Scholarlink

Introduction <p>Obstructive Sleep Apnea (OSA) is a highly prevelance sleep breathing disorder that imposes, significant public health and economic burdens through its untreated associated comorbidities. The gold standard for diagnosing OSA, polysomnography (PSG), retains several limitations because it is cumbersome to conduct in the sleep lab and requires labor-intensive efforts to annotate the measurement. Furthermore, current research focuses on the alternative of PSG on the general OSA population without specifically considering the OSA comorbid conditions, such as cardiovascular and cancer diseases. There is a need for methods to monitor OSA using features derived from consumer sleep technologies (CSTs), considering comorbid conditions associated with OSA.</p> Methodology <p>In this paper, we identify the features that can be collected from CSTs to predict the patient’s Apnea-Hypopnea Index (AHI) with the consideration of the comorbid chronic diseases, including cardiovascular disease and cancer. These sleep characteristics include the total recording time, the sleep period time, total sleep time, the onset of sleep, the efficiency of sleep, the wakefulness after the onset of sleep, the percentages of N1, N2, N3, REM sleep stages, the latency of REM from sleep onset and the latency of REM from lights off.</p> Results <p>Based on these features, we build stepwise regression, LASSO, and XGBoost models to evaluate the risk of a patient having OSA and classify individuals as normal versus OSA, achieving classification accuracies of 68.76%, 69.26%, and 70.4%, respectively. The proposed method has been validated based on the Wisconsin Sleep Cohort Database. To test the generalization of the method, the model was validated on the IRB-approved Sanford Obstructive Sleep Apnea Cancer dataset. XGBoost demonstrated superior performance with the highest Recall (0.9136) and F1 Score (0.7647), while regression and Lasso excel in interpreting the features with the Precision (0.865).</p> Conclusion <p>Such performance highlights their effectiveness in identifying OSA cases with comorbidity. Our proposed solutions introduce better interpretability of algorithms and improve the accuracy of OSA detection using CSTs while addressing the limitations of PSG.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Predicting the Apnea-Hypopnea index for the diagnosis of obstructive sleep apnea with cardiovascular and cancer comorbidity using wearable sleep-tracking devices

  • 期刊论文

  • 发表日期：

    2026年2月7日

查看全文
  • Hon Keung Tony Ng,
  • Uyen Nguyen,
  • Yusuf Akbulut,
  • Thuan Tran,
  • Toan Vo,
  • Harun Pirim,
  • Arveity Setty,
  • Trung Le

摘要

Introduction

Obstructive Sleep Apnea (OSA) is a highly prevelance sleep breathing disorder that imposes, significant public health and economic burdens through its untreated associated comorbidities. The gold standard for diagnosing OSA, polysomnography (PSG), retains several limitations because it is cumbersome to conduct in the sleep lab and requires labor-intensive efforts to annotate the measurement. Furthermore, current research focuses on the alternative of PSG on the general OSA population without specifically considering the OSA comorbid conditions, such as cardiovascular and cancer diseases. There is a need for methods to monitor OSA using features derived from consumer sleep technologies (CSTs), considering comorbid conditions associated with OSA.

Methodology

In this paper, we identify the features that can be collected from CSTs to predict the patient’s Apnea-Hypopnea Index (AHI) with the consideration of the comorbid chronic diseases, including cardiovascular disease and cancer. These sleep characteristics include the total recording time, the sleep period time, total sleep time, the onset of sleep, the efficiency of sleep, the wakefulness after the onset of sleep, the percentages of N1, N2, N3, REM sleep stages, the latency of REM from sleep onset and the latency of REM from lights off.

Results

Based on these features, we build stepwise regression, LASSO, and XGBoost models to evaluate the risk of a patient having OSA and classify individuals as normal versus OSA, achieving classification accuracies of 68.76%, 69.26%, and 70.4%, respectively. The proposed method has been validated based on the Wisconsin Sleep Cohort Database. To test the generalization of the method, the model was validated on the IRB-approved Sanford Obstructive Sleep Apnea Cancer dataset. XGBoost demonstrated superior performance with the highest Recall (0.9136) and F1 Score (0.7647), while regression and Lasso excel in interpreting the features with the Precision (0.865).

Conclusion

Such performance highlights their effectiveness in identifying OSA cases with comorbidity. Our proposed solutions introduce better interpretability of algorithms and improve the accuracy of OSA detection using CSTs while addressing the limitations of PSG.