Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SweetDeep: A Wearable AI Solution for Real-Time Non-Invasive Diabetes Screening

Ian Henriques, Lynda Elhassar, Sarvesh Relekar, Denis Walrave, Shayan Hassantabar, Vishu Ghanakota, Adel Laoui, Mahmoud Aich, Rafia Tir, Mohamed Zerguine, Samir Louafi, Moncef Kimouche, Emmanuel Cosson, Niraj K Jha

TL;DR

SweetDeep advances non-invasive diabetes screening by combining smartwatch-derived ECG, PPG, and BIA features with demographic data in a compact edge model. Trained on a decentralized home-data collection protocol, it achieves about 82% patient-level accuracy with well-calibrated probabilities and a practical abstention option to defer uncertain cases. The approach demonstrates real-world generalization across diverse populations and diurnal variations, highlighting the feasibility of rapid, low-cost screening in primary care and telehealth. Future work includes expanding to a three-class ND-PD-T2D model and validating across larger, more diverse cohorts.

Abstract

The global rise in type 2 diabetes underscores the need for scalable and cost-effective screening methods. Current diagnosis requires biochemical assays, which are invasive and costly. Advances in consumer wearables have enabled early explorations of machine learning-based disease detection, but prior studies were limited to controlled settings. We present SweetDeep, a compact neural network trained on physiological and demographic data from 285 (diabetic and non-diabetic) participants in the EU and MENA regions, collected using Samsung Galaxy Watch 7 devices in free-living conditions over six days. Each participant contributed multiple 2-minute sensor recordings per day, totaling approximately 20 recordings per individual. Despite comprising fewer than 3,000 parameters, SweetDeep achieves 82.5% patient-level accuracy (82.1% macro-F1, 79.7% sensitivity, 84.6% specificity) under three-fold cross-validation, with an expected calibration error of 5.5%. Allowing the model to abstain on less than 10% of low-confidence patient predictions yields an accuracy of 84.5% on the remaining patients. These findings demonstrate that combining engineered features with lightweight architectures can support accurate, rapid, and generalizable detection of type 2 diabetes in real-world wearable settings.

SweetDeep: A Wearable AI Solution for Real-Time Non-Invasive Diabetes Screening

TL;DR

SweetDeep advances non-invasive diabetes screening by combining smartwatch-derived ECG, PPG, and BIA features with demographic data in a compact edge model. Trained on a decentralized home-data collection protocol, it achieves about 82% patient-level accuracy with well-calibrated probabilities and a practical abstention option to defer uncertain cases. The approach demonstrates real-world generalization across diverse populations and diurnal variations, highlighting the feasibility of rapid, low-cost screening in primary care and telehealth. Future work includes expanding to a three-class ND-PD-T2D model and validating across larger, more diverse cohorts.

Abstract

The global rise in type 2 diabetes underscores the need for scalable and cost-effective screening methods. Current diagnosis requires biochemical assays, which are invasive and costly. Advances in consumer wearables have enabled early explorations of machine learning-based disease detection, but prior studies were limited to controlled settings. We present SweetDeep, a compact neural network trained on physiological and demographic data from 285 (diabetic and non-diabetic) participants in the EU and MENA regions, collected using Samsung Galaxy Watch 7 devices in free-living conditions over six days. Each participant contributed multiple 2-minute sensor recordings per day, totaling approximately 20 recordings per individual. Despite comprising fewer than 3,000 parameters, SweetDeep achieves 82.5% patient-level accuracy (82.1% macro-F1, 79.7% sensitivity, 84.6% specificity) under three-fold cross-validation, with an expected calibration error of 5.5%. Allowing the model to abstain on less than 10% of low-confidence patient predictions yields an accuracy of 84.5% on the remaining patients. These findings demonstrate that combining engineered features with lightweight architectures can support accurate, rapid, and generalizable detection of type 2 diabetes in real-world wearable settings.

Paper Structure

This paper contains 25 sections, 13 equations, 6 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. The Biochemical "Gold Standard"
  4. Biophysical Sensing Alternatives
  5. Model Reliability
  6. Methodology
  7. Framework Overview
  8. Data Collection Protocol
  9. Sensor Quality Control
  10. Clinical Feature Extraction
  11. Time-of-Day Representation
  12. Fold Creation and Model Training
  13. Prediction and Abstention
  14. Performance Evaluation Metrics
  15. Accuracy Measures
  16. ...and 10 more sections

Figures (6)

  • Figure 1: An overview of the SweetDeep real-time inference framework, consisting of three main stages: signal and questionnaire data collection with quality control, neural network inference (with two hidden layers), and optional confidence-based abstention.
  • Figure 2: Example delineation of many stable, filtered ECG beats from a single recording after quality control steps are applied. Notably, the R Onsets and T Offsets, two components of QTc, have consistent positions across heartbeats.
  • Figure 3: Time-of-day counts (clustered by hour) for instances in non-diabetic and diabetic cohorts. Most hours have $\ge$50 instances from both cohorts, supporting the inclusion of time-of-day features to account for circadian variations in physiological signals.
  • Figure 4: Toy calibration curves, with 10 evenly spaced bins. (green = well-calibrated, pink = poorly calibrated)
  • Figure 5: SweetDeep three-fold calibration curves. Points closer to the dotted diagonal line indicate better bin-wise calibration.
  • ...and 1 more figures