Deep Learning Meets Sleep Medicine: A Proof-of-Concept Clustering of Minute-Resolution CPAP Telemetry

Abstract

This proof-of-concept study demonstrates that minute-resolution telemetry from continuous positive airway pressure (CPAP) devices can be effectively repurposed for large-scale chronotype and adherence phenotyping. We collated 30 consecutive nights from n=200 de-identified ResMed patients into 30 x 1440 colour images, embedded each image with a frozen ResNet-50 convolutional neural network, and clustered the embeddings with k -means. Six distinct phenotypes emerged, capturing both sleep timing (early birds, typical sleepers, night owls) and adherence patterns (high, medium-high, inconsistent, fragmented, and non-adherent). The approach leverages routinely collected clinical data without the need for additional sensors, promising significant benefits for personalized sleep medicine. External validation against actigraphy and questionnaire-based chronotype measures is planned to further strengthen these findings.