運用穿戴裝置與機器學習預測安寧緩和醫療中的急性與死亡事件

Abstract

在安寧緩和的照護中，緊急的臨床狀況改變以及死亡都是常見而不可避免的。然而，這些事件的不確定性，包括發生的時機、症狀之輕重、是否需要至醫院接受及時照護，常會讓病人與家屬無所適從。因此對於在接受門診或是居家安寧緩和照護的病人，會下降其生活品質並帶來不安全感。然而目前預測這些事件並沒有很好的預測方式，運用新興科技如穿戴裝置或人工智慧來協助的臨床研究證據也不多。 本研究為一前瞻性研究，旨在探索將穿戴式裝置和人工智慧應用於安寧緩和癌症患者的照護上，預測死亡和緊急醫療需求的可能性。收案對象為於臺大醫院門診、病房或居家醫療接受安寧緩和照護的末期癌症患者。每位參與者均配備智慧手環，收集包括步數、心率、睡眠時間和血氧飽和度等生理數據。研究團隊同時每週進行臨床評估，追蹤持續到受試者死亡或是最長52週的時間。生理數據與臨床評估被合併為一資料集，並在資料前處理後利用基於機器學習的分類器和深度神經網絡模型，預測7天內的死亡或緊急醫療需求事件，評估指標包括接受者操作特徵曲線下面積、F1分數、準確度和特異度，之後並針對表現較好的模型進行Shapley值分析，以進一步了解人工智慧於預測上的判斷基礎。 於2021年9月至2022年8月期間，本研究共納入40名患者，總共收集了1657個數據點。在所使用的模型中，包括極限梯度提升、深度學習、隨機森林和K近鄰演算法在內的4個模型，在預測死亡事件和緊急醫療需求呈現出不錯的表現。其中，與死亡預測相關的關鍵特徵包括平均心率、每日行走步數、臨床功能、意識變化和臨床照護階段評估。對於緊急醫療需求的預測，重要特徵包括疼痛和睡眠品質，以及當天的最低心律。 本研究的結果顯示了將穿戴式裝置和人工智慧整合到安寧緩和照護之潛力，並透過可解釋性人工智慧為末期病患的病程變化提供可能的洞見。未來也需要進一步的研究，持續驗證模型之可靠性，並評估其實際應用對臨床照護品質的影響。

Death or emergent events are inevitable in the clinical course of terminal patients, but their uncertainty can diminish quality of life and create feelings of unsafety for patients and their families, especially those who are receiving outpatient or home-based care. Predicting these events is important yet challenging in palliative care. Limited evidence exists on the use of artificial intelligence (AI) and wearable devices in palliative cancer patients. This prospective study aimed to explore the potential of using wearable devices and artificial intelligence (AI) to predict death events and emergent medical needs in terminal cancer patients. Participants diagnosed with cancer and receiving palliative care at National Taiwan University Hospital were enrolled between September 2021 and August 2022. Each participant was provided with a smartwatch to collect physiological data, including steps taken, heart rate, sleep time, and blood oxygen saturation. Weekly clinical assessments were conducted, and participants were followed until the end of life or up to 52 weeks. Machine learning-based classifiers and deep neural network (DNN) models were employed to predict events occurring within 7 days, with evaluation metrics including area under the receiver operating curve (AUROC), F1 score, accuracy, and specificity. Shapley value analysis was performed to gain further insights into the models' performance. The study included a total of 40 patients, with 1657 data points collected throughout the study period. Among the models examined, including Extreme-gradient boost (XGBoost), DNN, Random Forest (RF), and K-nearest neighbors (KNN), fair performances were observed in predicting both death events and emergent medical needs. Key features associated with death prediction included average heart rate, steps taken, functional level, clinical care phase, and a change in consciousness level. For the prediction of emergent medical needs, minimal heart rate on the day, pain, and sleep quality were identified as significant features. These findings highlight the potential of integrating wearable devices and AI into clinical palliative care, providing valuable insights for personalized care at the end-of-life stage. Further research is needed to validate these findings and assess the impact on clinical care.