Towards interpretable machine learning for clinical decision support

A major challenge in delivering reliable and trustworthy computational intelligence for practical applications in clinical medicine is interpretability. This aspect of machine learning is a major distinguishing factor compared with traditional statistical models for the stratification of patients, which typically use rules or a risk score identified by logistic regression. We show how functions of one and two variables can be extracted from pre-trained machine learning models using anchored Analysis of Variance (ANOVA) decompositions. This enables complex interaction terms to be filtered out by aggressive regularisation using the Least Absolute Shrinkage and Selection Operator (LASSO) resulting in a sparse model with comparable or even better performance than the original pre-trained black-box. Besides being theoretically well-founded, the decomposition of a black-box multivariate probabilistic binary classifier into a General Additive Model (GAM) comprising a linear combination of non-linear functions of one or two variables provides full interpretability. In effect this extends logistic regression into non-linear modelling without the need for manual intervention by way of variable transformations, using the pre-trained model as a seed. The application of the proposed methodology to existing machine learning models is demonstrated using the Multi-Layer Perceptron (MLP), Support Vector Machine (SVM), Random Forests (RF) and Gradient Boosting Machines (GBM), to model a data frame from a well-known benchmark dataset available from Physionet, the Medical Information Mart for Intensive Care (MIMIC-III). Both the classification performance and plausibility of clinical interpretation compare favourably with other state-of-the-art sparse models namely Sparse Additive Models (SAM) and the Explainable Boosting Machine (EBM).