Hunting the Deceitful Turkey: Towards glucose predictions & Non-invasive monitoring

Chasing an elusive dream

At a close friend’s Christmas get together, we played a game where you share your dream, and the one which is the most outrageous is declared a winner. I thought I’ll share a professional one with the group, and in my head, pondered, it’s about time I wrote about this. For context I need to go back to my Master’s days. Professor Soumyo Mukherji from IIT Bombay told us in one of our first classes of Medical Sensors – if you want to become a billionaire in the healthcare space, either find a cure for baldness or solve for non-invasive glucose. The latter stuck with me.

Now like me, if you’ve ever followed the saga of glucose monitoring, you know it’s been a bit like chasing a deceitful turkey—just when you think you’re about to capture it, it slips away, leaving you with nothing but a story and a trail of false starts. Over the past few decades, researchers, entrepreneurs, and technologists have passionately pursued this “holy grail” of diabetes care. Which is why, inspired by Mark Twain’s book, John L. Smith’s tribute to the topic of non-invasive glucose monitoring is called “Hunt for the Deceitful Turkey - The pursuit of Noninvasive glucose”. Fascinating read on the human ambition in glucose sensing. 11 years ago, the first entrepreneur I worked with – Abhishek Sen at Biosense steered me in the direction of the book.

Last week, at Fitterfly we were awarded our first patent for continuous glucose post prandial prediction using AI and our proprietary Fitterfly PGR score. We’ve come a long way, as developers in this space. It’s unfortunate that John stopped updating his book. I believe it’s time to renew that hunt, armed with modern sensors, advanced forecasting, and of course a little AI sprinkled on top of it all.

From Bluetooth glucometers to modern sensors

Circa 2015-16, I had the joy of building one of India’s first Bluetooth-enabled glucometers. Back then, it was all about bridging the gap between a clunky finger-prick test and the promise of seamless, wireless data transfer. Aman, one of the co-founders of Biosense had designed a beautiful, all inclusive shell which had the lancing device, a drawer for lancets, an integrated bottle and the meter. All of it, form-factor-wise, was smaller than the iPhone 3g. The kicker was BLE, we were transmitting data from the meter to a phone. Data coming to the phone opens up a lot of avenues of what you can offer to the person living with diabetes. From the Ames reflectance meter to the Glucowatch, to gold and palladium based glucose strips and our very own Sync glucometer with carbon electrode strips. A lot has happened in Point of Care glucose sensing. No surprise since it’s a $3-5 billion market.

The Flash/ CGM glucose monitoring sensor, though, was probably the crown jewel. Abbot, Dexcom and Medtronic and others had 14 day wearable sensors that sample interstitial fluid, measuring glucose in 5 or 15 minute intervals. We went from random sporadic data to continuous and contiguous data. If you see someone with a white patch on the back of their arm, take a moment to marvel how that tiny device is transmitting glucose readings. The CGM market standalone is at $4-6 billion.

Fast forward to today, we also have off-label use of these devices. A person living with diabetes, especially a type 1, without glucose readings, not knowing how much to titrate their insulin, is a pilot flying blind (Some pilots can still land, albeit, with instrumentation). But a longevity-bro with HbA1c below 5 wearing a CGM sensor feels like a lot of over-engineering. Don’t get me wrong, I love Time in Range of glucose as a marker, but maybe not everyone can optimise at that level, with probably, most people ending up with analysis-paralysis unless they are data nerds.

Time & Attention is all you need

Since the 2017 paper, language processing and computer vision tasks have had non-linear advances using the multi-head attention approach. Even audio and video! Many groups have tried and have made some strides beyond LSTMs using similar approaches for the time series problem. But I fear prediction in time has not garnered the same love by the open source or academic community or even industry. Maybe for forecasting stock prices but not all use cases. There’s some cool approaches using  TTMs and TimesFM and other modified Transformers. Still, I feel time series forecasting needs its own “Attention is all you need” moment.

I lean on time, since we've gone from cross sectional finger prick measurements to CGM biosignals making this a temporal forecasting problem. Of course we would love if the sampling frequency goes down even further. Till then, one can dream. For everything else, there’s always XGBoost.

And we shouldn’t focus on the Non-invasive bit as much. The right model. The right sensor. The right use case. I feel we’re almost there. For me, non-invasive predicted continuous glucose as a SaMD that is clinically acceptable doesn’t seem that far (Note the stress on clinically acceptable - MARD <10%). By leveraging the latest in sensor technology and AI, we can finally give patients the gift of continuous, effortless monitoring that supports better management, improved outcomes, and a higher quality of life. From those early Bluetooth glucometers to the marvels of modern sensor tech and AI-driven forecasting—we are closer than ever to turning the dream of noninvasive glucose monitoring into reality. We owe it to every diabetic to deliver an experience that is pain-free, effortless, and unobtrusively integrated into daily life.

I’m writing from a place of privilege to have passion and profession intersecting at Fitterfly with Arbinder & Shailesh. With Fitterfly PGR™ - we help diabetics interpret CGM graphs into a single number that demystifies post prandial glucose. With vCGM, once the 14 day period of a CGM sensor is over, we help diabetics make low GI choices by predicting post prandial spikes based on meal choices using time series forecasting. We’re doing our bit by improving outcomes of people living with chronic non-communicable diseases with a bias for outcome focussed, patient centric efforts in digital therapeutics.