A computational framework that integrates clinical semantics, temporal spacing, and uncertainty quantification has identified five distinct pre-surgical care trajectories among patients undergoing total knee replacement, according to a study published in BioData Mining. Researchers developed the Time-Aware Sequence Clustering system, designated as TASC, to resolve the irregular and sparse timing characteristic of longitudinal electronic health records. By evaluating diagnostic pathways recorded prior to primary arthroplasty across 2,052 patients from the All of Us Research Program, the investigation established that progression to knee surgery follows systematically diverse healthcare pathways rather than a uniform clinical course. The unsupervised sequence analysis mapped heterogeneous trajectories spanning from rapid surgical intervention occurring within hundreds of days to protracted clinical sequences extending over decades. These distinct trajectories were strongly correlated with patient sociodemographic factors, chronic opioid usage patterns, comorbidity accumulation, and self-reported post-surgical recovery measures, demonstrating that pre-operative care histories embody structured behavioral and clinical archetypes.
Musculoskeletal disorders rank among the leading global causes of physical disability, frequently culminating in total joint arthroplasty to alleviate persistent pain and restore joint function in advanced osteoarthritis. More than 790,000 total knee replacements are performed each year in the United States alone, yet clinical research has historically focused on cross-sectional surgical indicators or short-term post-operative recovery rather than multi-year progression patterns. Prior investigations regularly monitored isolated risk factors, standardized quality-of-life indices such as the EQ-5D, or immediate perioperative protocols under enhanced recovery pathways. Consequently, relatively little has been known regarding how patients accumulate diagnoses, navigate non-surgical conservative treatments, and enter healthcare systems over extended periods preceding surgical intervention. The study authors noted that understanding these multi-year pathways is essential for evaluating access disparities, structuring targeted prehabilitation programs, optimizing surgical timing, and understanding why individual post-operative recovery trajectories diverge so sharply across patient populations.
Extracting meaningful progression patterns from real-world health records presents severe algorithmic difficulties due to irregular encounter spacing, wide variations in record length, and overlapping diagnostic terminology. Standard clustering methodologies, including conventional K-means algorithms and Gaussian mixture models, depend upon fixed-length numerical vectors and cannot natively represent the temporal ordering, duration, or intermittent recurrence of clinical encounters. Meanwhile, classical string-matching algorithms, such as unweighted Levenshtein edit distance and basic dynamic time warping, treat medical diagnosis codes as unrelated arbitrary symbols without semantic context. In musculoskeletal medicine, clinical entities such as localized joint pain, osteoarthritis, obesity, and mechanical soft-tissue disorders frequently co-occur throughout long-term care pathways. Disregarding the semantic relationships between diagnostic terms risks obscuring genuine clinical similarities across patient groups, while ignoring temporal intervals between visits fails to distinguish acute progressive deterioration from decades of intermittent, low-intensity conservative management.
To overcome these computational limitations, the investigators utilized registered and controlled electronic health records and participant survey responses from the All of Us Research Program collected between May 2018 and October 2023. The research team identified an initial cohort of 2,178 adult participants living in the United States who underwent primary total knee replacement, selecting the earliest recorded procedure date as the defining index event. Longitudinal records were filtered to exclude diagnoses unrelated to musculoskeletal pathology or relevant medical comorbidities, establishing an analytic window spanning from each individual patient’s earliest documented musculoskeletal diagnosis up to the surgical date. After excluding individuals with fewer than two qualifying diagnostic entries to eliminate uninformative single-event records, the final study cohort comprised 2,052 patients. Survey data provided linked sociodemographic indicators, including health insurance status and living arrangements, alongside longitudinal post-operative quality-of-life measurements.
The TASC architecture models each individual medical trajectory as an interleaved sequence of clinical event tokens and temporal interval markers. Event tokens captured curated musculoskeletal diagnostic categories, including lower-limb osteoarthritis, joint disorders, and soft-tissue conditions, alongside systemic comorbidities such as falls, skeletal fractures, obesity, and vitamin D deficiency. To address repetitive encounter-based billing entries where chronic conditions are redundantly documented without clinical change, the pipeline retained only the initial onset and final recorded occurrence of each condition code, preserving total disease duration while reducing documentation noise. Between successive clinical events, the framework inserted discretized time-gap tokens, labeled G1 through G10, representing deciles of the empirical distribution of inter-event intervals measured in elapsed days. This representation ensured that both the chronological ordering of diagnoses and the biological tempo of care-seeking were embedded directly within each patient’s serialized sequence string.
Pairwise sequence dissimilarity was computed using a customized weighted edit-distance algorithm that established variable costs for insertion, deletion, and token substitution operations. While insertions and deletions carried a uniform penalty of one unit, substitution costs between diagnosis tokens were mathematically derived from semantic embeddings trained on the complete sequence corpus using a continuous Word2Vec architecture. The substitution penalty between two distinct diagnostic events was calculated as one minus the cosine similarity of their learned vector representations, allowing clinically related conditions to incur negligible penalties when aligned. Time-gap substitutions were scored according to the absolute difference between their respective decile ranks in a shared continuous space, whereas alignments between a diagnosis token and a time token reflected cross-space distance. The resultant dissimilarity matrix served as the mathematical foundation for unsupervised K-Medoids sequence clustering, identifying representative medoid trajectories that defined baseline pathway prototypes.
