Forecasting new diagnoses from patient timelines (Synthea)

For each patient, predict which of 40 conditions are first diagnosed in the five years after an anchor date (last encounter minus five years), using only events before the anchor. 3,514 synthetic patients: 2,791 train / 365 validation / 358 test.

• Report: report/report.md (render with python -m src.make_report)
• Model: huggingface.co/sallamsaka/M31-Coding-Test
• Training curves: wandb project m31-patient-timeline
• Submission: outputs/predictions.csv

Setup

Python 3.13, CPU only. Put the provided data folders (train_val/, test/) and CSVs (patient_splits.csv, target_conditions.csv, test_anchors.csv) in the repo root.

pip install -r requirements.txt
python -m pytest -q                                   # tests (~5 min)
python -m src.cross_validate                         # 5-fold out-of-fold predictions
python -m src.calibrate --fit oof --model ens_lr_gbdt_tx --oof-model lr+gbdt+transformer
python -m src.run_baseline --with-transformer --recipe ens_lr_gbdt_tx   # -> predictions.csv

To regenerate predictions.csv from the published model files without training:

python -m src.reproduce --from-hub sallamsaka/M31-Coding-Test

.\run_all.ps1 -Smoke runs every code path in a few minutes.

Model

An average of three models' logits, with per-condition calibration:

• logistic regression and gradient-boosted trees on 3,320 count/lab/demographic features;
• a one-layer transformer over the event sequence (one token per event, with the time before the anchor and the patient's age at each event encoded), fused with the same features.

Model choices were made with 5-fold cross-validation on the training patients; the validation set was used for early stopping and monitoring.

Code

path	what it does
src/data/cohort.py	load the CSVs, compute anchors, build the event table
src/data/labels.py	labels: first-ever diagnosis inside the window
src/data/features.py	tabular features
src/data/sequences.py	vocabulary and token sequences
src/models/gpt.py	the transformer
src/train_baseline.py, src/train_finetune.py	training
src/cross_validate.py, src/evaluate.py	cross-validation and paired bootstrap comparisons
src/run_baseline.py, src/predict.py	final fit and predictions.csv
src/tx_sweep.py	the transformer experiments reported in section 2
src/hf_upload.py, src/log_shipped_run.py	Hugging Face upload; wandb-logged refit

Experiment tracking is opt-in ($env:WANDB=1); every run also logs to outputs/metrics.jsonl.