert care (we covered the first pilot data back in February at ATTD). In the study, patients were instructed by doctors – either based on their own clinical intuition or the Advisor’s recommendations – through two cycles of pump settings adjustments over the course of three months. There were no significant differences in time within range (70-180 mg/dl) between groups after three months, a good sign that software is performing as well as expert clinicians. Prof. Phillip did note a slight tendency for the Advisor to more effectively protect patients from hypoglycemia (see below). Both the control and Advisor groups included children, adolescents, and adults, and exhibited similar metabolic control at the study’s start (baseline A1c = 7.8%). The pilot data is very promising (non-inferiority in the bar, in our view), and we look forward to seeing the pivotal trial and ultimately a commercialized product. As a reminder, the pivotal trial will investigate the Advisor integrated within the Glooko platform, allowing providers to view the patient’s current therapy alongside suggested adjustments, including basal rates, carb ratio, correction factor, and active insulin time for pumpers. The Advisor also describes detected patterns related to patient behavior and insulin dosing using easily understood language. In his presentation, Dr. Phillip highlighted the “edit” feature on the Glooko platform, which facilitates physician alterations to the Advisor’s recommendations. As Prof. Phillip typically puts it, the Advisor is meant to be just another member of the care team, equivalent to knocking on the door of a colleague in search of a helpful tip for managing a patient. We love that positioning!

3. Topline Glooko MIDS Feasibility Results: Avg BG Drops 18 mg/dl, Readings in Range Up 9%; N=240 Study Underway

A poster presented more detailed findings from Glooko’s 14-day feasibility study (n=14 type 2s) showing use of the Mobile Insulin Dosing System (MIDS; a basal titration app) drives significant decreases in average blood glucose levels, as well as increased time-in-range and decreased time in hyperglycemia. Average blood glucose dropped 18 mg/dl (baseline 164 mg/dl), the proportion of in-range readings (80-180 mg/dl) increased by nine percentage points (baseline 64%), and the proportion of hyperglycemic readings (>250 mg/dl) decreased by 11 percentage points (baseline 14%), without a significant increase in the rate of hypoglycemic events. For all three metrics, variability decreased too. The rate of hypoglycemia readings (<70 mg/dl) did not differ significantly between the before and during study periods – it looked to be ~3% in both phases, though this was not quantified on the poster. (As we noted last week when the press release came out, this basal-only group with above-target average glucose was probably not seeing much hypoglycemia, hence little room for improvement.) Changes in in-range readings were positively correlated with titration cycle adherence, a great indication that the titration algorithm is efficacious. Fingerstick frequency was not reported, though we assume it was ~1-2 times per day. All participants were prescribed a personalized long-acting insulin (LAI) dosage treatment plan, ranging in complexity with titration periods varying from one to four days. At baseline, all of the people in the study were already on long acting insulin, so the results are an encouraging sign that adding MIDS brings additional benefits.

  • The poster noted that a larger clinical study investigating the system for a longer duration is currently underway (recruiting per CT.gov). The study aims to randomize 240 type 2 patients to usual care (STEP WISE degludec algorithm) or MIDS, with a primary outcome of A1c change at 16 weeks and an expected completion date of November 2018. We learned last month at the Novo Nordisk-sponsored HITLAB Symposium that MIDS is still under FDA review following submission in mid-May – implying there has been some back-and-forth with the Agency, presumably on human factors. We’re looking forward to this product coming to market, perhaps with the larger marketing push of partner Novo Nordisk!

