Sleep Disorder Pattern Identification Using Markov Chain Monte Carlo in Health Monitoring Devices
Annotatsiya
Consumer health-monitoring wearables increasingly report sleep-disorder risk as a single probability, a presentation that discards the distinction between a confident estimate (built on clean, consistent multi-night signal) and an uncertain one (built on noisy or contradictory data)-a distinction with direct consequences for whether a user should trust the estimate or seek clinical confirmation. Markov Chain Monte Carlo (MCMC) inference within a Bayesian Hidden Markov Model (HMM) framework estimates not a point value but a full posterior distribution over latent sleep-disorder states and model parameters, naturally quantifying this uncertainty as a by-product of the inference itself. Physiological data from a wearable device (HRV,$\mathbf{S p O}_{\mathbf{2}}$, accelerometry, estimated RR) were used to fit the 4-state Bayesian HMM, with 96 of the 480 participants (20.0%) also having undergone simultaneous gold-standard polysomnography (PSG) for validation. The Bayesian HMM has a higher AUC than the supervised Random Forest benchmark (88.4% vs. 88.4%), which lacks any quantification of posterior uncertainty, and higher than the rule detection with support vectors (80.4% vs. 70.4%), indicating an increased sensitivity and specificity for OSA relative to these methods (19.9% and 5.0% respectively). The model's 95% credible intervals are good for the true event count, as confirmed by PSG, with the true count being included in 91/96 (94.8%) of the validation nights compared to the 95% nominal level. With the resulting uncertainty estimates applied operationally, 15.0 per cent (1,296 of 8,640) of monitored nights are flagged for remonitoring because the credible intervals for these nights are large, and 74.7 per cent (968 of 1,296) of them are also independently identified as having poor signal quality, confirming that the posterior uncertainty reflects real data quality issues, not noise in the model. A separate, larger cohort of patients$(\mathbf{N}=\mathbf{6 4 2})$were flagged as algorithmic high-probability-OSA (APHS) and went on to have clinical PSG confirmed, with 504 (78.5%) of them being positive. Current limitations include model-complexity sensitivity, limits set by the accuracy of the sensors at the hardware level, the computational cost of carrying out full MCMC inference, and the degradation in performance of higher-BMI participants.
Hali tarjima qilinmagan