使用穿戴式裝置資料運用人工智慧模型預測雙相情緒障礙症病患之情緒徵兆

Abstract

在本研究中，我們將使用穿戴式裝置 (Garmin Vivosmart 4) 收集到的數位生物標誌物，其中包含心率、活動和睡眠狀態，通過機器學習演算法建立雙相情緒障礙症的預測模型。我們可以根據可解釋模型得出的見解提出治療建議或防止複發。 參與者為 20 至 65 歲被診斷為雙相情緒障礙症的患者，而雙相情緒障礙症在美國精神病學協會的精神疾病診斷和統計手冊 (Diagnostic and Statistical Manual of Mental Disorders)第五版中的描述為一種導致人的情緒、能量和功能能力極度波動的腦部疾病。使用到的數據包含兩部分：問卷數據和數位生物標誌物。問卷數據包含貝克憂鬱量表（Beck Depression Inventory）和楊氏躁症量表（Young Mania Rating Scale），它們分別用於評估鬱症和躁症的發作，數位生物標誌物則用於模型構建。我們使用了六種機器學習演算法來構建預測模型。 本研究共招募了24名參與者。鬱症的預測模型在測試集上達到了準確率86%、Area Under the Receiver Operating Characteristic curve 0.85和F1 score 0.56。通過可解釋模型 Shapely Additive exPlanations，我們發現相對較高的靜止心率、低活動和睡眠不足與鬱症相關，可能可以預測鬱症的發生。然而，相較於鬱症的發作，躁症的發作預測準確度較低。 總結來說，在本研究中，我們使用了從穿戴設備收集的數位生物標誌物來構建機器學習模型，該模型可以預測幾天後自我報告的憂鬱和躁症症狀。除此之外，我們還可以利用從可解釋模型中獲得的資訊來更早地提供臨床評估和治療，以降低復發風險。

In this study, we would use digital biomarkers collected by wearable devices (Garmin Vivosmart 4), including heart rate, activity, and sleep status, to build prediction models for the recurrence of manic or depressive symptoms in bipolar disorder with machine learning algorithms. Moreover, we could make treatment recommendations or prevent recurrence based on insights derived from the interpretable model. Participants were 20-65 years old patients diagnosed with bipolar disorder (BD), as described by the American Psychiatric Association’s Diagnostic and Statistical Manual of Mental Disorders (DSM-5). Data contained two parts: questionnaire data and digital biomarkers. Questionnaire data of the Beck Depression Inventory (BDI) and Young Mania Rating Scale (YMRS) were used to evaluate depressive and manic symptoms, respectively, and digital biomarkers were for model building. Six machine learning algorithms were used to construct prediction models. A total of 24 participants with BD were recruited. The prediction model for depressive symptoms achieved 86% accuracy, 0.85 AUROC, and 0.56 F1 score on testing data. With interpretable model Shapely Additive exPlanations (SHAP), we found that relatively high resting heart rate, low activity, and lack of sleep were associated with and may predict depressive symptoms. However, compared with predicting a depressive episode, the accuracy of predicting a manic episode was lower. In this study, we used digital biomarkers collected from wearable devices to construct machine learning models which could predict self-report depressive and manic symptoms several days later. Furthermore, important features derived from the interpretable model may provide insight for early detection of mood symptoms recurrence and reduce the risk of recurrence.