4. Verily’s Dr. Zisser Talks Current + Next Steps in AI Retinopathy Screening

Verily Diabetes Lead Dr. Howard Zisser provided an in-depth look at the company’s work in automated diabetic retinopathy detection, sharing that there are ongoing clinical trials in India. Notably, these studies go beyond the sensitivity/specificity of the deep learning algorithm (which is already strong, according to a ~year-old JAMA paper), but also on clinic workflow. Last week, we learned from Verily’s website that this is actually in partnership with imaging giant Nikon. Regarding technology vs. clinic workflow, Dr. Zisser said, “If you try to solve a problem without looking at the whole system, then you can be unsuccessful. We’re looking at ways to design image capture and workflow in the office. Today, many clinicians say they’re not interested because it’ll take up too much time.” It’s great to hear Verily is doing the necessary in-depth field work to ensure adoption success – how long does it take to get an image? Can you train a non-specialist to get an image? If you get a high-quality image, can you diagnose it quickly? – and not simply unleashing a powerful technology. The need is indeed dire: According to Dr. Zisser, India alone has a shortage of 127,000 (!) eye doctors, and 45% of patients suffer from vision loss before they are diagnosed. We’re excited to see Verily, IBM, and perhaps iDx launch products, particularly in nations with these types of dire shortages. In a similar vein, Verily is currently working on developing a screening device that is affordable, easy to use, not too big, but effective. In Q&A, Dr. Bruce Buckingham inquired if iPhone images were of sufficient quality – Dr. Zisser said, “I don’t know, but I’ll check,” and Dr. Buckingham said he would send a collection of images from Stanford to Verily for investigative purposes. It would be remarkable if these proved reliable for detection, as this would lower the cost and barriers to adoption even further. Then again, how would you ensure consistency in image quality? There might have to be an attachment that, for example, blocks out light and standardizes camera orientation and position. Also coming down the pike are algorithm feature updates (“Show me where the retinopathy is”) and expanded data sets to account for variation amongst different populations. Outside of the retinopathy realm, Verily is also undertaking deep learning studies in radiology and pathology. More broadly, we were glad to see Verily presenting separately at this year’s DTM, reinforcing its commitment to diabetes.

  • Dr. Zisser displayed images of ARDA (automated retinal disease assessment), which conveys intensity and grading scores in a few seconds. We found a better-quality image online (below) than what we snapped during the session. The severity-indicating “diabetic retinopathy grade” graph at the bottom left as great to see, as an IBM researcher previously told us that Google/Verily’s algorithm did not have this. Along with IBM Watson, this means Verily’s algorithm now offers granular classification of stages of diabetic retinopathy, a nice way to track progression.

  • We were also blown away by the algorithm’s ability to highlight problem areas in the images. In the below retinal fundus, moderately afflicted with retinopathy, the algorithm designated areas of damage in green. Dr. Zisser pointed out that it completely ignored the speck of dust in the center, something a human’s eyes would naturally be drawn to. 

  • One analysis of over 30 ophthalmologists suggested a lot of inconsistency: within-grader consistency came out to ~65%, and between-grader consistency was only ~60%. In other words, even when trained clinicians are available, they are likely to be less reliable than algorithms available today, and certainly take much longer. Explained Dr. Zisser, “Thinking about how we as physicians were taught to recognize and diagnose things, it’s based on criteria – so many hemorrhages, so many fields. It’s straightforward, but is it the best way? Machines look at edge patterns, blocks of colors. It’s based on annotated material, so higher quality input equals higher quality output.” To those who argue it’s a “black box,” Dr. Zisser says “not really, you just have to ask the right questions of the right algorithms.” Even assuming that humans are adequate screeners, patients are not necessarily screened annually (even in the US), and even if they are, the proper follow-up isn’t always seen through.
  • Tidepool VP Mr. Brandon Arbiter asked in Q&A where else in diabetes deep learning can be applied, to which Dr. Zisser replied: “We have to look in our menu of algorithms, and ask if there’s something that can match a given problem. For image recognition, the answer is yes – yes, yes, yes. For the diabetes space, we’re just starting to capture data. And I think a lot of the data we capture is incomplete – may ask a question and look at data set – we had CGM and activity data, but no insulin. There’s data missing. CareLink, you [Tidepool], Glooko – you have a lot of data. We want it to be high quality in format you can use. The main point is to ask the right questions, and design how you’re going to capture it.

