Purpose: This review aims to understand the clinical relevance of time series clustering analysis in the inpatient setting and the methodological decisions involved in applying time series clustering to inpatient data.

Background/significance: With a rise in machine learning research and improved accessibility of large time series data from inpatient hospital records, there has been recent interest in clinical phenotyping using time series clustering algorithms to capture dynamic heterogeneity that is often lost in cross-sectional analyses. Phenotypes derived from dynamic features (e.g., multiple time points) rather than static features (e.g., single time point) have been shown to better predict patient outcomes such as illness severity, treatment response, and mortality. Currently, there is a lack of literature that summarizes and synthesizes the existing research on time series clustering using data from hospitalized patients.

Methods: An extensive literature review search of the PubMed database was performed to identify relevant articles meeting inclusion criteria. The primary inclusion criteria were as follows: studies must analyze hospitalized patient data obtained from electronic health records (EHR) and studies must involve phenotyping or clustering of longitudinal data. Articles were excluded if they did not use time series data, were not derived from the EHR, involved genetic or epigenetic research, or were not an inpatient population. All articles meeting the inclusion criteria were analyzed, and their findings were synthesized to derive further insights.

Results: The initial search resulted in 2,568 articles. A total of 23 articles were identified for final inclusion. During data extraction and synthesis, four key areas were examined in each article: 1) type of time series data modeled, 2) patient population, 3) clustering approach applied, and 4) outcome variables used. Themes were then identified across these areas. Studies in this review used routinely collected hospital data for temporal phenotyping, primarily analyzing three types of time series data: vital signs and physio-markers, lab values, and organ dysfunction or severity scores. Most studies applied traditional clustering methods to identify clinical phenotypes, including group-based trajectory modeling, partitioning, hierarchical and density-based clustering. A smaller subset used advanced methods like hidden Markov models (HMMs) and deep learning techniques. Descriptive analyses of these clusters were then performed to characterize phenotypes and predict patient outcomes. These outcomes were grouped into three main categories, mortality, length of recovery, and physiological response, to evaluate the clinical relevance of identified clusters.

Conclusion/implications: The studies reviewed demonstrate a diverse range of clinical relevance, attesting to the clinical utility and widespread application of these techniques. Additionally, methodological decisions widely vary. These decisions impact the resulting clinical phenotypes identified through clustering techniques, which must be fully explored before integrating this type of technology into clinical decision making. Further investigation into the impact of methodological decisions of time series clustering is essential to support clinicians in leveraging routinely collected inpatient data as clinical decision support. This review provides a comprehensive understanding of current practices, informing both researchers and clinicians about the clinical utility of this analytical approach.

Reanna Panagides MSDS, BSN, RN

Appearances