JMIR Publications

ABSTRACT

Background:

Objective assessment of insomnia remains challenging, and most physiological investigations have focused on nocturnal sleep. Whether daytime electrophysiological signals contain discriminative information related to insomnia remains insufficiently explored.

Objective:

This exploratory cross-sectional study aimed to investigate the feasibility of using daytime multimodal electrophysiological signals for insomnia assessment and to identify the physiological features contributing most strongly to machine-learning classification.

Methods:

A total of 152 participants were included. Daytime multimodal electrophysiological recordings comprising electrocardiography (ECG), right-frontal electroencephalography (EEG; F4–M1), and bilateral pulse-wave signals were collected. Candidate digital biomarkers characterizing waveform morphology, temporal characteristics, variability, and nonlinear complexity were extracted. Support vector machine (SVM), logistic regression, random forest, and k-nearest neighbors models were developed and evaluated using AUC, accuracy, sensitivity, specificity, precision, and F1-score.SHapley Additive exPlanations (SHAP) analysis was applied to interpret the optimal model. The top SHAP-ranked features were further compared between the insomnia and normal-control groups, with false discovery rate (FDR) correction and effect-size estimation.

Results:

SVM demonstrated the best overall classification performance among the evaluated models [AUC = 0.87; accuracy = 0.80]. Among the 10 most influential SHAP features, six were derived from pulse signals, two from ECG, and two from right-frontal EEG. Nine of these features showed lower values in the insomnia group and negative SHAP associations, whereas one pulse systolic-wave-height feature was higher and showed a positive SHAP association. All 10 features differed significantly between groups after FDR correction (FDR-adjusted P < 0.05), with effect sizes ranging from r = 0.22 to 0.57. Several highly ranked features were related to multiscale entropy, indicating that altered multiscale organization of electrophysiological signals contributed importantly to insomnia classification.

Conclusions:

Daytime EEG, ECG, and bilateral pulse-wave signals contain complementary physiological information associated with insomnia. The concordance between SHAP-derived feature contributions and group-level statistical differences suggests that insomnia may be characterized by altered multiscale and dynamic organization across central and peripheral physiological systems. These findings provide exploratory evidence supporting interpretable multimodal electrophysiological approaches for objective daytime assessment of insomnia, although external and longitudinal validation is required.