TGDS-HF: AI Model Predicts Heart Failure Risks from EHRs
A new machine learning approach, Trajectory-Guided Discharge Stratification for Heart Failure (TGDS-HF), has been created by researchers to forecast one-year risks of clinical instability or death in heart failure patients using electronic health records (EHRs). This model analyzes a patient’s in-hospital trajectory, including diagnoses, vital signs, lab results, medications, and procedures, through a compact short-context autoregressive Transformer. It was tested on a Swedish cohort of 42,820 patients, predicting outcomes at the time of initial heart failure diagnosis during hospitalization. Comprising category-level tokenization, recency-weighted temporal representation, and sequence model configuration, TGDS-HF underwent ablation experiments for component validation. The research, published on arXiv (ID: 2511.16839), aims to enhance discharge planning for heart failure, leveraging routinely collected EHR data for scalability in clinical environments.
Key facts
- TGDS-HF is a methodology for heart failure discharge stratification using EHRs.
- It uses a compact short-context autoregressive Transformer.
- The model predicts one-year risks of clinical instability or mortality.
- The cohort includes 42,820 Swedish heart failure patients.
- Predictions are made at the initial heart failure diagnosis in-hospital.
- TGDS-HF has three components: tokenization, temporal representation, and sequence model configuration.
- Ablation studies were performed on the components.
- The paper is available on arXiv with ID 2511.16839.
Entities
Institutions
- arXiv
Locations
- Sweden