Explainable, Fair, Reproducible and Collaborative Surgical Artificial Intelligence: Integrating data, algorithms and clinical reasoning for surgical risk assessment (XAI-IDEALIST)

点击翻译按钮获取中文摘要

Project Summary In the United States, the average American can expect to undergo seven surgical operations during a lifetime. Each year 150,000 surgical patients die, and 1.5 million develop a complication after surgery. Progress in medical Artificial Intelligence (AI) remains halted by limited datasets and models with insufficient interpretability, transparency, fairness, and reproducibility that are difficult to implement and share across institutions. In the previous funding period, in addition to 98 publications and 3 patents, a real-time intelligent surgical risk assessment system was successfully implemented at University of Florida. The overall objective of this renewal application is to develop a new conceptual framework for “Explainable, Fair, Reproducible, and Collaborative Medical AI” to provide a foundation for clinical implementation at scale. It will leverage the OneFlorida, a large clinical consortium of 22 hospitals serving 10 million patients in Florida, the nation’s third largest state. The overall objective will be achieved by pursuing three specific aims. (1) External and prospective validation of novel interpretable, dynamic, actionable, fair and reproducible algorithmic toolkit for real-time surgical risk surveillance. (2) Developing and evaluating explainable AI platform (XAI-IDEALIST) for real-time surgical risk surveillance using human-grounded benchmarks. (3) Implementing and evaluating a federated learning approach with advanced privacy features for collaborative surgical risk model training. The approach is innovative, because it represents the first attempt to (1) build the first surgical FAIR (Findable, Accessible, Interoperable, Reproducible) AI-ready, large multicenter multimodal dataset, (2) Novel computational approaches accompanied by assessing fairness and reproducibility, (3) a multifaceted and full-stack explainable AI framework, and (4) federated learning capacity for privacy-preserving model trainingacross institutions. The proposed research is significant since it will address several key problems and critical barriers, including (1) lack of AI-ready large surgical datasets, (2) lack of interpretable, dynamic, actionable, fair and reproducible surgical risk algorithms, (2) lack of a medical AI explainability platform, and (4) lack of a systematic approach for collaborative model training and sharing across institutions. Ultimately, the results are expected to improve patient outcomes and decrease hospitalization costs, as well as lifelong complications.