Deep Learning-Based Assessment of Sleepiness from Fingertip Pulse Wave Analysis During Objective Daytime Testing.

Abstract

Purpose: Objective evaluation of daytime sleepiness is a key factor in the management of sleep disorders. However, objective testing is limited due to the labor-intensive nature of standard methods; the Multiple Sleep Latency Test (MSLT) and the Maintenance of Wakefulness Test (MWT). Deep learning analysis of a simple fingertip pulse wave (photoplethysmography, PPG) recording could be a viable option for the routine assessment of sleepiness in patients with suspected sleep disorders.

Methods: A deep learning-based sleep staging model was tested in MSLT (n = 143) and MWT (n = 127) datasets. The primary aim was to analyze the classification of subjects as sleepy versus non-sleepy using PPG-based sleep analysis. As an intermediate technical evaluation, we assessed PPG-based sleep-wake classification and mean sleep latency (MSL).

Results: Classification of sleepiness status based on automatic MSL detection from PPG showed promising results in both MSLT (accuracy: 80%, sensitivity: 65%, specificity: 84%) and MWT (accuracy: 83%, sensitivity: 62%, specificity: 86%). Sleep probability curves estimated from the fingertip pulse wave differed significantly between sleepy and non-sleepy groups, both in MSLT (p = 0.0002) and in MWT (p = 0.0012) datasets. In MSLT, the model detected sleep with moderate agreement and balanced performance (accuracy 81%, precision 0.71, recall 0.80). In MWT, overall accuracy (88%) was strongly influenced by class imbalance (98% wakefulness), and direct second-by-second sleep detection showed limited precision (precision 0.10, recall 0.72). Automated MSL estimates showed correlation with manual scoring in both MSLT and MWT; however, larger errors and outliers were observed in MWT.

Conclusion: Fingertip pulse wave analysis combined with deep learning shows feasibility for automated assessment of daytime sleepiness, particularly in MSLT. However, direct zero-shot application of a model trained on overnight PSG to MWT is limited by low sleep precision requiring methodological adaptation and daytime-specific optimization.