Diagnosis of obstructive sleep apnea in children based on the XGBoost algorithm using nocturnal heart rate and blood oxygen feature.

PURPOSE: Obstructive sleep apnea (OSA) is a serious type of obstructive sleep-disordered breathing (SDB) that can cause a series of adverse effects on children's cardiovascular, growth, cognition, etc. The gold standard for diagnosis is polysomnography (PGS), which is used to assess the prevalence of OSA by obtaining the apnea-hypopnea index (AHI), but this diagnosis method is expensive and needs to be performed in a specialized laboratory, making it difficult to be of benefit to children with suspected OSA on a large scale. Our goal was to use a machine learning method to identify children with OSA of varying severity using data on children's nighttime heart rate and blood oxygen data. METHODS: This study included 3139 children who received diagnostic PSG with suspected OSA. Age, sex, BMI, 3 % oxygen depletion index (ODI), average nighttime heart rate and fastest heart rate were used as predictive features. Data sets were established with AHI ≥ 1, AHI ≥ 5, and AHI ≥ 10 as the diagnostic criteria for mild, moderate and severe OSA, and the samples of each data set were randomly divided into a training set and a test set at a ratio of 8:2. An OSA diagnostic model was established based on the XGBoost algorithm, and the ability of the machine learning model to diagnose OSA children with different severities was evaluated through different classification ability evaluation indicators. As a comparison, traditional classifier Logistic Regression was used to perform the same diagnostic task. The SHAP algorithm was used to evaluate the role of these features in the classification task. RESULTS: We established a diagnostic model of OSA in children based on the XGBoost algorithm. On the test set, the AUCs of the model for diagnosing mild, moderate, and severe OSA were 0.95, 0.88, and 0.88, respectively, and the classification accuracy was 90.45 %, 85.67 %, and 89.81 %, respectively, perform better than Logistic Regression classifiers. ODI is the most important feature in all classification tasks, and a higher fastest heart rate and ODI make the model tend to classify samples as positive. A high BMI value caused the model to tend to classify samples as positive in the mild and moderate classification tasks and as negative in the classification task with severe OSA. CONCLUSION: Using heart rate and blood oxygen data as the main features, a machine learning diagnostic model based on the XGBoost algorithm can accurately identify children with OSA at different severities. This diagnostic modality reduces the number of signals and the complexity of the diagnostic process compared to PSG, which could benefit children with suspected OSA who do not have the opportunity to receive a diagnostic PSG and provide a diagnostic priority reference for children awaiting a diagnostic PSG.