Building robust methods for model explainability for healthcare

点击翻译按钮获取中文摘要

With the recent rise in medical applications of artificial intelligence (AI), decision makers are now demanding more transparency from predictive models. Explainability (XAI) is a complex challenge that stems from various technical fields. Its main goal is to build explanation model that make model decisions readily interpretable. We distinguish two types of XAI approaches: (i) local model approximations where a simple model, g(x), is fitted to predict the black boxmodel, f(x), in a neighborhood around the prediction point x, and (ii) additive feature attribution methods where a model-free estimator describes how the model outcome at a point x changes when some of its features are removed and sampled from a reference distribution. To this day, there is no consensus in the field of XAI when it comes to finding the right method for a use case. Further, methods are not reliable due to their instability and lack of robustness. Ultimately, a limited number of XAI methods are based on a causal reasoning, even though clinicians often seek causal explanations. Our aim is to find methodological improvements to bridge the gap between statisticians and clinicians who want to routinely use transparent, fair AI tools for prediction. We make use of statistical learning theory, robust statistics and knowledge of healthcare applications to build new, innovative approaches. The focus of our first research endeavor was to tackle the issue of locality in local explanation models and build a more robust approach to Shapley values that can resist adversarial attacks. In parallel, we have been developing methods for understanding poly-victimization in multi-outcome causal models. Working on both model explainability and causal inference will hopefully enable us to design causal XAI methods. Long term, we hope to bridge the gap between the two subfields. Such technical advances in model explainability and causal inference can have high social impact, as they can improve the methodology for developing risk scores, and in particularly medical risk scores. Such measures are built according to feature attributions in predictive, non-causal models. Ultimately, we aim to study various alternatives to evaluating XAI methods. This task is challenging by definition, as there is no ground truth for evaluating these methods. Our aim is to borrow ideas and concepts from unsupervised learning and adapt them to the purpose of model explainability. This project falls within two EPSRC research areas: "Artificial Intelligence and robotics" and "Healthcare technologies".