Machine Learning Clinical Order Recommendations for Specialty Consultation Care

点击翻译按钮获取中文摘要

Summary: Machine Learning Clinical Order Recommendations for Specialty Consultation Care A future vision of clinical decision support must transcend constraints in scalability, maintainability, and adaptability. The shortage of 100,000 physicians by 2030 reflects unmet (and unlimited) demand for the scarcest healthcare resource, clinical expertise. Over 25 million in the US alone have deficient access to medical specialty care, with delays contributing to 20% higher mortality. There is no quality without access. Our goal is to develop a radically different paradigm for outpatient specialty consultations by inductively learning clinical workups embedded in clinical data. We focus on predicting the concrete clinical orders for medications and diagnostic tests that result from specialty consultations. This can power a tier of fully automated guides that will enable clinicians to initiate care that would otherwise await in-person specialty visits, opening access for more patients. The major scientific barriers are advances in data science and decision support methods for collating clinical knowledge, with continuous improvement through clinical experience, crowdsourcing, and machine learning. Our innovative approach is inspired by collaborative filtering algorithms that power “Customers like you also bought this...” recommender systems with the scalability to answer unlimited queries, maintainability through statistical learning, and adaptability to respond to evolving clinical practices. Our team uniquely combines expertise in clinical medicine, electronic medical records, clinical decision support, statistics and machine learning to enhance medical specialty consultations through aims that seek to: (1) Develop methods to generate clinical decision support by predicting the clinical orders that will result from Endocrinology and Hematology specialty consultations; (2) Evaluate and iteratively design clinical collaborative filtering prototypes based on clinical user input on usability and acceptability; and (3) Determine which consult clinical order patterns are associated with better results through reinforcement learning and causal inference frameworks. Completion of these aims will yield a sustained, powerful impact on clinical information retrieval and knowledge discovery for synthesizing clinical practices from real-world data. By addressing grand challenges in clinical decision support, adoption of these methods will fulfill a vision that empowers clinicians to practice to the top of their license, making healthcare more scalable in reach, responsiveness, and reproducibility