運用穿戴式裝置與機器學習建構慢性病的精準健康管理服務

Abstract

本研究提出一個具可擴充性之精準健康服務，主要用於慢性病患的健康促進和異常事件預防。透過整合穿戴式裝置、開放環境數據、室內空氣品質感測裝置、背景偵測定位之智慧型手機應用程序以及人工智慧輔助決策個案管理平台，實踐對慢性病患的生活型態、生活環境、生活軌跡與臨床評估資料等資訊全天候即時監測。而人工智慧輔助決策個案管理平台可提供給專業醫療人員，其照護之慢性病患最詳盡的院外整合型風險資訊，同時基於每日所蒐集之即時資料與過往疾病發展史，平台將產出個人化的異常事件風險預測值，供醫護人員決策輔助之用，期以達到對於不同慢性病族群的異常事件早期偵測與即時提醒介入之目的。 在24個月的研究期間，本研究前瞻地收集了 1,767 名慢性病患者的所有數據，蒐集之數據被用於訓練模組化的慢性疾病預測模型。機器學習為主要訓練預測模型的演算法基礎，截至目前通過驗證的模組化慢性疾病預測模型包括肥胖症、恐慌症、大腸癌和慢性阻塞性肺病，驗證後的預測模型平均準確率可達84.8%，靈敏度達75.5%，F1-score達80.6%，具備模型可信度與公平性。 與過去的研究相比，本研究建立了一種有效收集生活型態、生活環境、生活軌跡和病症記錄的方法，同時透過添加客觀的生活型態與生活環境數據，大幅提高了預測模型的預測成效。此外研究結果也闡述了生活型態和生活環境因素與慢性病患的健康與異常事件發生高度相關；全面性資料特徵項收集輔以訓練預測模型的方法相較於以往多數研究僅利用問卷數據的方法具備更好的預測能力，有較大的潛力達到準確地預測未來異常事件、提供個人化健康促進建議與照護服務。此外本研究為將模型落實於真實世界，也另建立具成本效益的預測模型，用最少特徵項完成預測任務，對於真實世界的AI預測模型部署幫助顯著。

This thesis proposes an integrated and scalable precision health service to prevent chronic diseases and promote overall health. Combining the wearable technology, open environmental data API, air quality sensors, an app with location provider, and an AI-based telecare platform, we implement the monitoring on the environmental and lifestyle factors continuously. The AI-based telecare platform provides in-depth analysis of patients' lifestyle, environmental and medical records and can accurately predict abnormal events of chronic diseases. During a 24-month follow-up period, data were collected from 1,767 patients, revealing 416 abnormal episodes. Machine learning algorithms were used to train modular chronic disease models, which were externally validated and demonstrated an average accuracy of 84.8%, a sensitivity of 75.5%, and an F1 score of 80.6% for the prediction of obesity, panic disorder, colorectal cancer, and chronic obstructive pulmonary disease, respectively. Our approach effectively collects lifestyle, life trajectory, symptom records, and environmental factors compared to previous studies. Incorporating objective, comprehensive data and using feature selection techniques can improve the performance of prediction models. In addition, our results also indicate that lifestyle and environmental factors strongly correlated with patient health and may be more effective at predicting future abnormal events than using questionnaire data alone. Additionally, we have proposed a cost-effective prediction model that requires only several features to complete the prediction task, making it suitable for deploying the real-world modular prediction model.