This project harnesses the growing amount of data that is captured in the electronic health record to discover the optimal diagnostic pathway for an individual patient. A multidisciplinary team with expertise in decision modeling, radiology/clinical practice, informatics, and machine learning will investigate approaches that transform data into actionable knowledge, enabled by a new class of clinical decision support algorithms that actively learn from available clinical data. The objective of this project is to develop and evaluate a data-driven framework for decision support that helps clinicians to deliver individualized patient care by discovering optimal sequences of actions and to diagnose patients in a timely, accurate, and cost-effective manner. The project addresses challenges related to finding relevant information from large, longitudinal patient data; learning sequences of actions from past patient cases; and handling uncertainty that is inherent to the practice of medicine. The new algorithms will improve how observational clinical data can be used to generate evidence that improves healthcare delivery, efficiency, and ultimately, realizes precision medicine and improves patient outcomes. A diverse group of graduate students will be trained in an interdisciplinary manner to translate algorithms and data science concepts into applications that have real-world clinical utility, and with a clear understanding of the technical and cognitive challenges.
The proposed research will create a generalizable framework for learning from healthcare data to discover optimal actions for individual patients with the following objectives: (1) to determine what combination of diagnostic procedures (e.g., imaging, labs, biopsy) should be used to achieve an accurate and timely diagnosis, and in what sequence; and (2) to demonstrate that learning such pathways can be done using available data, allowing the new methodology to be applied in a wide range of clinical domains. Novel aspects of this project are three-fold: (1)Dealing with an environment that is unknown and changing over time in unpredictable ways through a novel adaptive learning approach to discover the most informative features that are predictive of subsequent actions taken in real-time. This builds upon earlier work in relevance learning to dynamically elucidate the relationships between clinical features and possible actions. (2)Developing a new type of bandit algorithm that not only discovers the next best diagnostic test to order, but also identifies additional information that is needed to make a definitive diagnosis. The accuracy and value of diagnostic tests are dependent on many factors (e.g., technology, patient characteristics). The team will assess how prior information from similar patients can speed-up learning given these factors. (3)Providing confidence bounds about the risks and benefits of selecting a specific diagnostic exam to perform. These confidence bounds can be easily understood by clinicians. The performance and utility of this approach will be demonstrated using a prospective study that solicits physician feedback about specific recommendations for a given patient case and learns from situations in which the physician does not follow the system's recommendation.