Radiomics ML

Radiomics / Classical-ML Skill

Purpose

Radiomics + tree-ensemble studies (features → random forest / XGBoost → a clinical outcome) are the most common solo-doable clinical-ML workflow — no GPU, no engineer — and the most commonly over-optimistic: hundreds-to-thousands of features on tens of patients, hyperparameters tuned on the same folds the performance is reported from, features selected on the whole dataset, unstable features never filtered, and discrimination (AUC) reported without calibration. This skill produces the pipeline correctly and audits an existing one, so the clinical result survives review (Lambin 2017; CLEAR; TRIPOD+AI; PROBAST-AI).

It sits beside the imaging-DL lane: where /model-scaffold builds a deep network, radiomics-ml covers the feature-based classical-ML path. It integrates scikit-learn / xgboost / pyradiomics (referenced in the emitted code); it does not reimplement them and never runs a model on real patient data.

When to use

• You have a radiomics or clinical/tabular feature table and want to build a random-forest / XGBoost clinical prediction model that will pass statistical review.
• You want to audit an existing radiomics/ML pipeline for the failure modes below.

When NOT to use

• Deep-learning imaging models → /model-selection → /model-scaffold → /model-assessment.
• Classical inferential statistics / a regression model as the estimand → /analyze-stats.
• Interpretability of a trained network → /model-assessment.
• Reimplementing scikit-learn / xgboost / pyradiomics → out of scope (this skill wires and audits them).

The failure modes (what the gate enforces)

1. No nested CV. Tuning and reporting on the same folds inflates performance. Use nested CV or a held-out test set.
2. High dimensionality, low events. Candidate features ≥ events overfits — the classic radiomics trap. Declaring dimensionality_reduction does not clear it: n_features already counts what is left after outcome-blind reduction. The gate's p ≥ events rule (events = the minority class) is a floor that catches the worst case, not a sample-size criterion: size the study with /calc-sample-size Test 12 (Riley criteria, pmsampsize), counting every candidate feature that reaches outcome-driven selection or fitting. At C = 0.75 and 35% prevalence that is about 48 patients per candidate parameter: 100 candidates need N = 4,755 (1,665 events), and the gate passes 100 features on 105 events. Reduce candidates without the outcome first (stability, redundancy, clinical prior); LASSO or other penalisation does not substitute for sample size, because the shrinkage it estimates is itself unstable at small n (Riley et al., J Clin Epidemiol 2021; Van Calster et al., Stat Methods Med Res 2020).
3. Selection outside the fold. Feature selection fit on the whole dataset leaks the held-out folds. Nest selection inside each training fold.
4. No feature stability. Radiomics features are unstable across acquisition/segmentation — filter to reproducible features (ICC / test-retest).
5. No calibration. A clinical prediction model needs calibration (slope/intercept + a flexible curve), not discrimination alone.
6. No external validation. A single-cohort model needs external / temporal validation for a clinical claim.

Not in the gate, because the manifest cannot show it: rows of one patient on both sides of a split. Lesion-level tables with row-wise folds let the model recognise the patient; on null synthetic data this took nested-CV AUROC from 0.48 to 0.99. Split by patient and check the fold table (Phase 2).

Workflow

Phase 1 — Extract features (integrate, don't reimplement)

For radiomics, extract with pyradiomics under reproducible, IBSI-aligned settings (fixed bin width, resampling, normalisation) — record them. For clinical/tabular data, assemble the feature table with a patient/subject ID and the outcome. One row per lesion or ROI is fine; the ID is what the folds are split by. See references/radiomics_ml_guide.md.

Phase 2 — Build the pipeline correctly

• Feature stability — with test-retest / multi-rater data, keep features with ICC ≥ 0.75.
• Nested cross-validation — outer folds estimate performance, inner folds tune; do feature selection and scaling inside each training fold (never on the whole dataset). The CV unit is the patient: split both loops with StratifiedGroupKFold(groups=patient_id) so one patient's lesions never straddle folds, write the fold table (patient_id,split), and prove it with /model-assessment's check_split_leakage.py --splits cv_folds.csv --seed <seed> --strict (skeleton in the guide §4).
• Dimensionality — reduce the candidate set without the outcome (ICC stability, |r| redundancy filter, clinical prior; PCA fit inside the fold), then size the study for the candidates that remain with /calc-sample-size Test 12 (pmsampsize). LASSO selects inside the fold but does not make a small sample large enough; report the shortfall as a limitation if N falls below the Riley minimum.
• Model — pick from the full classical family for the task; a simple baseline (penalised logistic) is mandatory alongside any complex learner:
  • penalised regression — LASSO / ridge / elastic-net logistic (also the baseline)
  • margin / kernel — linear or RBF SVM
  • instance-based — k-NN
  • probabilistic / discriminant — naive Bayes, LDA / QDA
  • trees & bagging — decision tree, random forest, extra-trees
  • boosting — XGBoost, LightGBM, CatBoost, HistGBM, AdaBoost
  • shallow neural — MLP
  • meta — stacking / voting ensembles
  • unsupervised (upstream) — PCA / UMAP for reduction, k-means / hierarchical / GMM for phenotyping The gate below is learner-agnostic — it audits the pipeline (nested CV, leakage, dimensionality, calibration), so it applies identically to any of these. See the full method map in docs/method_coverage_map.md.
• Report — discrimination and calibration (slope/intercept + flexible curve, via the /analyze-stats calibration guide) and clinical utility (decision curve). SHAP for interpretation.