0292 Detecting Sleep Deficiency with Voice Biomarkers and Machine Learning

Abstract Introduction Accurate biomarkers of insufficient sleep have been a central interest of sleep scientists. Given advancements in artificial intelligence, researchers have explored non-invasive digital biomarkers from human voices. In this study, we conducted a within-participant counterbalanced controlled trial of chronic sleep restriction (CSR) and leveraged machine learning to investigate voice biomarkers for detecting sleep deficiency. Methods Healthy young adults completed a 32-day in-patient protocol. The protocol included 5 days of baseline 8-hour time-in-bed (TIB) followed by 5 days of CSR (5-hour TIB), during which their light exposure, activity, and diets were controlled. Every 4 hours during waking episodes, a voice measurement (VM) was administered via computer. During each VM, the participant read 10 sentences at their habitual volume, pitch, and pace. The sentences were phonetically balanced and tailored for professional speech quality assessment. Each VM used different sentences to prevent memorization. An omnidirectional microphone with an adjustable stand was used. 85 common acoustic features, such as fundamental frequency, formants, and mel-frequency cepstral coefficients, were extracted from each VM. Additionally, 11 features from the Cepstral Spectral Index of Dysphonia were extracted, forming a 96-dimensional vector. Within each participant, the vectors were z-scored to remove personal vocal traits. After dimension reduction via principal component analysis, under a leave-one-out procedure, we trained a support vector machine (SVM) to classify recordings into those taken during CSR or baseline. We tested the SVM on one unseen participant every iteration. Results After excluding low-quality data, we analyzed 196 VMs (100 during CSR and 96 during baseline) from 5 participants, including 1 non-native English speaker. The SVM reached a sensitivity of 0.74 (95% CI: 0.71-0.78), a specificity of 0.71% (95% CI: 0.68-0.75), and an area under the ROC curve of 0.76 (95% CI: 0.73-0.80) in classifying CSR vs. baseline conditions. Conclusion These preliminary findings reveal that vocal characteristics may represent a target for non-invasive, objective sleep deficiency biomarkers. Additional studies in both controlled conditions and field settings should be carried out to explore whether features of the human voice can be developed for non-intrusive monitoring of sleep status for sleep disorders patients, fitness-for-duty evaluations, and other applications in ambulatory settings. Support (if any)