A Wrist-Worn Photoplethysmography Platform for Blood Glucose Estimation Using Embedded CNN Inference and CNN–SVR Modeling

A wrist-worn reflective PPG prototype (MAX30102 optical front end, ESP32-S3 touch-display microcontroller, coin-cell supply, rear-aperture wrist enclosure) runs a CNN–SVR blood-glucose estimation workload at the edge, with 256-sample windows at 50 Hz and a deterministic signal-quality gate that rejects poor windows before inference. The authors frame the contribution as a reproducible embedded system design, explicitly not a clinical claim of non-invasive glucose-monitoring accuracy.

• DOI: 10.1109/jsen.2026.3747957
• Prototype: MAX30102 PPG front end + ESP32-S3, 256-sample windows at 50 Hz
• Hybrid CNN–SVR on 18-record wrist validation: 18.21 mg/dL MAE, 16.37% MARD, Clarke A/B/D 13/5/0
• Authors frame it as a system-boundary reference design, not a clinical accuracy claim

This paper presents a wrist-worn reflective photoplethysmography (PPG) platform for translating a CNN–SVR blood-glucose estimation workload from controlled PPG recordings to a bounded wearable edge-inference setting. The contribution is a reproducible system boundary, not a clinical claim of non-invasive glucose-monitoring accuracy: optical acquisition, signal-quality admission, provenance logging, and microcontroller resource accounting are specified together. The prototype combines a MAX30102 PPG front end, an ESP32-S3 touch-display microcontroller, a coin-cell supply, and a compact rear-aperture wrist enclosure. Firmware uses interrupt/FIFO readout to form 256-sample windows at 50 Hz, records sensor-register and preprocessing metadata, rejects FIFO/sample-integrity failures before preprocessing, and applies a deterministic first-failure signal-quality gate that rejects windows containing non-finite samples, insufficient variability, extreme within-window z-scores, flat-segment content, excessive baseline drift, low pulse-band SNR, or weak pulse periodicity. The workload extracts a 64-D representation with a 1-D CNN and compares a dense CNN regression head with an SVR head. The CNN extractor requires 4.21 MMAC and 80,160 weights per window; the SVR head adds support-vector storage and O(NSVd) kernel operations. On an 18-record wrist-device validation partition, the hybrid CNN–SVR achieved 18.21 mg/dL MAE, 23.89 mg/dL RMSE, 16.37% MARD, −0.83 mg/dL bias, and 13/5/0 Clarke Zone A/B/D counts, lower error compared to the CNN-only and standalone SVR alternatives on this split. CNN-only avoids support-vector memory traffic and is therefore the lower-complexity deployment candidate, and the hybrid is an experimental alternative. These results provide a watch-scale reference design and model-selection evidence for larger subject-wise wrist validation.

Verbatim from OpenAlex (opens in a new tab) · Abstract via OpenAlex (publisher-deposited); © the authors/publisher

No results from this paper on ChatPPG leaderboards yet. Browse the leaderboards.