A Scalable Deep Learning Analytics Pipeline for Converting Longitudinal Real-World Data Into Predictive Disease Trajectories

Abstract

Real-world healthcare data encode longitudinal patterns reflecting evolving patient trajectories, but these patterns are fragmented, irregular, and sparse. Static classifiers and heuristic trigger systems rely on explicit diagnosis codes or narrow event heuristics, limiting their ability to detect early trajectory transitions. This work presents a scalable temporal modeling framework that transforms tokenized claims and laboratory records into predictive trajectory states. The approach reconstructs longitudinal encounters into episodic structures and Lines of Therapy, embeds categorical and laboratory variables, and encodes patient histories using a gated recurrent network with time-gap decay. A discrete-time hazard layer estimates future transition probabilities, generating Predictive Clinical Signals that summarize emergent or escalating disease states. Applied to oncology and rare-disease cohorts, the framework achieved a time-dependent AUC of 0.83, a median lead-time of 18 days relative to first ICD-10 code attribution, and precision/recall values of 0.78/0.70 compared with rule-based baselines. The pipeline supports multi-vendor token ecosystems, operates exclusively on de-identified data, and scales to millions of encounters in weekly inference cycles. These results illustrate a practical direction for temporal representation learning over heterogeneous healthcare streams. The model outputs analytic trajectory indicators and is not intended for clinical decision support or patient-level intervention.

Ähnliche Arbeiten