Deep Neural Network Algorithm Using the Electrocardiogram for Detection of Obstructive Sleep Apnea

Abstract

Background

Although highly prevalent, obstructive sleep apnea (OSA) remains largely underdiagnosed, thus justifying the need for high-performing screening tools.

Objectives

The authors sought to develop a machine learning–powered algorithm to identify OSA from the 12-lead electrocardiogram (ECG), a routine clinical test.

Methods

A retrospective population of 11,299 patients who completed sleep evaluation and underwent 12-lead ECG at Mayo Clinic were included. OSA was defined as an apnea-hypopnea index ≥5. A deep convolutional neural network model was constructed to detect OSA from the ECG (artificial intelligence [AI]-ECG). Predictive performance of the algorithm in the total sample and separately in males and females was evaluated using the receiver-operating characteristic curve with area under the curve (AUC).

Results

The population consisted of 7,170 patients with OSA and 4,129 controls (53.7% males, median [Q1-Q3] of age 58 [47-68] years). The AUC of the AI-ECG model for identification of OSA in the test sample was 0.80 (95% CI: 0.77-0.83), with accuracy, sensitivity, and specificity of 73.7%, 77.0%, and 68.6%, respectively. The model showed better discriminatory performance in females (AUC: 0.82; 95% CI: 0.79-0.86) than in males (AUC: 0.73; 95% CI: 0.68-0.78; P < 0.001). Sensitivity analyses showed that the predictive abilities of the model were robust across different time intervals and even when including ECG recordings manifesting cardiac abnormalities.

Conclusions

Our AI-ECG model demonstrated good diagnostic performance as an ECG-based screening tool for OSA in a clinical population, particularly among females. Incorporating this algorithm in medical practice may enable widespread low-cost screening for OSA, optimizing early diagnosis and therapy.