mu-glioma-post-peak-forward-model

Peak Forward Model

This folder contains the exported probability-capable bundle for the best observed corrected forward split.

• Source run: results/repeated_forward_hybrid_basic_corrected/seed_62
• Modalities: t1c, flair
• Clinical feature set: hybrid_basic
• Output probability column: progression_risk_probability

Files

• model_bundle.pkl: pickled classifier, preprocessor, calibrator, and threshold.
• metadata.json: exported metrics and selected features.
• train_predictions.csv: out-of-fold probability outputs for the training split.
• test_predictions.csv: probability outputs for the held-out split.
• shap_feature_importance.csv: test-set mean absolute SHAP values for the exported logistic model.
• shap_modality_share.csv: modality-level SHAP aggregation.

Usage

import pickle
bundle = pickle.load(open('model/peak_forward_seed62/model_bundle.pkl', 'rb'))

The training predictions are exported as out-of-fold scores so they stay comparable to held-out test outputs.

Use bundle['classifier'] on preprocessed features for raw probabilities, then pass them through bundle['calibrator'] to obtain calibrated progression risk probabilities.