Artificial intelligence-enabled detection and assessment of Parkinson's disease using nocturnal breathing signals

Abstract

There are currently no effective biomarkers for diagnosing Parkinson's disease (PD) or tracking its progression. Here, we developed an artificial intelligence (AI) model to detect PD and track its progression from nocturnal breathing signals. The model was evaluated on a large dataset comprising 7,671 individuals, using data from several hospitals in the United States, as well as multiple public datasets. The AI model can detect PD with an area-under-the-curve of 0.90 and 0.85 on held-out and external test sets, respectively. The AI model can also estimate PD severity and progression in accordance with the Movement Disorder Society Unified Parkinson's Disease Rating Scale (R = 0.94, P = 3.6 × 10^-25). The AI model uses an attention layer that allows for interpreting its predictions with respect to sleep and electroencephalogram. Moreover, the model can assess PD in the home setting in a touchless manner, by extracting breathing from radio waves that bounce off a person's body during sleep. Our study demonstrates the feasibility of objective, noninvasive, at-home assessment of PD, and also provides initial evidence that this AI model may be useful for risk assessment before clinical diagnosis.

Fig. 1. Overview of the AI model for PD diagnosis and disease severity prediction from nocturnal breathing signals.

The system extracts nocturnal breathing signals either from a breathing belt worn by the subject, or from radio signals that bounce off their body while asleep. It processes the breathing signals using a neural network to infer whether the person has PD, and if they do, assesses the severity of their PD in accordance with the MDS-UPDRS.

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.

Fig. 3. PD severity prediction from nocturnal breathing signals.

a, Severity prediction of the model with respect to MDS-UPDRS (two-sided t-test). The center line and the shadow denote the mean and 95% CI, respectively. b, Severity prediction distribution of the model with respect to the H&Y stage; a higher H&Y stage indicates increased PD severity (Kruskal–Wallis test). 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. c, Test–retest reliability of PD severity prediction as a function of the number of nights per subject. The dots and the shadow denote the mean and 95% CI, respectively. The model achieved a reliability of 0.97 (95% CI (0.95, 0.98)) with 12 nights of data. d, Correlation of the AI model predictions with MDS-UPDRS subpart I (two-sided t-test). e, Correlation of the AI model predictions with MDS-UPDRS subpart II (two-sided t-test). f, Correlation of the AI model predictions with MDS-UPDRS subpart III (two-sided t-test). g, Correlation of the AI model predictions MDS-UPDRS subpart IV (two-sided t-test). The center line and the shadow denote the mean and 95% CI, respectively. Data in all panels are from the wireless dataset (n = 53 subjects).

Fig. 4. Model evaluation for PD risk assessment before actual diagnosis, and disease progression tracking using longitudinal data.

a, Model prediction scores for the prodromal PD group (that is, undiagnosed individuals who were eventually diagnosed with PD) and the age- and gender-matched control group (one-tailed Wilcoxon rank-sum test). b, The AI model assessment of the change in MDS-UPDRS over 6 months (one-tailed one-sample Wilcoxon signed-rank test) and the clinician assessment of the change in MDS-UPDRS over the same period (one-tailed one-sample Wilcoxon signed-rank test). c, The AI model assessment of the change in MDS-UPDRS over 12 months (one-tailed one-sample Wilcoxon signed-rank test) and the clinician assessment of the change in MDS-UPDRS over the same period (one-tailed one-sample Wilcoxon signed-rank test). d, Continuous severity prediction across 1 year for the patient with maximum MDS-UPDRS increase (Kruskal–Wallis test; n = 365 nights from 1 September 2019 to 31 October 2020). For each box in a–d, 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.

Fig. 5. Interpretation of the output of the AI model with respect to EEG and sleep status.

a,b, Attention scores were aggregated according to sleep status and EEG bands for PD patients (n = 736 nights from 732 subjects) and controls (n = 7,844 nights from 6,840 subjects). Attention scores were normalized across EEG bands or sleep status. Attention scores for different EEG bands between PD patients and control individuals (one-tailed Wilcoxon rank-sum test) (a). Attention scores for different sleep status between PD patients and control individuals (one-tailed Wilcoxon rank-sum test) (b). On each box in a and b, 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.

Extended Data Fig. 1. Nocturnal breathing data collection setup.

a, Data from the breathing belt is collected by wearing an on-body breathing belt during sleep. b, Data from wireless signals is collected by installing a low-power wireless sensor in the subject’s bedroom, and extracting the subject’s breathing signals from the radio signals reflected off their body. c, d, Two samples of full-night nocturnal breathing from breathing belt and wireless signal and their zoomed-in versions.

