TASC: a time-aware sequence clustering framework with uncertainty quantification for electronic health record trajectories.

Longitudinal electronic health record (EHR) trajectories are highly heterogeneous, sparse, and irregular, making unsupervised temporal pattern discovery and uncertainty quantification challenging. We developed a Time-Aware Sequence Clustering (TASC) framework that integrates temporal spacing, clinical semantic similarity, and probabilistic subtype characterization to identify longitudinal trajectory patterns from EHR data. We demonstrate the framework using pre-surgical musculoskeletal care trajectories preceding total knee replacement (TKR). As a proof-of-concept application, we applied TASC to EHR and survey data from 2,052 patients undergoing primary TKR in the All of Us Research Program. Using electronic health records and survey data, we constructed their temporally ordered sequences of musculoskeletal-related diagnoses and comorbidities. We applied a weighted edit-distance algorithm that incorporates both clinical semantic similarity and inter-event time gaps, followed by K-Medoids clustering and subsampling-based stability analysis. Clusters were compared across sociodemographic variables, quality-of-life measures, prevalence of musculoskeletal-related diagnostic events, time to surgery, and pre- and post-surgical opioid use. We further quantified cluster membership uncertainty by computing probability-based subtype assignments from distances between partial or full trajectories and cluster medoids using a SoftMax transformation. TASC identified stable and interpretable trajectory structures from heterogeneous longitudinal EHR data in patients undergoing TKR. A five-cluster solution achieved the best silhouette score of 0.22 and demonstrated good stability under subsampling with mean Adjusted Rand Index of 0.71. The resulting trajectory clusters demonstrated distinct sociodemographic profiles, diagnostic/comorbidity patterns, and time-to-surgery distributions, supporting the framework's ability to capture heterogeneous longitudinal care-utilization patterns. Patients with fast progression (median = 252 days) to surgery were younger (mean = 53.6 years, SD =16.5 years) and had minimal comorbidity, whereas joint pain-obesity (median = 766 days) and joint pain-fracture (median = 1154 days) subtypes showed intermediate progression timelines. The soft-tissue-dominant subtype was characterized by prolonged pre-surgical trajectories (median = 2900 days) and elevated chronic opioid use, while the complex, chronic subtype comprised older patients (mean = 62.7 years, SD = 10.0 years) with multimorbidity, long delays to surgery (median = 5457 days), and the lowest post-surgical quality-of-life scores. Opioid use declined after surgery across all clusters, with the largest reductions observed in clusters with less complexity and shorter pre-surgical trajectories (e.g., fast progression to surgery and joint pain-obesity subtypes, 38% and 31% reduction, respectively). Probability-based analysis showed that fast progression subtype was identifiable early and with high confidence, joint pain-obesity and joint pain-fracture subtypes reflected risk patterns already present in earlier musculoskeletal history, whereas soft-tissue-dominant and complex chronic subtypes required longer longitudinal care records to become distinguishable. TASC provides a flexible framework for uncertainty-aware clustering of longitudinal EHR trajectories and demonstrates how temporal and semantic sequence information can reveal heterogeneous care-utilization patterns in musculoskeletal disease. These findings support future applications of TASC to other longitudinal clinical domains and event representations.