Framework models full patient hospital stays from electronic health records

A team of researchers has developed EHR2Path, a machine learning framework designed to trace and forecast how patients' conditions evolve during hospital stays. The system integrates multiple types of clinical information into a unified prediction model, moving beyond approaches that focus on single, narrow tasks.

The Challenge#

Hospitals generate vast amounts of patient data spanning notes from physicians, vital signs, medication records, lab results, imaging reports, and bedside observations. Existing prediction systems typically handle only a few data types at once or work with limited time windows—missing the fuller picture of how a patient's status changes over days or weeks. This fragmentation limits the ability to anticipate deterioration, recovery timelines, treatment transitions, and shifts in care needs.

How EHR2Path Works#

The framework converts diverse clinical inputs—radiology reports, clinical notes, vital signs, medication patterns, laboratory values, and continuous monitoring data—into a shared temporal format. To handle lengthy patient histories without overwhelming the model with tokens, the team introduced a Masked Summarization Bottleneck: a mechanism that compresses long-term history into compact, task-focused summary tokens while keeping recent data intact for context.

This design enables both better performance and computational efficiency, allowing the system to reason over broader patient information while managing computational demands.

Validation and Results#

In experiments using MIMIC-IV, a large real-world hospital dataset, EHR2Path demonstrated the ability to forecast the next step in a patient's hospital journey and to simulate complete trajectories iteratively. The framework outperformed comparable baseline methods on directly comparable prediction tasks, suggesting that pathway-level modeling—tracking full patient arcs rather than isolated snapshots—is feasible and beneficial.

Clinical Implications#

By integrating heterogeneous, longitudinal EHR data into a single predictive model, the approach could support more anticipatory clinical decision-making. Rather than reacting to individual events, clinicians might better prepare for where a patient is headed.