A novel interpretable machine learning system to generate clinical risk scores: An application for predicting early mortality or unplanned readmission in a retrospective cohort study

Risk scores are widely used for clinical decision making and commonly generated from logistic regression models. Machine-learning-based methods may work well for identifying important predictors to create parsimonious scores, but such ‘black box’ variable selection limits interpretability, and variable importance evaluated from a single model can be biased. We propose a robust and interpretable variable selection approach using the recently developed Shapley variable importance cloud (ShapleyVIC) that accounts for variability in variable importance across models. Our approach evaluates and visualizes overall variable contributions for in-depth inference and transparent variable selection, and filters out non-significant contributors to simplify model building steps. We derive an ensemble variable ranking from variable contributions across models, which is easily integrated with an automated and modularized risk score generator, AutoScore, for convenient implementation. In a study of early death or unplanned readmission after hospital discharge, ShapleyVIC selected 6 variables from 41 candidates to create a well-performing risk score, which had similar performance to a 16-variable model from machine-learning-based ranking. Our work contributes to the recent emphasis on interpretability of prediction models for high-stakes decision making, providing a disciplined solution to detailed assessment of variable importance and transparent development of parsimonious clinical risk scores.Author summary: Risk scores help clinicians quickly assess the risk for a patient by adding up a few scores associated with key predictors. Given the simplicity of such scores, shortlisting the most important predictors is key to predictive performance, but traditional methods are sometimes insufficient when there are a lot of candidates to choose from. As a rising area of research, machine learning provides a growing toolkit for variable selection, but as many machine learning models are complex ‘black boxes’ that differ considerably from risk scores, directly plugging machine learning tools into risk score development can harm both interpretability and predictive performance. We propose a robust and interpretable variable selection mechanism that is tailored to risk scores, and integrate it with an automated framework for convenient risk score development. In a clinical example, we demonstrated how our proposed method can help researchers understand the contribution of 41 candidate variables to outcome prediction through visualizations, filter out 20 variables with non-significant contribution and build a well-performing risk score using only 6 variables, whereas a machine-learning-based method selected 16 variables to achieve a similar performance. We have thus presented a useful tool to support transparent high-stakes decision making.