Advancing Neurosymbolic AI for Smart and Trustworthy Healthcare

University Of Pennsylvania, Philadelphia PA

Investigators

Abstract

Project Summary Artificial intelligence (AI) is poised to transform healthcare, particularly in clinical decision support for treatment selection in a range of diseases. However, AI is underutilized in clinical care, in part due to two unique challenges: 1) noisy and complex health data, such as multi-modal electronic health record (EHR) data, and 2) demand for a high degree of trustworthiness in treatment selection use cases. Explainability is a core principle of trustworthy AI, but may be lacking in current state of the art (SOTA) AI models for healthcare, particularly for analysis of multi-modal EHR data. Most notably, current approaches to explainable AI in healthcare based on post-hoc explanations of black-box models are unlikely to offer clinically meaningful and actionable explanations to clinicians and end-users, thereby decreasing trust and buy-in. Furthermore, treatment effect estimation and selection based on inaccurate explanations may cause patient harm. As such, there is an urgent need for inherently explainable AI/ML models in healthcare that also enjoy the robustness and expressivity of modern deep learning models. Our preliminary work suggests that incorporation of explanations based on existing clinician domain knowledge would be a key facilitator for trusting AI models that predict treatment benefit. However, to the best of our knowledge, no existing SOTA AI models effectively incorporate clinical domain knowledge into explanations. To address these gaps, we propose to develop Med-Scallop, a novel methodology and associated software tool based upon the emerging paradigm of neurosymbolic AI. This paradigm can effectively integrate deep neural models with symbolic domain knowledge from human experts (e.g. physicians), yielding neurosymbolic programs that are more robust, data-efficient, and interpretable than their neural or symbolic counterparts alone. Our preliminary quantitative and qualitative results suggest that Med-Scallop can generate clinically meaningful explanations for clinical decision support (CDS) and facilitate AI-physician collaboration. We propose to assess the performance of our proposed neurosymbolic AI models using two clinical use cases for guiding optimal individualized treatment selection and management for patients with non-small cell lung cancer and sepsis, using multi-institutional EHR datasets for training and validation. Our specific aims are to 1) develop inherently explainable neurosymbolic AI architectures for treatment selection using temporal EHR data; 2) integrate multi-modal clinical data into a neurosymbolic framework for faithful and accurate treatment selection; and 3) develop Med-Scallop, an open-access software platform for neurosymbolic AI for CDS, as well as a comprehensive mixed-methods methodology for assessing clinical validity and usability of explanations from AI models. Our approach represents a new, transformative approach to harness the benefits of both deep neural networks and interpretable symbolic models without suffering their limitations and the power of rich yet complex EHR data.