利用貝氏主題分群方法探索來自穿戴式裝置之自由生活身體活動資料中的活動模式

Abstract

透過自由生活身體活動的分析讓研究者有機會探索身體活動與疾病或健康事件間的關係，然而透過穿戴式裝置所蒐集的的活動資料並無活動類型或標籤的相關訊息。應用非監督分群分析可以讓我們有機會探索可能的活動類型或標籤。因此，基於曲線形狀的非監督分群分析在這類型的科學應用中至關重要。然而，現有函數資料分群方法常側重於振幅(amplitude)差異，卻忽略了對身體活動(physical activity)資料分析影響顯著的相位(phase)變化。識別活動曲線之間的相似性需要同時考量相位與振幅的差異變化。儘管分群分析在發現活動模式方面具有潛力，但相關研究仍然有限。為彌補這些不足，本研究提出一種貝氏分群方法，從穿戴式裝置採集的自由生活身體活動資料識別出不同的活動模式(motif)。 我們將24小時活動曲線分割成固定時間段。接著，應用彈性函數資料分析，透過彈性距離(elastic distance)矩陣量化曲線之間的差異性，並將相位與振幅距離的權重總和定義為曲線間的差異距離，再分別使用Von Mises distribution和Gamma distribution建模。應用Dirichlet process貝氏無母數分群方法(Bayesian nonparametric cluster method)架構，得到分群結果與分群數量的自動推斷。最後識別出的群集，可以為進一步分析身體活動資料時，用於定義新數位生物標記的選擇。 我們透過實際資料的應用，驗證所提出方法的表現。各群集皆展現出不同的特徵，其中部分群集較依賴振幅距離作為群集特徵，而另一些則較依賴相位距離。本研究提出的貝氏函數資料分群方法，透過後驗分析確定最佳群集數量，不需要預先指定群集數量。此方法將透過應用實際資料來進行驗證。此架構為函數資料分析提供了一個穩健的data-driven分群方法，有助於關聯性研究，並透過揭示有意義的活動模式來加強健康事件研究。

Analyzing free-living physical activity (PA) offers researchers a valuable opportunity to explore the relationship between PA and various health outcomes. However, a significant challenge arises from the free-living PA data collected by wearable devices: it often lacks information regarding activity type or labels. To address this, applying unsupervised cluster analysis becomes crucial, allowing researchers to identify potential activity types or labels within the unlabeled PA data. Hence, unsupervised cluster analysis on curve shapes is a significant problem for this scientific application. However, many functional clustering methods focus on amplitude differences, neglecting the considerable impact of phase variations, particularly in physical activity data. Analyzing the similarity between two activity curves necessitates considering phase and amplitude variations for meaningful insights. Moreover, research discussing or proposing methods for studying PA data through cluster analysis to identify activity patterns remains limited. Hence, this study aims to develop a novel Bayesian motif-based clustering method to uncover distinct activity patterns (motifs) within free-living PA data collected from wearable devices. Initially, we segment the 24-hour activity curve into fixed-time intervals, using the fundamental time unit of an activity as the basis for segmentation. Subsequently, elastic shape analysis is employed for these activity segments, and elastic distance is utilized to quantify curve dissimilarity. This curve dissimilarity is decomposed into phase and amplitude distance components, which are modeled using the Von Mises and Gamma distributions, respectively. A Bayesian nonparametric clustering framework with a Dirichlet process was proposed. We derive cluster results and the posterior distribution, thereby inferring the posterior distribution of the number of clusters. Finally, these identified activity clusters could be used to define new digital biomarkers as PA features for further analysis. This study developed a novel Bayesian functional clustering methodology for activity pattern discovery, leveraging elastic functional data analysis to compare activity curves and eliminating the necessity for pre-specifying the number of clusters. By employing a flexible prior on the space of data partitions and analyzing the resulting posterior distribution, our approach will effectively determine the optimal clustering configuration. Each cluster exhibits distinct characteristics, with some relying more on the amplitude distance component, while others are more dependent on the phase distance component. We validated the performance of our proposed method through real-world applications. We hope this framework will provide a practical solution for real-world datasets in prospective applications. This method will facilitate data-driven partitioning within functional data analysis, thereby making a significant contribution to ongoing research, such as association studies, by enabling the discovery of meaningful patterns and relationships within health events. Moreover, it is hypothesized that the motifs identified through this method can serve as a foundational basis for establishing digital biomarkers, subsequently advancing research in physical activity analysis.