Explainable and Actionable Machine Learning Models for Electronic Health Record Data

State-of-the art machine learning (ML) methods show immense promise for medical applications, offering significant improvement in prediction capabilities for electronic health record (EHR) data. However, the models’ black-box nature make it difficult for an end-user to trace the decision-making process, a key challenge for the application of machine learning in healthcare. Explainable machine learning (ex-ML) methods is a complementary tool, identifying key features associated with a model’s predictions. Extending from ex-ML methods, we show how EHR data can be used to make explainable machine learning predictions, and generate insights to minimise an unfavourable healthcare outcome. Firstly, our clinical explainability step outlines how feature importance values change across the range of possible physiological variables, for a single patient and across a population. The output is a target feature value for which the feature importance values are minimised. Secondly, we show that explanations across different machine learning models are not guaranteed, and we provide a method for aggregation. Finally, we develop a set of actionable recommendations on features that can be changed, providing a target treatment order that doctors can follow, to reduce a patient’s likelihood of disease.