Why Do I Trust Your Model? Building and Explaining Predictive Models for Peritoneal Dialysis Eligibility

Abstract

Achieving fairness, accountability and transparency is of vital importance when using machine learning (ML) techniques in the health-care realm. Yet, the myths of the "black box" of ML algorithms still exist among healthcare professionals. In this research, we developed a ML model for the eligibility of patients for peritoneal dialysis and employed various interpretability techniques to explain the models to nephrologists to gain their trust in the model. We compared different model-specific and model-agnostic ML interpretability strategies with traditional statistical analysis methods and we analyzed their applicability in healthcare domain.