5. DTS Awaits FDA Response on BGM Surveillance Data; Repeated Study is Unlikely If No Policy Action Taken

Following Dr. Michael Kohn’s presentation of the illuminating DTS BGM Surveillance data (read our full report here), the Q&A session revealed that DTS is waiting for movement from the FDA before deciding whether to repeat the study. Dr. Klonoff suggested that the Agency’s response would ideally come in the form of punishing/inspecting poorly-performing meters. It’s difficult for the FDA to take an already-cleared device off the market, though perhaps there is also potential to influence payers or even Congress. Dr. Klonoff said he doesn’t necessarily expect immediate FDA action, but in response to Adam’s question (“Will you repeat this on an ongoing basis?”), he clarified, “if nothing has happened in a year or two, I don’t think I’d want to do another study.” While the study’s value is strong – showing that 12 of 18 meters did not pass accuracy criteria – Dr. Klonoff noted it was incredibly demanding and required four years to complete. Ultimately, the study was conducted not just out of academic interest, but also to affect public policy. If the FDA doesn’t/can’t take action, the data, regardless of its objective value, will not live up to its full potential to improve patient safety. Dr. Klonoff believes the study does “deserve to be done again,” especially given the landscape’s dynamic nature – a couple of the meters included in the study are actually off the market now. Dr. Kohn agreed, claiming that repeating the study annually or every other year would be optimal, but that, “if you don’t take action, there’s no point in measuring the issue again.” We agree with this analysis! Dr. Alberto Gutierrez, who just recently retired from his post at the FDA, mentioned that the Agency is interested in looking into the DTS study, but cautioned that taking already-cleared products off the market is a lengthy and difficult process. We too hope to see the FDA or at least payers (especially Medicare!) take some action to punish inaccurate meters – it is our view that if a product wouldn’t get approved today, but was approved years ago when standards were lower and options were fewer, then it has no business being on the market.

6. 8,000 Suggestic App Users; Food Personalization and Insight for More Precision Eating

Mr. Victor Chapela, CEO of Suggestic, highlighted three key features he believes are critical for “precision eating” or “food as medicine”): (i) personalization; (ii) actionable insight; and (iii) know-how. The Suggestic app, which recently won the Health 2.0 Launch competition and leverages machine learning, aims to help people eat their diet of choice – see screenshots below. Still in beta, Suggestic currently has 8,000 users. Mr. Chapela noted that there is already lots of knowledge on how to eat well; however, there’s not enough “how to” and a community-centric marketplace for people to “shop around” for nutritional plans. Mr. Chapela hopes to change this with better personalization. Suggestic provides a variety of diets (e.g., paleo, low carb, Mediterranean, DASH, ADA), allows users to read reviews, find foods to avoid, recommends recipes, helps with ordering in restaurants, and even overlays augmented reality on a restaurant menu – very cool! The company also has some promising partners that might help personalize dietary choices based on genetics or microbiome (23andMe, Helix and Pathway Genomics). In the background, Suggestic creates hidden chains of activity, linking users’ food choices with their health goals like weight loss or A1c reductions. The app learns which sequences of activities are most likely to lead to achievement of user goals. Deep neural networks are utilized to filter for optimal ingredients and nutritional recommendations. These algorithms (the know-how) will also be made publically available, a nice open ecosystem move. Things are still early, but many seemed impressed with the app – will users find enough value to take it out multiple times a day to determine if food is ok to eat and then log it? (This remains difficult, even for highly motivated individuals.) How will Suggestic keep people engaged long-term? Or better yet, how long using Suggestic would it take to train and motivate someone to “self-manage” food as medication? For more details on the app, read our profile from Health 2.0 here.

  

Insulin, Insulin Delivery, and Insulin Monitoring

1. Lutz Heinemann Shares Early Concerning Data: Highly Variable Insulin Concentrations in U100 Regular+NPH Vials Procured at Pharmacies

Prompted by the French-led inspection of Biocon, Dr. Lutz Heinemann generated a stir by calling into the question the quality of insulin that end users are actually injecting: He presented preliminary results from a basic mass spectrometry study in which colleague Dr. Alan Carter et al. found that only one of 18 U100 regular and NPH samples procured at different pharmacies approached a concentration of 100 units per ml – the others were much lower (see figure below). All of the nine regular insulin samples hovered between just 13.9-28.7 units per ml (suggesting someone could be injecting 1/8 as much insulin as he/she thought!) while the NPH samples were much more variable (equally concerning), ranging from 35.1-94.2 U/ml. FDA requires that the concentration of U100 insulin is at least 95 U/ml before it leaves the manufacturer, and Dr. Heinemann doesn’t doubt that the manufacturers are compliant. Rather, his intuition is that the cold supply chain may not working as expected – at some point during the vial’s journey from manufacturer cold room to pharmacy, something may be going awry. Dr. David Rodbard suggested that perhaps the vials get “beat up” during transit, which could cause degradation. Whatever the reason, Dr. Heinemann called for a deeper analysis into the issue (he also acknowledged it could be measurement error), causing meeting organizer Dr. David Klonoff to wonder – is there a need for an ongoing Insulin Post-Market Surveillance Program (similar to the recently reported BGM Surveillance Program)?