Extended Data Fig. 2. Cumulative distributions of the prediction score for PD diagnosis.

a, Results for breathing belt data (n = 6,660 nights from 5,652 subjects). b, Results for wireless data (n = 2,601 nights from 53 subjects). For both data types, fixing a threshold of 0.5 leads to good performance (that is, sensitivity 80.22% and specificity 78.62% for breathing belt, and sensitivity 86.23% and specificity 82.83% for wireless data).

Extended Data Fig. 3. Disease progression tracking using a different number of nights.

a, b, 6-month and 12-month change in MDS-UPDRS as assessed by a clinician and predicted by the AI model, both using a single night and multiple nights of data. 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. Similar to the clinician assessment, when using only a single night of data, the AI model cannot detect statistically significant changes over a year (p = 0.751 for 6 months, p = 0.235 for 12 months, one-tailed one-sample Wilcoxon signed-rank test). This indicates that the reason why the AI model can achieve statistical significance for progression analysis while MDS-UPDRS cannot stems from being able to combine measurements from multiple nights, which substantially reduces measurement noise and increases sensitivity.

Extended Data Fig. 4. Performance of the AI model on differentiating subjects with Parkinson’s disease (PD) from subjects with Alzheimer’s disease (AD).

a, The model’s output scores differentiate PD subjects from AD subjects (p = 3.52e-16, one-tailed Wilcoxon rank-sum test). b, Receiver operating characteristic (ROC) curves for detecting PD subjects against AD subjects (n = 148). The model achieves high AUC for differentiating PD from AD (AUC = 0.895).

Extended Data Fig. 5. Visualization examples of the attention of the AI model.

a, b, Full-night attention distribution (left) and its zoomed-in version (right) overlayed on the corresponding qEEG bands, sleep stages, and breathing. Graphs show a control subject in (a) and a PD patient in (b). For the control subject, the model’s attention focuses on periods with high Delta waves, which correspond to deep sleep. In contrast, for the PD subject, the model attends to periods with relatively high Beta or Alpha waves, and awakenings around sleep onset and in the middle of sleep.

Extended Data Fig. 6. Ablation studies for assessing the benefit of multi-task learning and transfer learning.

a, PD diagnosis performance on breathing belt data (n = 6,660 nights from 5,652 subjects) and wireless data (n = 2,601 nights from 53 subjects), for the model with all of its components and the model without the qEEG auxiliary task (that is, without multi-task learning) and without transfer learning. Each bar and its error bar indicate the mean and standard deviation across 5 independent runs. The graphs indicate that the qEEG auxiliary task is essential for good performance and eliminating it reduces the AUC by almost 40%. Transfer learning also boosts performance for both breathing belt data (7.8% improvements) and wireless data (8.3% improvements), yet is not as essential as multi-task learning. b, Pearson correlation of the PD severity prediction and MDS-UPDRS. The correlation is computed for subjects in the wireless datasets (n = 53 subjects) since their MDS-UPDRS scores are available. Each bar and its error bar indicate the mean and standard deviation across 5 independent runs. The results indicate that transfer learning is useful, but multi-task learning (that is, the qEEG auxiliary task) is essential for good performance.

Extended Data Fig. 7. Performance comparison of the model with two machine learning baselines: Support Vector Machine (SVM) and a neural network based on ResNet and LSTM.

a, PD diagnosis performance on breathing belt data (n = 6,660 nights from 5,652 subjects) and wireless data (n = 2,601 nights from 53 subjects). Each bar and its error bar indicate the mean and standard deviation across 5 independent runs. b, Pearson correlation of PD severity prediction and MDS-UPDRS. The correlation is computed for subjects in the wireless datasets (n = 53 subjects) since their MDS-UPDRS scores are available. Each bar and its error bar indicate the mean and standard deviation across 5 independent runs.

Extended Data Fig. 8. Evaluation results for predicting the full-night qEEG summary from nocturnal breathing signals.

a, One prediction sample of a full-night qEEG. The time resolution of the predicted qEEG is 1 second. b, Distribution of the prediction errors across four EEG bands (n = 6,660 nights from 5,652 subjects). The AI model made an unbiased (that is, median-unbiased) estimation of EEG prediction for all bands. 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. c, Cumulative distribution functions of the absolute prediction error across four EEG bands (n = 6,660 nights from 5,652 subjects).

Extended Data Fig. 9. Neural network architecture of the AI-based model.

a, The neural network takes as input a night of nocturnal breathing. The main task of PD prediction consists of a breathing encoder, a PD encoder, a PD classifier and a PD severity predictor. We also introduce an auxiliary task of predicting the subject’s qEEG during sleep. We include also two discriminators for domain-invariant transfer learning. b, The detailed architecture of each neural network module in the model.