Cardiovascular Risk Assessment: An Interpretable Machine Learning Approach

Abstract

Unsustainable health costs impose a new health care paradigm, where the support to clinical decision making assumes a critical importance. In this context, several machine learning risk assessment models have been developed in order to support a proper patients’ stratification. Although their superior performances, machine learning-based risk assessment models have faced strong difficulties to obtain the trust of professionals in their application in daily clinical practice. This work proposes a strategy able to address some of the major limitations of such models: i) interpretability; ii) personalization; iii) ability to incorporate new knowledge/new risk factors.

An hybrid scheme is developed, combining knowledge-driven methods (to create an interpretable set of rules for the general population) with data-driven methods (to select the most suitable subset of rules for each individual). Three main steps can be identified: i) derivation of an initial set of rules directly from current clinical evidence and/or data, ii) personalized scheme where a subset of the initial rules is identified as the most adequate one to classify a given patient; iii) an ensemble voting strategy based on the outputs of the previously selected rules. Moreover, the strategy demonstrates a high flexibility to incorporate new risk factors (in this case the inflammation biomarker), through the definition of additional rules.

This strategy was applied in the context of cardiovascular disease, namely on the risk stratification of Acute Coronary Syndromes patients. It was validated based on a real dataset composed of N = 1544 patients, admitted in the Cardiology Unit of Coimbra Hospital and Universitary Centre, achieving a SE = 0.763 and SP = 0.778.