zinahghulam/m31-patient-timelines

M31 Patient Timelines

Transformer-based longitudinal EHR model for forecasting newly diagnosed clinical conditions from patient timelines.

Model

• Architecture: EHRTransformer
• Embedding dimension: 128
• Transformer layers: 3
• Attention heads: 4
• Maximum sequence length: 512
• Output: 40-condition multi-label prediction
• Vocabulary: 209 clinical event tokens
• Training: AdamW + Focal Loss with positive-class weighting
• Early stopping based on validation macro-AUROC

The model was trained from scratch using the provided synthetic EHR dataset. No external datasets or pretrained weights were used.

Validation Performance

The following results are from the held-out validation cohort at the optimal early-stopping checkpoint (Epoch 20):

Metric	Validation
Macro AUROC	0.6642
mAP	0.1344
Macro F1	0.1380
Brier Score	0.1953

The test-set ground truth is withheld for independent scoring.

Code

Source code and reproducibility instructions:

https://github.com/zinahghul/m31-patient-timelines