2 months ago
- Non-invasive glucose monitoring is more difficult than blood oxygen monitoring due to glucose's low concentration (1000x less than hemoglobin), lack of distinct absorption peaks, and inability to use the AC/DC separation method.
- Various techniques have been explored: near-infrared (NIR) achieves ~20-25% MARD due to weak signals; mid-infrared (MIR) shows ~12% MARD but requires miniaturization; Raman spectroscopy reaches ~11.7-14.3% MARD but needs high laser power; photoacoustic spectroscopy is infeasible due to skin variability; fluorescence sensors like Eversense achieve ~8.5% MARD but are implantable; PPG with deep learning initially reports high accuracy but fails in large cohorts.
- Recent progress includes Apple's foundation model trained on 141,000 PPG/ECG waveforms, which successfully detected hypertension but remains unproven for glucose.
- The most likely path forward combines advanced machine learning with hardware advances (MIR or Raman), as indicated by Apple shrinking a tabletop prototype to iPhone size.