  • The study’s early findings caused a lot of hallway chatter in the subsequent coffee break – mostly in the form of shock (Dr. David Kerr referred to it as “the scariest data of the meeting,” but there was also a healthy dose of skepticism floating around. People called into question the group’s analytical methods, sample preparations, vial handling, etc. Dr. Heinemann, aware of the possible implications of this study, was also healthily skeptical: “I encourage other people to repeat such measurements in order to see – have we done anything wrong? Probably, I don’t know. You’re always skeptical about yourself when you do this the first time. The best explanation we see is the need to look into the details of the cold chain. [Dr. Carter] and I would be happy to have more brains working on this.”
  • At the end of the day, insulin vial quality variability gets heaped on to the list of factors that contribute to uncertainty in diabetes therapy. Is this a real issue? We’ll be interested to follow this at subsequent meetings, and we wonder if the data below also applies to analog insulin. Presentations on continuous insulin monitoring later in the day also made us wonder about the possible benefits of quantifying circulating insulin levels, and therefore characterizing insulin absorption and quality status – a metric Close Concerns hereby coins “IAQS” (rhymes with “kayaks”).

2. Dr. Zhen Gu Introduces New Lab Projects – Smart Glucagon Patch, Artificial Beta Cell, More

UNC/NC State’s Dr. Zhen Gu introduced a number of new projects his lab is working on (including a smart glucagon patch for hypoglycemia prevention and an artificial beta cell) and presented new data on existing projects (red blood cell-tethered insulin). Dr. Gu is best known for his early-stage work on developing a smart insulin patch. What’s next for his lab? (i) Optimize smart insulin patch with stability, carrying capacity, response speed, and biocompatibility; and (ii) test the smart insulin patch on pigs with CGM. It seems as if every time we hear him talk, Dr. Gu has another host of ideas to share – and it usually comes with preliminary data. One problem, as UCSF’s Dr. Gerald Grodsky pointed out, is that the work is high potential, but there have yet to be studies in larger animals. Dr. Grodsky is concerned about sensitivity to falling glucose levels. Dr. Gu responded that he has been running a larger animal study “for a couple months with encouraging preliminary data, but it needs further validation and optimization.”

  • The other patches under development are: (i) a smart glucagon patch for hypoglycemia prevention (glucagon is released from the microneedle array when insulin concentrations get too high; see below) – hopefully the team can learn from Zosano’s failures with microneedle patch-delivered glucagon; (ii) a smart cell patch, in which external beta cell capsules which rest on top of the microneedle patch; picture); (iii) an “anti-obesity” patch, which aims to convert white fat into brown fat and ideally improve diabetes too. As a reminder, the transcutaneous microneedles are 600 um-800 um long (not long enough to reach nerves, so essentially pain-free; see pictures here).

  • Recently, Dr. Gu has been developing synthetic artificial beta cells, where he modifies a signaling process: When glucose comes in through a channel, pH is locally reduced, causing polymers on the surface of insulin-containing vesicles to unbind, exposing other proteins on the surface that allows the insulin-containing vesicles to fuse with the “cell” membrane, spilling their contents outside. Dr. Gu displayed preclinical data (below) showing that the fully-intact artificial beta cell induces a rapid fall in blood glucose that lasts five hours before ramping back up (presumably in mice treated with STZ). Work on this project was published in Nature Chemical Biology last month, with Dr. John Buse listed as a co-author.

  • Dr. Gu has also published his first study of “hijacking” red blood cells for smart insulin delivery. In a mouse with STZ-induced type 1 diabetes, on the left below, the red blood cell-bound insulin is sufficient to lower blood glucose for a longer period of time than glycated insulin alone or insulin and red blood cells injected together; and on the right, it is clear that the method yields glucose-responsive insulin release. We’d love to know what the next steps are in this study. This work was published in Advanced Materials in March, again with Dr. Buse as a co-author.
    • The basic principle, as we understand it, is that red blood cells are isolated from a donor and linked to a glucosamine-insulin complex. After being injected intravenously back into the donor, the glucosamine-insulin complex dissociates from the blood cell and binds GLUT receptors. In hyperglycemic environments, the insulin then leaves the bound glucose molecule to complex with circulating glucose, where it can then lower blood glucose on a systemic level.

