matlab-stats-ml-v2

MATLAB Statistics and Machine Learning Toolbox Skill

Version: R2025b | Focus: Biomedical-specific workflows and template scripts

This skill provides expert guidance for applying the Statistics and Machine Learning Toolbox to biomedical data analysis. The model already has strong knowledge of the toolbox API — this skill adds domain-specific workflows, clinical best practices, and ready-to-use template scripts.

What Makes Biomedical Stats Different

Biomedical data analysis has unique challenges that generic ML tutorials do not address:

1. Class imbalance is the norm — Disease prevalence is often 1-10%. Always use cost-sensitive learning or resampling. Never report accuracy alone; use sensitivity, specificity, AUC, and likelihood ratios.

2. p >> n is common — Genomics/proteomics panels have thousands of features and hundreds of samples. Feature selection and dimensionality reduction are mandatory, not optional.

3. Data leakage is catastrophic — All preprocessing (normalization, feature selection, imputation) must happen within cross-validation folds. Leakage inflates performance and leads to clinical failures.

4. Clinical interpretability matters — Clinicians need odds ratios, hazard ratios, and feature importance — not just predictions. Use fitglm for odds ratios, coxphfit for hazard ratios, shapley for model explanation.

5. Missing data is expected — Use fillmissing(X, 'knn') for clinical imputation (requires Stats & ML Toolbox, not Bioinformatics Toolbox).

6. Censored data requires special handling — Survival analysis with ecdf(..., 'Censoring', ...) and coxphfit — not standard regression.

7. Multiple testing correction is essential — When testing many biomarkers, use Benjamini-Hochberg FDR, not just Bonferroni.

Read Before Coding

Task	Start Here
Build a diagnostic classifier	scripts/template_svm_classification.m or scripts/template_random_forest_ensemble.m
Analyze survival / time-to-event data	scripts/template_cox_survival_analysis.m + cards/survival-analysis.md
Discover biomarkers	cards/biomedical.md (consensus feature selection)
Optimize hyperparameters	scripts/template_bayesopt_hyperparameter.m + cards/bayesian.md
Compare treatment groups	scripts/template_hypothesis_testing.m + cards/hypothesis-testing.md
Cluster patients into subtypes	scripts/template_kmeans_patient_clustering.m
Fit distributions to clinical data	scripts/template_distribution_fitting.m
Reduce dimensions (genomics, etc.)	scripts/template_pca_feature_reduction.m
Interpret a black-box model	scripts/template_shapley_interpretability.m
Handle missing clinical data	scripts/template_missing_data_handling.m
Build a regression model	scripts/template_glm_regression.m
Cross-validate properly	scripts/template_cross_validation_pipeline.m

Template Scripts

The scripts/ directory contains 12 ready-to-use MATLAB scripts covering the most common biomedical analysis workflows:

Script	Purpose
template_svm_classification.m	SVM classifier with RBF kernel, standardization, optimization
template_random_forest_ensemble.m	Random Forest with feature importance and OOB error
template_cox_survival_analysis.m	Kaplan-Meier curves, Cox regression, hazard ratios
template_bayesopt_hyperparameter.m	Bayesian hyperparameter optimization with bayesopt
template_hypothesis_testing.m	T-test, ANOVA, nonparametric tests with effect sizes
template_kmeans_patient_clustering.m	Patient subtyping with optimal k selection
template_distribution_fitting.m	Multi-distribution fitting with AIC/BIC comparison
template_pca_feature_reduction.m	PCA pipeline for high-dimensional biodata
template_shapley_interpretability.m	SHAP values for model explanation
template_missing_data_handling.m	Missing data strategies (kNN, interpolation)
template_glm_regression.m	GLM with odds ratios and clinical reporting
template_cross_validation_pipeline.m	Nested CV for unbiased performance estimation

Biomedical Research Workflow

%% Complete Biomedical Classification Pipeline
% 1. Load and explore clinical data
data = readtable('patient_data.csv');
summary(data);
grpstats(data, 'Diagnosis', {'mean', 'std'});

%% 2. Handle missing data
data = fillmissing(data, 'knn');  % kNN imputation (Stats & ML Toolbox)

%% 3. Feature selection (within training fold only!)
cv = cvpartition(data.Diagnosis, 'Holdout', 0.2, 'Stratify', true);
XTrain = data(training(cv), :);

[idx, scores] = fscmrmr(XTrain(:, predictors), XTrain.Diagnosis);
selectedFeatures = predictors(idx(1:10));

%% 4. Train classifier with cross-validation
Mdl = fitcensemble(XTrain(:, selectedFeatures), XTrain.Diagnosis, ...
    'Method', 'Bag', 'NumLearningCycles', 100, ...
    'OptimizeHyperparameters', 'auto', ...
    'HyperparameterOptimizationOptions', struct('ShowPlots', false));

%% 5. Evaluate on held-out test set
XTest = data(test(cv), :);
[label, score] = predict(Mdl, XTest(:, selectedFeatures));
[X, Y, T, AUC] = perfcurve(XTest.Diagnosis, score(:,2), 'Positive');

%% 6. Interpret model
explainer = shapley(Mdl, XTrain(:, selectedFeatures));
plot(explainer);

Knowledge Cards

See knowledge/INDEX.md for the complete card index. Key cards:

• cards/survival-analysis.md — Kaplan-Meier, Cox regression, coxphfit patterns
• cards/biomedical.md — Diagnostic classifiers, biomarker discovery, class imbalance
• cards/bayesian.md — bayesopt, MCMC sampling, Bayesian model comparison

See Also (Related MATLAB Toolboxes)

• Deep Learning Toolbox — For CNNs, U-Net, YOLO (trainnet, unet, dlnetwork)
• Medical Imaging Toolbox — For medical data I/O and radiomics (medicalVolume, radiomics)
• Image Processing Toolbox — For image-based feature extraction (regionprops, imgaussfilt)
• Wavelet Toolbox — For multiresolution feature extraction (wavedec2, wdenoise2)