利用快速傅立葉轉換進行穿戴式裝置資料特徵變數之選取

Abstract

近年來，穿戴式裝置被廣泛用於人類活動監測上。在早期的醫學研究中，裝置較常配戴於臀部上，而隨著科技的進度，腕部穿戴式逐漸受到歡迎。由於裝置攜帶方便且紀錄數值具有可信度，因此近年來被用於記錄身體活動。根據裝置中裝設的感測器不同，可紀錄的資料類型也不同，如：加速度、旋轉角度等。此類型資料在分析時需考量多種潛在問題，例如：由於此類型資料為長期追蹤資料，每個人在自身的紀錄中存在著自身的變異，如何在分析時將變異納入考量，以及如何比較不同廠牌裝置的資料等。本研究以頻域的角度分析穿戴式裝置資料，透過活動特徵選取降低資料量，並根據特徵的解釋性對於活動進行解讀及探討其應用的可能性。研究中使用三個資料集(英國人體生物資料庫、DEPRESJON資料集、PSYKOSE資料集)進行分析，其中英國人體生物資料庫中的穿戴式裝置資料透過R軟體GGIR套件(版本.2.3.0)進行品質管理和前處理，DEPRESJON資料集、PSYKOSE資料集則制定數個篩選標準，確保資料品質。本研究將對三個資料集進行活動特徵選取、建立廣義線性混合模型探討族群在平日和假日之間是否具有活動差異和建立族群分類模型。研究結果發現，在頻域中使用5%的週期波即可足夠代表原始的時序資料，且結果並不會因為手錶品牌、資料紀錄時間單位不同而有影響。在英國生物資料庫中，憂鬱症患者及健康受試者的活動受到平日和假日的影響，而性別、年齡、肥胖程度對於活動的影響是較為明顯的。在DEPRESJON資料集、PSYKOSE資料集中建立分類模型，結果可發現加入活動變數可提升模型表現力並可以達到較好的分類效果。由於本研究並無考量週期波之間可能存在的相關性，建議未來研究於頻域分析時可將此納入考量。

In recent medical research, wearable devices have been commonly used to monitor human activities. A wearable device could be placed on different parts of the body, previous studies mainly used hip-worn devices to collect data, but wrist-worn devices have become more popular, with their more reliable recordings and smaller device size, allow people to monitor their daily activities more conveniently. Wearable device with different sensors can record various data types, including acceleration, angular velocities, and orientations. However, many problems exist in this kind of data, including how to put the within-subject variability into consideration for longitudinal data analysis and how to compare the data collected by different brands of wearable devices. Our research focuses on analyzing wearable device data on frequency domain and uses tri-accelerometers data to propose a method that selects activity features containing information of amplitude and frequency, thus help explain the meaning of the activity features and build statistical models to explore activity patterns between different populations. We used three datasets (UK Biobank, DEPRESJON dataset, PSYKOSE dataset) for our analysis. The R package GGIR (ver.2.3.0) was used to perform quality control and UK Biobank acceleration data preprocessing, then we set up the selection criteria for the DEPRESJON dataset and the PSYKOSE dataset to ensure data quality. After data preprocessing, we then extracted features and used them to build generalized linear mixed effect models and classification models. Our results suggested that using 5% of periodic waves can represent the original time series data, and the results would not be affected by different brands of wearable devices and different time record units. In analysis on data from UK Biobank, we found out there exist different activity patterns between weekdays and weekends for major depressive disorder (MDD) patients and healthy participants. We also find that gender, age, and BMI are significant factors for human activity. The result for the classification models in the DEPRESJON dataset and the PSYKOSE dataset showed that activity features could improve the model performance and achieve better results. Our research did not consider the correlation between each periodic waves, which we recommend that it can be considered in future research.