# Spectral analysis of ECG and SpO₂ for machine learning classification of Sleep-Disordered breathing

## TL;DR

This study explores using ECG and SpO₂ data with machine learning to detect sleep-disordered breathing in hospitalized patients.

## Contribution

A novel machine-learning classifier using ECG and SpO₂ for non-invasive SDB detection in hospitalized patients is proposed.

## Key findings

- The SVM-1 classifier showed high sensitivity and specificity for detecting respiratory events.
- The combined model achieved 72.3% sensitivity and 73.3% specificity in the validation set.
- Performance was affected by arrhythmias and lack of respiratory effort measures.

## Abstract

Sleep-disordered breathing (SDB), particularly obstructive sleep apnea (OSA), is prevalent in hospitalized patients, yet remains underdiagnosed due to limited screening tools. This pilot study aimed to develop and validate a machine-learning classifier using electrocardiogram (ECG) and pulse oximetry (SpO₂) waveforms to detect SDB, leveraging routinely collected physiological data to enable non-invasive, scalable screening in hospitalized patients.

We utilized data from the Multi-Ethnic Study of Atherosclerosis (MESA) Sleep cohort, which included full overnight polysomnography (PSG). A total of 122 participants were randomly selected, enriched for severe OSA cases. Spectral analysis of R-R intervals and SpO₂ variability was performed, and a support vector machine (SVM) classifier was trained using a subset of subjects (n = 30). Performance was evaluated in a validation set (n = 92) using standard classification metrics with 95% CIs calculated.

The first-stage classifier (SVM-1) demonstrated high sensitivity (98.7%) and specificity (99.0%) in identifying respiratory events at the window level. The second-stage classifier (SVM-2) correctly classified OSA presence with 100% accuracy in the training set. When applied to the validation set, the combined model achieved a sensitivity of 72.3%, specificity of 73.3%, and F1-score of 73.1%. Performance was impacted by arrhythmias, and lack of direct respiratory effort measures.

A machine-learning model using routinely collected ECG and SpO₂ data shows promise for SDB detection but may require additional cardiopulmonary parameters for clinical decision-making in the ICU. Further validation with data collected in real-world ICU settings is warranted.