EUVIMEDEUVIMED entdecken
CrossrefInfokarte

A Disease Trajectory Representation Framework With Risk-Aware Pseudotime Analysis for Electronic Health Records: Application to Heart Failure Risk Stratification

Jhen-Li Huang, Cing-Han Yang, Tun-Wen Pai

JMIR Medical Informatics · 2026-11-10 · 14 · e95929

DOI: 10.2196/95929 ↗
AbstractBibliografische Angaben

Abstract

Background Heart failure develops through complex longitudinal interactions among multiple chronic diseases, making longitudinal disease trajectories an important source of clinical information. Longitudinal disease trajectories contain rich structural information that reflects disease evolution and comorbidity interactions. However, this information is often lost when electronic health record (EHR) data are summarized using conventional pooling strategies or sparse International Classification of Diseases (ICD) representations. Objective This study aimed to develop an interpretable trajectory representation that preserves structural characteristics of longitudinal disease evolution for heart failure risk stratification using EHRs. Methods Longitudinal ICD trajectories were represented using Node2Vec embeddings and summarized with information gain (IG)–weighted geometric features, including historical means, directional shifts, dispersion, and deviation relative to high-risk disease representations. The resulting maximum IG-weighted deviation geometric (mIGwDG) representation was evaluated using 10-fold cross-validation against conventional embedding aggregation strategies, sparse ICD one-hot encoding, and a long short-term memory (LSTM) model. To facilitate interpretation of longitudinal disease evolution, we further developed a risk-aware pseudotime framework to organize patient trajectories along a model-inferred risk continuum and used a framework combining IG with the positive pointwise mutual information (PPMI) framework to characterize clinically relevant comorbidity patterns. Results The proposed framework achieved an out-of-fold (OOF) area under the precision-recall curve (AUPRC) of 0.667 and an area under the curve (AUC) of 0.888, outperforming conventional embedding aggregation methods. Compared with sparse ICD one-hot encoding, the proposed framework reduced feature dimensionality by approximately 75% while maintaining comparable predictive performance. Combining mIGwDG with sparse ICD one-hot features further improved both logistic regression (LR) and extreme gradient boosting (XGBoost), suggesting that the proposed representation captures complementary longitudinal disease trajectory information. In addition, the risk-aware pseudotime framework organized heterogeneous patient trajectories along a model-inferred risk continuum, while IG×PPMI analysis highlighted clinically relevant comorbidity patterns that supported the interpretation of longitudinal disease organization. Conclusions The proposed framework provides a compact and interpretable representation of longitudinal disease trajectories while maintaining comparable predictive performance. By combining geometric trajectory characterization with risk-aware pseudotime visualization, the framework facilitates the representation of longitudinal EHRs, the visualization of disease evolution, and risk stratification.

Wissen, das zu Ihrer Arbeit passt.

Vom kostenlosen Einstieg über die persönliche Recherche bis zur vollständigen Research Suite.

Tarife ansehen →

Sind alle Verlagsvolltexte enthalten?

Nein. Bibliografische Nachweise, Abstracts und Volltexte sind unterschiedliche Inhalte. Ob ein Volltext verfügbar ist, hängt von der Quelle, Open Access und Ihren vorhandenen Lizenzen ab. Ein EUVIMED-Tarif ersetzt keine zusätzlichen Verlagsrechte.

Tarife ansehen →