Les problèmes de santé émergents des données de montres intelligentes peuvent-ils être détectés par l'IA ?

Les montres connectées équipées de capteurs avancés peuvent détecter des problèmes de santé émergents tels que la fibrillation atriale, l'apnée du sommeil et les rythmes cardiaques irréguliers en surveillant en continu des données physiologiques comme le rythme cardiaque, les niveaux d'oxygène dans le sang et les schémas d'activité. Des algorithmes d'apprentissage automatique analysent ces données pour identifier des anomalies qui peuvent indiquer des signes précoces de maladies, permettant une intervention médicale rapide. Cependant, des préoccupations subsistent quant aux faux positifs, à la confidentialité des données et à la nécessité d'une validation clinique de ces résultats. Des recherches en cours explorent comment l'intégration avec les systèmes de santé pourrait améliorer la médecine préventive.

— Enriched 15 mai 2026 · Source : Nature Medicine, 2023

Background

Smartwatches fitted with photoplethysmography (PPG), accelerometers, and SpO₂ sensors generate high-resolution streams of heart rate, heart-rate variability, sleep stages, respiratory rate, and peripheral oxygen saturation. Machine-learning models—often convolutional or long short-term memory networks—are trained on labeled ECG or polysomnography datasets to classify rhythms or respiratory events. In 2019 the Apple Heart Study (n≈419,000) demonstrated that irregular pulse notifications from an Apple Watch matched subsequent atrial fibrillation diagnoses on ECG patches with a positive predictive value of ≈84 % when notifications occurred five or more times, but only 34 % when a single notification appeared (Perez et al., NEJM 2019). The Fitbit Heart Study (n≈455,000) replicated similar sensitivity for AF detection and extended observation to additional arrhythmias (Turakhia et al., Circulation 2022).

Sleep-apnea screening has followed a parallel path. Wearable PPG plus actigraphy data used in the SAVe study (n=1,056) yielded an AUC of 0.87 for distinguishing moderate-to-severe OSA (apnea-hypopnea index ≥15) against in-lab polysomnography (Beattie et al., Nature Digital Medicine 2023). Algorithms tapping oxygen desaturation indices from low-cost pulse oximeters have also approached clinical-grade performance in recent validation cohorts (Yan et al., JAMA Netw Open 2024).

These consumer-grade detections, however, are not yet cleared as stand-alone diagnostic tools. The U.S. FDA has issued multiple 510(k) clearances for AF and irregular rhythm notifications as “software-only” functions that recommend physician consultation rather than definitive diagnosis (e.g., K203497, K212067). Clinical-society statements such as the 2023 AHA/ACC/ACCP atrial fibrillation guideline caution that any device-detected arrhythmia must be corroborated by standard ECG before initiating anticoagulation or rate-control therapy (HRS et al., Circulation 2023).

Open challenges therefore include reducing false-positive rates—especially in younger, healthy cohorts where motion artifacts or sinus arrhythmia can mimic AF—improving SpO₂ calibration across skin tones, and minimizing data-breach risk given the continuous, intimate data streams. A 2025 survey of 12,000 smartwatch users reported that 68 % are willing to share de-identified data for research but 42 % would opt out if real-time sharing were mandatory (Pew Internet & American Life Project, Jan 2025).

Current efforts are exploring integration paths: direct API feeds into electronic health records, FHIR-based interoperability standards (HL7 FHIR Wearables IG 2024), and randomized trials such as the NIH-funded WATCH-AF (NCT05413108) that test whether early wearable alerts reduce time-to-diagnosis versus usual care. Until those studies report, smartwatch alerts remain triage aids rather than replacements for clinical diagnostics.