恐慌發作的預測與可解釋模型：運用穿戴式裝置及機器學習之研發與世代研究

Abstract

提供恐慌症患者睡眠和身體活動建議是相當重要的，但過去卻較少被仔細研究。本研究是一項三年期前瞻性世代研究及研發計畫，目標在預測未來七天內的恐慌發作（PA）、狀態焦慮（SA）、特質焦慮（TA）以及恐慌症嚴重程度（PDS）。我們從一家綜合醫院納入了148名恐慌症患者，使用DSM-5、MINI和臨床評估問卷：BDI、BAI、PDSS-SR、STAI以及Garmin智慧手環，從2020年6月16日至2023年6月10日記錄他們的每日睡眠、身體活動和心跳。團隊應用RNN、LSTM、GRU深度學習和SHAP可解釋模型來分析數據。有99人完成了本研究。從LSTM（長期和短期記憶）模型計算的7天預測準確率分別為PA、SA、TA、PDS：92.8％、83.6％、87.2％ 和75.6％。透過SHAP可解釋模型，PA較容易復發於：初始BDI及BAI分數較高、共病憂鬱症、泛焦慮症或懼曠症的病人。然而，PA在以下情況有所下降：每日平均心跳在72-87次/分鐘區間、每日爬升大於9層樓、總睡眠時間介於6小時23分至10小時50分、以及深度睡眠期大於50分鐘。綜合以上，本試驗透過機器學習與深度學習分析了穿戴式手環和問卷資料，預測七天後恐慌發作的正確機率約為75.6-92.8%，而且，研究顯示充足的睡眠以及每日爬升樓層數可降低PA復發機率。此外，運用個案管理合併上述預測及回饋系統，相較於常規治療，顯著改善了恐慌症維持期治療 (p < 0.05) 的結果。

Sleep and physical activity suggestions for panic disorder (PD) patients are critical but less surveyed. This three-year prospective cohort study aims to predict panic attacks (PA), state anxiety (SA), and trait anxiety (TA), panic disorder severity (PDS) in the upcoming seven days. We enrolled 148 PD patients from one general hospital. We used DSM-5, MINI, and clinical app questionnaires: BDI, BAI, PDSS-SR, STAI, and wearables, recording their daily sleep, physical activities, and heart rates from June 16, 2020, to June 10, 2023. Our teams applied RNN, LSTM, GRU deep learning, and SHAP explainable methods to analyze the dataset. Ninety-nine completed this study. The 7-day prediction accuracies for PA, SA, TA, and PDS were 92.8%, 83.6%, 87.2%, and 75.6%, respectively, from LSTM (Long and Short-Term Memory) model. By SHAP explainable model, higher initial BDI.BAI score, comorbidities with depressive disorder, general anxiety disorder, or agoraphobia predict increased chances of PAs. However, PA decreased in the following conditions: daily average heart rates fall between 72-87 bpm, daily floors climbing more than 9 floors, total sleep duration between 6h23m-10h50m, deep sleep over 50m. Deep learning provides a 75.6-92.8% accuracy of 7-day PA prediction through smartwatch and questionnaire data. Recurrent PA chances decrease if participants' daily sleep time falls between 6h23m-10h50m or daily floors climb more than nine floors. Furthermore, case management with this system improves maintenance treatment (p < 0.05) compared with treatment as usual.