Table 1.

Key considerations and example use cases for learning healthcare systems

	Key considerations	Example use cases
Prediction	Prediction algorithm(s) designed by a multidisciplinary team with knowledge of the clinical target and optimal approaches, given data limitations. Target of prediction is a clinically relevant endpoint with potential for intervention. Overfitting is avoided through cross-validated assessment of prediction performance, as well as external validation.	Early disease detection. Generation of differential diagnoses. Clinical risk–benefit prediction for competing interventions. Quantification of expected utility, given risk–benefit predictions and patient-reported preferences. Early warning systems in critical care patients.
Interpretation/Inference	Measure feature (variable) influence in prediction. Easily interpreted visualisations of dependence of predictions on each feature. Predictions of intervention effects with validated measures of uncertainty. Identification of plausible causal pathways consistent with observations.	Identification of risk factors, which, on their own or through interactions, have the greatest impact on the prediction of clinical outcomes. Choosing an intervention that targets specific risk factors to optimise the risk–benefit according to the individual’s preferences.
Communication	Intuitive decision-support tools that present aspects of the data relevant to the specific decision under consideration. Integration of intuitive data visualisations in the electronic medical record and patient portals with links to original clinical notes, labs and images.	Explaining predicted risks to patients. Discussion of patient-specific, modifiable risk factors to intervene on. Involvement of clinicians and patients in the design and implementation of tools for learning healthcare systems, as well as discussions of the ethical consideration.