Fig. 2: PD diagnosis from nocturnal breathing signals.

a, ROC curves for detecting PD from breathing belt (n = 6,660 nights from 5,652 subjects). b, ROC curves for detecting PD from wireless data (n = 2,601 nights from 53 subjects). c, Test–retest reliability of PD diagnosis as a function of the number of nights used by the AI model. The test was performed on 1 month of data from each subject in the wireless dataset (n = 53 subjects). The dots and the shadow denote the mean and 95% CI, respectively. The model achieved a reliability of 0.95 (95% CI (0.92, 0.97)) with 12 nights of data. d,e, Distribution of PD prediction (pred.) scores for subjects with several nights (n₁ = 1,263 nights from 25 PD subjects and n₂ = 1,338 nights from 28 age- and gender-matched controls). The graphs show a boxplot of the prediction scores as a function of the subject ids. On each box, the central line indicates the median, and the bottom and top edges of the box indicate the 25th and 75th percentiles, respectively. The whiskers extend to 1.5 times the interquartile range. Points beyond the whiskers are plotted individually using the + symbol. f, ROC curves for detecting PD on an external test set from Mayo Clinic (n = 1,920 nights from 1,920 subjects). The model has an AUC of 0.851 with a sensitivity of 80.12% and specificity of 72.65%. g, Cross-institution PD prediction performance on SHHS (n = 2,630 nights from 2,630 subjects). h, Cross-institution PD prediction performance on MrOS (n = 3,883 nights from 2,875 subjects). In this analysis, all data from one institution was held back as test data, and the AI model was retrained excluding all data from that institution. Cross-institution prediction achieved an AUC of 0.857 with a sensitivity of 76.92% and specificity of 83.45% on SHHS, and an AUC of 0.874 with a sensitivity of 82.69% and specificity of 75.72% on MrOS.