Abstract Body

The advent of big data and artificial intelligence opens new perspectives for diabetes mellitus monitoring and management. Yet, complexity and uniqueness of the human body hardly allow to define a general statistical approach able to accurately predict blood glucose variations in all patients. Our technology, MIND, is a personalized, patient-specific blood glucose level (BGL) prediction service integrated in a diabetes management platform. For each patient, a machine Learning model is created using his/her historical data, including BGL and insulin inputs collected through wide-spread devices and meals recorded by the patient when available.

The CDDIAB study conducted in 2018 demonstrated that our prediction technology is accurate enough to allow safe therapeutical decisions ; an extension of this study conducted in 2019 showed that accurate BGL predictions drive better decision making on treatment options than patient alone.

The positive outcomes of this study triggered new research cases. First, the technology is tested on another patient cohort to assess its robustness; then, the confidence in the prediction provided is studied. The second point is essential for large-scale industrialization. Hence the study of a confidence index and an envelope curve, to provide visual insights of the accuracy and an additional security on the predictions. Based on our results, the next challenge is to predict accurate bolus doses given the historical data and the predicted BGL. Another challenge is to build a system able to detect and reconstruct meals, thus meal management would be streamlined by avoiding manual inputs.

Abstract Body

The Bio-inspired Artificial Pancreas (BiAP) is an advanced hybrid closed-loop insulin delivery system based on mathematical modelling of the pancreatic beta-cell physiology. BiAP incorporates an innovative adaptive meal-insulin bolus calculator which uses artificial intelligence to provide adaptive and individualised mealtime insulin dosing by learning from past post-prandial glycaemic outcomes, user behaviour, and controller’s functioning.

The BiAP control algorithm is designed for embedded low-power solutions. It has been implemented in a dedicated microchip-based handheld device and, more recently, in an iPhone connected to a Dexcom G6 continuous glucose sensor, a Tandem t:slim AP insulin pump, and a dedicated remote web-based platform. BiAP has been successfully assessed in-clinic and an ambulatory crossover randomised controlled trial is planned to evaluate its longer-term clinical effectiveness.

In this talk, we review the latest algorithmic and software developments within BiAP, as well as the results of a realistic in-silico head-to-head comparison of the BiAP controller and an open-source do-it-yourself (DIY) artificial pancreas controller. Finally, the latest developments on enhancing BiAP with a machine learning-based meal detection algorithm will be introduced.

Abstract Body

Diabetes-related retinal conditions can be detected by using a fundoscope or fundus camera to examine the posterior part of the eye. By contrast, examining or imaging the anterior part of the eye can reveal conditions affecting the front of the eye, such as changes to the eyelids, cornea, or crystalline lens. In this work, we studied whether external photographs of the front of the eye can reveal insights into both diabetic retinal diseases and blood glucose control. We developed a deep learning system (DLS) using external eye photographs of 145,832 patients with diabetes from 301 diabetic retinopathy (DR) screening sites in one US state, and evaluated the DLS on three validation sets containing images from 198 sites in 18 other US states. In validation set A (n=27,415 patients, all undilated), the DLS detected poor blood glucose control (HbA1c > 9%) with an area under receiver operating characteristic curve (AUC) of 70.2% (95%CI 69.4-70.9); moderate-or-worse DR with an AUC of 75.3% (95%CI 74.4- 76.2); diabetic macular edema with an AUC of 78.0% (95%CI 76.4-79.6); and vision-threatening DR with an AUC of 79.4% (95%CI 78.1-80.8). For all 4 prediction tasks, the DLS’s AUC was higher (p<0.001) than using available self-reported baseline characteristics (age, sex, race/ethnicity, years with diabetes). In terms of positive predictive value, the top 5% of patients based on DLS-predicted likelihood had a 67% chance of having HbA1c > 9%, and a 20% chance of having vision threatening diabetic retinopathy that needed ophthalmology followup (vs. 54% and 14%, respectively for baseline characteristics). Similarly, the odds ratio per standard deviation increase in the DLS prediction was 2.0 for HbA1c > 9% and 1.6 for vision threatening diabetic retinopathy after adjusting for baseline characteristics (p<0.001 for both). The results generalized to patients with dilated pupils in validation set B (n=5,058 patients) and to patients at a different screening service (validation set C, n=10,402 patients). Our results indicate that external eye photographs contain information useful for healthcare providers managing patients with diabetes, and may help prioritize patients for in-person screening. Further work is needed to validate these findings on different devices (e.g., computer web cameras and front-facing smartphone cameras) and patient populations (those without diabetes) to evaluate its utility for remote diagnosis and management.