Logo do repositório

Percorrer por data de Publicação, começado por "2026-09-17"

Filtrar resultados por ano ou mês
A mostrar 1 - 1 de 1
  • Deep learning-based estimation of respiration from PPG: from rate extraction to apnea phenotype detection in wearable devices
    Publication . Rodrigues, Sandro Canada; Mendonça, Fábio Rúben Silva; Durmer, Jeffrey Scott
    Sleep apnea a ects nearly one billion people worldwide, yet diagnosis still relies on polysomnogra phy, which is costly and complex. Wearable photoplethysmographic (PPG) sensors could support large-scale screening, but current methods mostly estimate respiratory rate or detect apnea within a single dataset, and phenotype analysis and cross-sensor generalization remain limited. This work presents a deep learning pipeline for sleep apnea detection, severity estimation and phenotype classi cation from PPG signals. A CNN-LSTM model rst showed that respiratory information can be extracted from PPG, with mean absolute errors of 1.70 and 1.80 breaths per minute on two held-out public test sets (BIDMC and CapnoBase). The CoTELMo architecture, a CNN-BiLSTM model trained on event level annotations from 2,056 patients of the Multi-Ethnic Study of Atherosclerosis (MESA), then screened moderate-to-severe apnea with an Area Under the Receiver Operating Characteristic Curve (AUC) of 0.909, and a two-stage regression estimated the apnea-hypopnea index with a Pearson correlation of 0.853. On the Cleveland Family Study (N=324), performance fell to an AUC of 0.568 without adapta tion, while ne-tuning with a frozen backbone and patient-level labels recovered an AUC of 0.832. In an exploratory pilot with a commercial wearable ring (N=46, of which 12 were pathological), direct inference reached an AUC of 0.917 for severe obstructive apnea, and CoTELMo-derived features separated obstructive from central apnea in all 12 pathological cases. These pilot results require validation in larger cohorts. The results suggest that PPG can support apnea screening and severity estimation, but that transfer to new sensors and populations requires supervised adaptation and local validation.
    2026-09-17Dissertação de mestrado Acesso aberto Ver mais