Why non-invasive glucose monitoring is hard (www.empirical.health)

Continuous non-invasive glucose monitoring remains a challenging goal for tech companies like Apple, despite advancements in health tracking features on smartwatches. Various technologies for measuring glucose levels have been explored, including near-infrared (NIR), mid-infrared (MIR), Raman spectroscopy, and photoplethysmography (PPG) enhanced by machine learning. A significant hurdle is glucose's low concentration in blood, lack of distinct absorption peaks, and its distributed presence in various bodily fluids, which complicates accurate measurements. Current approaches have produced varying degrees of accuracy, with techniques like PPG reaching 60-94% accuracy in small studies, but falling short in larger cohorts, while MIR shows promise with around 12% mean absolute relative difference (MARD) in clinical trials. For the AI/ML community, the ongoing efforts highlight the potential for integrating advanced machine learning techniques to improve glucose monitoring capabilities in consumer wearables. By utilizing PPG data in conjunction with deep neural networks, there may be opportunities to unlock insights previously thought unattainable. Nonetheless, the path forward likely requires both hardware advancements and innovative modeling to accurately capture glucose signals, making this a key area of research as developers aim for future FDA-approved non-invasive solutions. This pursuit underscores the intersection of health technology and machine learning, emphasizing the need for interdisciplinary approaches to overcome significant scientific and engineering challenges.