3. Diasome Prandial Insulin Slated for Late 2021 FDA Approval; Basal in Phase 2 Into 2020

Diasome’s Dr. Doug Muchmore shared a slightly delayed timeline for the company’s three ongoing phase 2 trials of its Hepatocyte Directed Vesicle (HDV) insulin lispro – with phase 3 in 2019-2020, an NDA filing in 2021, and possible approval in late 2021. He also gave timing on two phase 2 studies of an HDV basal insulin: a dosing study in late 2018 (ending in early 2019), paving the way for a phase 2b (ISLE-2) study to begin in 2019 and wrap up in late 2020. Dr. Muchmore noted that there is no CVOT requirement for either insulin, as it stands. Both are delivered subcutaneously. As a reminder, Diasome’s technology consists of nano-vesicles added to insulins to target insulin to the liver. There was substantial chatter at this meeting about the benefits of supplying insulin to the liver, better mimicking physiology and overcoming some of the big drawbacks of peripheral delivery. Will Diasome be the first to bring something to market? Will it seek a larger insulin player as a partner for phase 3 or commercialization?

4. Aptitude Developing Fingerstick and Continuous Insulin Monitoring; Dr. Kerr on Why Continous Insulin Monitoring Would Be Clinically Beneficial

In a new-to-us area of diabetes tech, Santa Barbara-based Aptitude is developing a product to monitor insulin levels in the blood, both as a fingerstick and as a continuous sensor. Aptitude is a spinout from UCSB and uses next-generation “aptamers” immobilized on an electrochemical sensor. When insulin binds to the specifically tailored sensor, it causes the “aptamer” (a synthetic antibody) to change shape. The product has shown feasibility with a chemotherapy drug, and the team has done a lot of work on measuring insulin. Near-term it is developing the INSIGHT “personal insulin meter” – an episodic insulin sensor that exploits a high performance aptamer (see picture below). Just like a BGM, a fingerstick measurement would allow someone to measure insulin concentrations on the spot – test time is a fairly lengthy “<2 minutes,” though we assume this could be improved. Dr. Ferguson noted the system is low cost, compatible with established glucose monitoring technology (combination with strips?), and can currently differentiate between prandial and basal insulin analogs – the current version can only measure Novolog, Humalog, and Apidra, though another aptamer could be developed for basal insulin analogs. He showed a prototype of the fingerstick meter, and work is ongoing to complete a manufacturing-ready prototype, optimize the sensor, and deploy a development kit in clinical research settings. Notably, Yale’s Dr. Eda Cengiz is a clinical collaborator, where work will proceed to demonstrate a use case – utilizing patient-specific insulin pharmacokinetic measures (we assume relevant for closing the loop). The acknowledgment slides also listed Drs. David Kerr and Howard Zisser. We’ll continue to watch this early technology – as Dr. Kerr’s enthusiastic commentary and Aptitude’s slide notes (below), there is some interesting clinical value in insulin monitoring for closing the loop/dose titration, better understanding how insulin is working, detecting/monitoring hyperinsulinemia in type 2 diabetes, etc. The bigger question is whether the additional value will be worth paying for, given the current state of underpenetrated CGM use and shrinking BGM margins.

  • Sansum’s Dr. David Kerr also provided a concise list of reasons why continuous insulin monitoring might be beneficial:
    • Add value to technologies (e.g. closed loop; smart pen/CGM-driven insulin titration)
    • Help figure out why CSII and MDI have different outcomes
    • Known when to intensify to insulin therapy
    • Help overcome clinical inertia
    • Make clinician’s job easier in figuring out why real-world efficacy isn’t as good as it should be
    • Ameliorate some of the psychological consequences associated with diabetes
    • When to start insulin in type 2 diabetes
    • Optimizing once-daily basal insulin
    • “If I could measure insulin, we could wind back and have clinical discussions with the individual about determinants of health and ways of dealing with it”
    • Precision medicine (e.g. Why do different races have different outcomes?)
    • Prediction: Knowing insulin levels will allow us to use less insulin (Joslin medalists tend to use less insulin – safer?)
  • Aptitude’s slide deck shared a similar list of potential clinical indications for insulin monitoring:

 

-- by Brian Levine, Maeve Serino, Adam Brown, and Kelly Close