Predicting All-Cause Mortality in Patients With Obstructive Sleep Apnea Using Sleep-Related Features: A Machine-Learning Approach

Abstract

Background and purpose: Obstructive sleep apnea (OSA) is associated with an increased risk of adverse outcomes, including mortality. Machine-learning algorithms have shown potential in predicting clinical outcomes in patients with OSA. This study aimed to develop and evaluate a machine-learning algorithm for predicting 10- and 15-year all-cause mortality in patients with OSA.

Methods: Patients with OSA were stratified into deceased and alive groups based on mortality outcomes. Various sleep-related features were analyzed, including objective sleep measures and the heart-rate variability during various sleep stages. The light gradient-boosting machine (LGBM) algorithm was employed to construct a risk-stratification model. The predictive performance of the model was assessed using the area under the receiver operating characteristic curve (AUC) for predicting mortality over 10 and 15 years. Survival analysis was conducted using Kaplan-Meier plots and Cox proportional-hazards model.

Results: This study found that parasympathetic activity was higher in OSA patients with worse outcomes than in those with better outcomes. The LGBM-based prediction model with sleep-related features was moderately accurate, with a mean AUC of 0.806 for predicting 10- and 15-year mortality. Furthermore, survival analysis demonstrated that LGBM could significantly distinguish the high- and low-risk groups, as evidenced by Kaplan-Meier plots and Cox regression results.

Conclusions: This study has confirmed the potential of sleep-related feature analysis and the LGBM algorithm for evaluating the mortality risk in OSA patients. The developed risk-stratification model offers an efficient and interpretable tool for clinicians that emphasizes the significance of patient-specific autonomic responses in mortality prediction. Incorporating survival analysis further validated the robustness of the model in predicting long-term outcomes.

Keywords: autonomic nervous system; machine learning; mortality; obstructive sleep apnea; sleep stages.

Fig. 1. Bar plot of selected HRV variables across different sleep stages for OSA and non-OSA patients. The X-axis represents the NREM N2, N3, and REM sleep stages. The Y-axis represents the values of RMSSD (A), LF/HF ratio (B), LFnu (C), and HFnu (D). Error bars indicate 95% confidence intervals. ^*p<0.05; ^**p<0.005. HF, power in the high-frequency range (0.15–0.40 Hz); HFnu, normalized HF power; HRV, heart-rate variability; LF, power in the low-frequency range (0.04–0.15 Hz); LFnu, normalized LF power; N2, NREM stage 2; N3, NREM stage 3; NREM, nonrapid eye movement; OSA, obstructive sleep apnea; REM, rapid eye movement; RMSSD, square root of the mean-squared differences of successive differences.

Fig. 2. ROC curves and AUCs with 95% confidence intervals for the five machine-learning models. The ROC curves indicate the performance in predicting all-cause mortality in patients with OSA within 10 years (A) and 15 years (B). AUC, area under the ROC curve; LGBM, light gradient-boosting machine; OSA, obstructive sleep apnea; ROC, receiver operating characteristic; XGBoost, extreme gradient-boosting.

Fig. 3. Plots of the top-20 predictors ranked by LGBM based on SHAP values. The bar chart of the mean absolute SHAP values for global feature importance (left side) and the SHAP summary plot for the local explanation (right side) are illustrated for predicting 10-year mortality (A) and 15-year mortality (B). Features are ordered from top to bottom according to their complexity. Each point on each feature line represents a specific individual, and the color of that point reflects the value of the predictor, ranging from red (high) to blue (low). Points on the X-axis indicate the SHAP values corresponding to the effect of the variable on the prediction for each individual; dots toward the right and left indicate higher and lower risks, respectively. AHI, apnea-hypopnea index; BMI, body mass index; HF, high-frequency; HFnu, HF power; LF, low-frequency; LFnu, LF power; LGBM, light gradient-boosting machine; N2, NREM stage 2; N3, NREM stage 3; NNI, normal-to-normal interval; NREM, nonrapid eye movement; REM, rapid eye movement; SDNN, standard deviation of normal-to-normal intervals; SHAP, SHapley Additive exPlanations; sq, indicator of sleep quality; WASO, wake time after sleep onset.