Abstract

Machine learning (ML) and in particular deep neural networks (DNNs) have the potential to transform diagnosis, prognosis and treatment planning in oncology. Uninterpretability of DNNs, however makes their deployment in safety critical domains, such as oncology, challenging. This is evidenced by the scepticism of clinical community about ML systems. In this talk, I present REM , a model extraction approach that allows extracting rules from deep neural networks. Unlike interpretability methods, such as feature importance and sample importance, model extraction allows human-simulatability that is an especially valuable feature for an interpretability method suited for use in clinical domain. It allows clinicians to inspect the predictions of a ML model and contrast them with their expert knowledge to verify the biological relevance as well as to identify the unintended bias. It further allows checking the impact of perturbation in input on the output and adjusting the model accordingly.

Register here