以慣性感測器配合監督式機器學習量測日常活動時人體動態平衡

Abstract

跌倒是全球意外性傷害死亡的第二大原因，會導致行動受限、功能能力降低、失去獨立性和生活品質下降。失去平衡是跌倒主因，然而目前不存在能長期監控動態平衡狀況的系統。利用可穿戴式慣性感測器數據進行平衡控制的機器學習評估是一個相對較新的研究領域。大多數研究集中在跌倒檢測上，但很少涉及到動態平衡。機器學習技術的應用是處理多通道、高度非線性慣性感測器數據進行平衡監測的可行方法。

本研究透過人工智慧訓練和分類，由慣性感測器的資料特徵分類出不同動作，預測出動態平衡參數，再將動態平衡參數映射到標準靜態平衡控制參數。開發出一個創新、個人化的可穿戴式平衡監測系統，該系統整合人工智慧，用於長期監測在各項日常活動中的平衡控制。並且在動作分類上達成89%的分類準確度，在身體質量中心(COM)-足底壓力中心(COP)的傾角(IA)的估算上相對均方根誤差(rRMSE)達到9.34%、10.54%，並且對傾角變化率(RCIA)也達到15.04%及9.92%。最後則是對於計算動態平衡及其等效之靜態平衡時有約19.7%的誤差，並且其誤差和真實結果之間的皮爾森相關係數為0.59。

本研究所提出之方法為一創新之方法，尤其對於動態平衡之等校靜態平衡計算上也是無先例，因此即使在某些部分相對沒有那麼高的準確度，也依然能夠提供未來我們在慣性感測器對於量測日常活動時人體動態平衡這項主題上更多的參考以及潛能。

Falling is the second leading cause of accidental injury death worldwide and can result in restricted mobility, decreased functional ability, loss of independence, and reduced quality of life. Loss of balance is a major cause of falls. However, there is no system available to monitor the dynamic balance condition over the long term. Machine learning using wearable inertial sensor data for balance control evaluation is a relatively new field of study. The application of machine learning techniques is a feasible method for monitoring balance using multi-channel, highly nonlinear inertial sensor data. This study classified different movements based on data features from inertial sensors through artificial intelligence training and classification, predict dynamic balance parameters, and then map these parameters to standard static balance control parameters. This will lead to the development of an innovative, personalized wearable balance monitoring system that integrates artificial intelligence and is used for long-term monitoring of balance control in various daily activities. The accuracy of classification is up to 89%. Meanwhile, the was 9.34%、10.54% for sagittal and frontal IAs, and 15.04%及9.92% for the RCIAs. The results of the equivalent static balance corresponding to the dynamic balance showed an error of 19.7%, where the Pearson coefficient was 0.59. The method proposed in this study is an innovative approach, particularly with regard to the calculation of dynamic balance and isometric static balance. Even though certain aspects may exhibit relatively lower accuracy, it still offers valuable references and potential for future research in the area of human dynamic balance during the measurement of daily activities using inertial sensors.