ADAPT in SC Thrust 1

XAI-Enabled Biomedical Devices for Diagnostic and Planning Applications 

Explainability is crucial to rationalizing and cross-checking model outcomes to ensure that the AI-informed decisions made are reliable and trustworthy. The major challenge to implementing AI in biomedical devices is a tradeoff between explainability and accuracy of the AI models. Highly accurate, complex models like deep neural networks (DNNs) trained by significant amounts of data tend to be less explainable, but explainable algorithms like decision trees usually lead to low accuracy for complex tasks. As an emerging field in trustworthy AI, XAI endeavors to find explanations for complex models using ante hoc or post hoc methods. Ante hoc methods address explainability from the beginning, whereas post hoc methods rely on an external explainer of an already trained model. We will use XAI to improve explainability of the diagnostic or treatment decisions made from multimodal clinical data to provide insights into important causal factors and to obtain domain experts’ trust, high prediction accuracy, and safe, continuous workflows from initial diagnosis to treatment end.We will use multimodal, longitudinal, and cross-sectional data from the available clinical database to develop an XAI-driven biomedical device for decision making. To address SC’s healthcare deficiency, we will explore XAI-enabled biomedical device development through implementation of XAI in two novel devices chosen from those being developed by the Thrust 1 team: one for impedance-based wound assessment and the other for microchamber-based antibiotics analyses. We will advance the fundamental science and engineering that enables optimal sample preparation, applications of novel sensors, and point-of-care imaging techniques with the capacity to increase objectivity and efficiency of clinical decision making, especially for use in diagnostics. 

Goals for Thrust 1