In my thesis, I address the above barriers to medical AI development and deployment, namely robustness, data-efficiency and model trust, by presenting three different works that improve upon the current state-ofthe-art for various modalities. First, I propose observational supervision, a novel supervision paradigm wherein we use passively collected, auxiliary metadata to train AI models. I show that leveraging observational supervision for structured Electronic Health Records using audit logs improves performance and robustness of AI models trained to predict clinical outcomes, even with limited labeled training data. Second, I present domain-specific augmentation strategies for self-supervised frameworks that enable large scale, label-efficient training of AI models. I show that such strategies improve performance over datahungry, fully supervised models in chest X-ray classification and generalize to both unseen populations and out-of-distribution data. Third, I present TRUST-LAPSE, an explainable, post-hoc and actionable trust-scoring framework for continuous AI model monitoring. I show that TRUST-LAPSE can determine when amodel’s prediction can and cannot be trusted with high accuracy, can identify when the model encounters classes unseen during training or a change in distribution, and can accommodate various types of incoming data (vision, audio and clinical EEG). Together, these works pave the way for developing and deploying robust, data-efficient and trustworthy medical AI models to improve clinical care. 39 Nandita Bhaskhar Computer Vision News
RkJQdWJsaXNoZXIy NTc3NzU=