Skip to content

kaggle_training.md

Latest commit

 

History

History
67 lines (46 loc) · 2.08 KB

kaggle_training.md

File metadata and controls

67 lines (46 loc) · 2.08 KB
Raw

Kaggle Model Training

The production model is trained on Kaggle with the CDC Diabetes Health Indicators dataset, a cleaned BRFSS 2015 dataset with 253,680 survey responses and 21 health indicator features.

Why This Dataset

The original repo used a tiny synthetic CSV, which is useful for smoke tests but not strong enough for a portfolio-grade ML project. The CDC/BRFSS dataset adds:

  • enough rows for meaningful train/test evaluation
  • lifestyle, access, demographic, and clinical indicator features
  • a binary diabetes/prediabetes risk target
  • a public Kaggle workflow that produces reproducible artifacts

Training Outputs

Running kaggle/kaggle_train.py creates:

  • artifacts/risk_model.joblib
  • artifacts/metrics.json
  • artifacts/model_card.md

The API loads artifacts/risk_model.joblib automatically through MODEL_ARTIFACT_PATH. If the artifact is missing, the app falls back to data/patients.csv for local demos and tests.

Run On Kaggle

Production training is meant to run on Kaggle, not on your laptop:

./scripts/kaggle_run.sh

That script:

  1. syncs kernel-metadata.json with your Kaggle username
  2. pushes kaggle/kaggle_train.py
  3. waits for the kernel to finish
  4. downloads risk_model.joblib, metrics.json, and model_card.md into artifacts/

Manual commands:

kaggle kernels push -p kaggle
kaggle kernels status aditya2402/healthcare-ai-diabetes-risk-training
kaggle kernels output aditya2402/healthcare-ai-diabetes-risk-training -p artifacts

Local Mirror (optional)

scripts/train_diabetes_model.py mirrors the Kaggle script for CI or offline debugging only. Use Kaggle for the portfolio artifact.

Model Selection (v4 pipeline)

The training script:

  1. Compares logistic regression, histogram gradient boosting, LightGBM, and random forest
  2. Ranks models with 5-fold stratified CV (ROC AUC)
  3. Runs RandomizedSearchCV on the top 2 models
  4. Applies isotonic calibration to all candidates
  5. Picks the winner on the holdout test set (ROC AUC, then F1)

Metrics include CV scores, Brier score, and tuned classification threshold.