Clin-JEPA: A Multi-Phase Co-Training Framework for EHR Patient Trajectories
Researchers have introduced Clin-JEPA, a novel multi-phase co-training framework designed for Joint-Embedding Predictive Architecture (JEPA) pretraining on Electronic Health Record (EHR) patient trajectories. Addressing the instability of naive co-training in existing JEPA models, this framework employs a five-phase curriculum to stabilize the joint training of a Qwen3-8B-based encoder and a latent trajectory predictor. The system aims to create a single backbone capable of forecasting patient trajectories and performing diverse risk-prediction tasks without requiring per-task fine-tuning. Evaluations on MIMIC-IV ICU data demonstrate significant improvements over baseline methods. Clin-JEPA uniquely achieves convergent latent rollout drift over 48-hour horizons and learns a clinically discriminative latent geometry where deteriorating patients are distinctly separated from stable ones. Furthermore, the model outperforms strong tabular and sequence baselines in multi-task downstream evaluations, achieving mean AUROC scores of 0.851 on ICareFM EEP and 0.883 across eight binary risk tasks. This advancement represents a significant step in applying advanced AI architectures to clinical data for improved patient outcome predictions.
Wire timeline
Clin-JEPA: A Multi-Phase Co-Training Framework for EHR Patient Trajectories
Researchers have introduced Clin-JEPA, a novel multi-phase co-training framework designed for Joint-Embedding Predictive Architecture (JEPA) pretraining on Electronic Health Record (EHR) patient trajectories. Addressing the instability of naive co-training in existing JEPA models, this framework employs a five-phase curriculum to stabilize the joint training of a Qwen3-8B-based encoder and a latent trajectory predictor. The system aims to create a single backbone capable of forecasting patient trajectories and performing diverse risk-prediction tasks without requiring per-task fine-tuning. Evaluations on MIMIC-IV ICU data demonstrate significant improvements over baseline methods. Clin-JEPA uniquely achieves convergent latent rollout drift over 48-hour horizons and learns a clinically discriminative latent geometry where deteriorating patients are distinctly separated from stable ones. Furthermore, the model outperforms strong tabular and sequence baselines in multi-task downstream evaluations, achieving mean AUROC scores of 0.851 on ICareFM EEP and 0.883 across eight binary risk tasks. This advancement represents a significant step in applying advanced AI architectures to clinical data for improved patient outcome predictions.
cs.AI updates on arXiv.org