Cluster validation confirmed that a five-cluster configuration provided the optimal mathematical separation across candidate values of k, achieving a peak silhouette score of 0.22. Although modest in absolute terms, this silhouette magnitude reflected defined structure within high-dimensional, irregular electronic health record data, where naive unweighted edit-distance baselines collapsed patients into two uninformative groups separated merely by record length. The robustness of the five-cluster partition was established through subsampling-based stability testing across 300 independent perturbation iterations, wherein ten percent of the cohort was randomly removed before recomputing distances and cluster medoids. Aligning perturbed cluster solutions using the Hungarian maximum-overlap algorithm yielded a mean Adjusted Rand Index of 0.71 with a standard deviation of 0.23, demonstrating strong architectural stability. Complementary Apriori frequent-pattern mining verified these partitions by uncovering distinct combinations of co-occurring diagnostic motifs uniquely enriched within individual cluster groupings.
The five derived subtypes demonstrated markedly disparate pre-surgical timelines, demographic profiles, and diagnostic compositions across the 2,052 analyzed participants. Cluster 0, termed the fast progression subtype, represented younger individuals with a mean age of 53.6 years and minimal comorbidity who transitioned rapidly to surgery with a median pre-operative timeline of 252 days and higher rates of private health insurance. Cluster 1, characterized by joint pain and obesity, exhibited intermediate pre-surgical timing with a median of 766 days and a high prevalence of public insurance at 63 percent, yet achieved the highest post-operative quality-of-life scores. Cluster 2 combined joint pain with acute leg or hip fractures over an intermediate median duration of 1,154 days, frequently involving self-reported emotional health challenges. Cluster 3, labeled the soft-tissue-dominant subtype, displayed prolonged pre-surgical management spanning a median of 2,900 days alongside elevated chronic opioid therapy, while Cluster 4 captured older, multimorbid patients with a mean age of 62.7 years, 74 percent female representation, widespread reliance on public insurance, a median pre-surgical duration of 5,457 days, and the lowest post-surgical quality-of-life indices.
To evaluate pathway assignment certainty as longitudinal medical records accrue over time, the researchers transformed sequence distances into continuous class probabilities using a temperature-scaled SoftMax function. Trajectory uncertainty was evaluated across 1,436 patients who had documented knee osteoarthritis records by analyzing partial trajectories split into periods preceding and following initial osteoarthritis documentation. The fast progression subtype proved readily identifiable early in care with high probabilistic confidence and near-perfect concordance between partial and full-trajectory assignments. Similarly, obesity-associated and fracture-associated subtypes exhibited discernible risk patterns in early musculoskeletal records prior to formal osteoarthritis documentation. In contrast, soft-tissue-dominant and complex chronic subtypes showed low classification concordance when evaluated using truncated early records, confirming that diffuse and multimorbid care pathways consolidate their distinct trajectory identity only after prolonged clinical accumulation across multiple years of routine healthcare encounters.
Pre-surgical chronic opioid exposure was widespread across all five clusters, ranging between 57 percent and 68 percent of participants, but reached its highest frequencies in the prolonged soft-tissue and complex chronic subtypes at 66 percent and 68 percent, respectively. While chronic opioid use decreased across every cluster following surgical joint replacement, significant disparities persisted in post-operative management, with Cluster 0 dropping to 25 percent chronic use compared to 47 percent in Cluster 3 and 54 percent in Cluster 4. The authors emphasized that the observed associations remain observational and do not indicate direct causal mechanisms between pre-surgical trajectory membership and post-operative recovery outcomes. Furthermore, trajectories were anchored at the earliest observable musculoskeletal record rather than biological disease onset, diagnostic embeddings were cohort-specific, and clinical practice variability could not be tracked across specific clinics due to de-identification protocols. Future research will explore whether tailoring conservative prehabilitation interventions to specific trajectory subtypes can reduce persistent post-surgical opioid dependency and enhance long-term physical recovery.
Subject of Research: Unsupervised sequence clustering and uncertainty quantification for longitudinal electronic health record trajectories preceding total knee replacement
Article Title: TASC: a time-aware sequence clustering framework with uncertainty quantification for electronic health record trajectories
Article References: Yan, A. Y., Cudjoe, T. K. M., & Taylor, C. O. (2026). TASC: a time-aware sequence clustering framework with uncertainty quantification for electronic health record trajectories. BioData Mining, 19(1), Article 60. https://doi.org/10.1186/s13040-026-00569-7
Image Credits: AI Generated
DOI: 10.1186/s13040-026-00569-7
Keywords: total knee replacement, electronic health records, sequence clustering, uncertainty quantification, musculoskeletal disorders, osteoarthritis, weighted edit distance, K-Medoids, All of Us Research Program, trajectory analysis, chronic opioid use, word2vec embeddings, Adjusted Rand Index, frequent-pattern mining, prehabilitation, clinical informatics
News Source: Teresa Odom. (November 11, 2026). Machine Learning Framework Uncovers Five Distinct Pre-Surgical Trajectories Before Knee Replacement. Scienmag.
