Clinical Machine Learning with Python
1 of 10 modules available · 0 completed · ≈ 35 min available now · ≈ 6 h planned
Modules
- 1. From Clinical Problem to Prediction Problem35 min
- 2. Know Your Clinical DataNot yet released
- 3. Build Your First Clinical ModelNot yet released
- 4. Is the Model Actually Any Good?Not yet released
- 5. Can You Trust the Predicted Risk?Not yet released
- 6. How Models Fool YouNot yet released
- 7. Who Does the Model Work For?Not yet released
- 8. Why Is the Model Making This Prediction?Not yet released
- 9. Will It Work in the Real World?Not yet released
- 10. Capstone: Should We Deploy ReadmitAI?Not yet released
Modelling that survives contact with a clinic
Clinical Machine Learning with Python
A ten-module track in development, built around one recurring question: can we trust this prediction enough to use it on patients? Module 1 — turning a clinical ambition into a precise prediction task — is available now and needs no Python. The remaining nine modules (clinical data, a first model in pandas and scikit-learn, metrics, calibration, the ways models mislead, subgroup performance, explanation, real-world transportability and a capstone appraisal) are planned and not yet released. All work uses one fictional readmission case and synthetic data.
1 of 10 planned modules are released so far; the rest are in development.
From Clinical Problem to Prediction Problem
Before any code: turn “reduce readmissions” into a prediction task with a defined population, index time, horizon, outcome and intended action — and separate predicting risk from claiming an intervention works.
Know Your Clinical Data
Provenance, unit of analysis, repeated patients, label quality, missingness and prevalence — the properties that decide whether a dataset can answer the question at all, and the split design that follows from them.
Build Your First Clinical Model
A pandas and scikit-learn workflow from table to fitted model: a deliberately simple baseline, logistic regression, and internal validation done in a way that respects patient, site and time structure.
Is the Model Actually Any Good?
Confusion matrix, sensitivity and specificity, PPV and NPV, ROC/AUROC and precision–recall — read at a chosen threshold and under the prevalence you actually have, with uncertainty shown rather than assumed away.
Can You Trust the Predicted Risk?
Calibration as a separate question from discrimination: calibration plots, calibration-in-the-large and slope, the Brier score, and what a mis-calibrated risk does to a threshold-based clinical decision.
How Models Fool You
Leakage by target, time, patient and site; overfitting and optimistic validation; splits that quietly answer an easier question; missing-data traps and shortcut learning — and what suspiciously high performance usually means.
Who Does the Model Work For?
Subgroup performance and fairness without reducing either to a single number: small-subgroup uncertainty, the limits of intersectional analysis, and the difference between an observed performance gap and its explanation.
Why Is the Model Making This Prediction?
Coefficients, global importance and local attributions — what each one can and cannot support. Explanation as traceability of the model's inputs and behaviour, not as an account of why the patient had the outcome.
Will It Work in the Real World?
External validation and transportability across sites and time, dataset shift, clinical utility against operational capacity, the workflow the prediction lands in, and the monitoring that decides recalibration, updating or retirement.
Capstone: Should We Deploy ReadmitAI?
A fictional vendor dossier with strong-looking headline performance and deliberately incomplete evidence. Appraise it as a whole and defend a decision: reject, request specific further evidence, or pilot under stated conditions.