使用穿戴式裝置資料運用人工智慧模型預測自閉症類群障礙症病患之情緒徵兆

Abstract

在這項研究中，我們使用從穿戴式設備（Garmin Vivosmart 4）收集的生理數據，其中包含心率、活動和睡眠時間的長度，通過機器學習演算法構建自閉症譜系失調的預測模型。我們可以根據可解釋模型的結果推薦治療方法或預防方法。 本研究招募了14名受試者。參與者是20至55歲的患者，他們被診斷患有自閉症類群障礙症，該疾病基於美國精神醫學會（American Psychiatric Association, APA）的精神疾病診斷和統計手冊（DSM-5）。使用的數據由兩部分組成。問卷數據包括貝克憂鬱量表（Beck Depression Inventory, BDI）和躁鬱症評定量表（Young Mania Rating Scale, YMRS），分別用於評估憂鬱症和躁症，以及用於模型構建的數位生物標記。我們使用 6 種機器學習演算法來構建預測模型。 預測模型在憂鬱發作的測試集上達到了79%的準確率，0.88 AUROC和0.88 f1得分。通過可解釋的模型SHAP我們發現相對較低的靜息心率、高活動與憂鬱症有關，並可能預測憂鬱症的發作。 總之，我們可以使用從可解釋模型中獲得的資訊來提供早期的臨床評估和一些預防治療。

In this study, we utilized digital biomarkers collected from wearable devices (Garmin Vivosmart 4), which include heart rate, activity level, and length of sleeping hours, to construct predictive models of Autism Spectrum Disorder using machine learning algorithms. Based on the results obtained from interpretable models, we can recommend preventive methods. Fourteen participants were recruited for this study. These participants were patients aged 20 to 55 who had received a diagnosis of ASD, according to the criteria outlined in the American Psychiatric Association's Diagnostic and Statistical Manual of Mental Disorders (DSM-5). The data used for analysis consisted of two parts: questionnaire data, including the Beck Depression Inventory (BDI) and the Young Mania Rating Scale (YMRS), used to assess depression and mania respectively, and digital biomarkers for model development. We employed six machine learning algorithms to construct a predictive model. The predictive model achieved 79% accuracy, 0.88 AUROC, and 0.88 F1 score on the test set of depressive episodes. Using the explainable model SHAP, we discovered that relatively low resting heart rate and high activity were associated with depression and could potentially predict the onset of depressive episodes. In conclusion, the information obtained from interpretable models can be used to provide earlier clinical evaluations for prevention.