Obstructive sleep apnea screening from unprocessed ECG signals using statistical modelling

This paper introduces a novel feature generation method using only the time-domain representation of electrocardiogram (ECG) signals to detect obstructive sleep apnea (OSA) based on statistical modelling. It is shown that ECG segments have heteroskedastic properties. Therefore, the autoregressive integrated moving average and exponential generalized autoregressive conditional heteroskedasticity (ARIMA-EGARCH) model for their description, which can capture this characteristic correctly, is used to describe them. Initially, ECG signals are divided into 1 min segments. To show that ECG segments are heteroskedastic, the ARCH/GARCH test is applied. Then, ARIMA-EGARCH parameters are estimated from these segments using maximum likelihood estimation. The efficiency of the proposed method is assessed using five different classifiers: support vector machine, artificial neural network, quadratic discriminant analysis, linear discriminant analysis, and k-nearest neighbor. To evaluate the proposed approach, 34 single-lead ECG signals from the Physionet Apnea-ECG database are used. Experimental findings show that using ARIMA-EGARCH coefficients as a feature vector make it possible to classify apneic and normal ECG segments, and the new ARIMA-EGARCH parameter-based method achieves a performance comparable to other approaches, while using only eight features. Using the cross-validation approach, the accuracy of the proposed method is 81.43% and 97.06% for per-minute and per-subject classification, respectively. The method is particularly promising because no transformation is applied to the ECG signals, which can enable its application to the diagnosis of other diseases. In addition, it can be effectively implemented in-home monitoring systems owning to its low computing load.

Automated summary

This paper presents a novel method for feature generation using the time-domain representation of ECG signals to detect obstructive sleep apnea. The method models ECG segments and utilizes the ARIMA-EGARCH model to describe their heteroskedastic properties, achieving performance comparable to other approaches with only eight features. The proposed method shows promising accuracy for per-minute and per-subject classification and has potential for application in diagnosing other diseases due to its low computing load.