How can data analytics assist in clinical decisions?
Clinical data is collected from a variety of sources, including electronic health records, disease registries, personal health monitoring devices and patient surveys. The healthcare industry is increasingly interested in using this data to assist in doctors' clinical decisions. For example, when deciding what to do about a torn rotator cuff, should the clinical decision be physical therapy or surgery? Frequently, the initial clinical decision involves first using physical therapy to treat patients with a torn rotator cuff, but some patients with severe tears may still need surgery.
Those deferred surgery decisions waste unnecessary physical therapy and leave patients suffering from pain for quite a while. University of Michigan researchers set out to develop an evidence-based clinical decision support system. Ph.D. graduates Weihong (Grace) Guo and Kamran Paynabar (currently an assistant professor at Georgia Institute of Technology), professor Judy Jin of the industrial and operations engineering department and Dr. Bruce Miller and Dr. James Carpenter of the orthopedic surgery department present their findings in the article "A Decision Support System on Surgical Treatments for Rotator Cuff Tears."
Their system operates with a prediction model that can assess the probabilistic degree of a patient eventually requiring surgery. By using their system, a physician simply compares his or her subjective judgment with the model's predicted probability of requiring a surgery. If they are consistent, it would confirm the physician's expert decision; otherwise, it would give a warning and suggest that the physician double check the listed attributes' information used by the model. Therefore, their system not only improves a physicians' decision to reduce the cost of unnecessary physical therapy treatment, it also helps reduce physicians' occasional mistakes due to oversight.
Their decision support system is developed by integrating advanced data analytics methods, including missing data imputation, variable selection and classification/regression analysis. Their newly proposed missing data imputation method is able to handle heterogeneous mixed-type missing data, which is typically a challenging issue in using clinical data. They also recommended a practical approach to determine the probabilistic decision threshold for satisfying the decision risk constraint. The developed methodology has the potential to be extended for other evidence-based clinical decisions.
Their research presents potential broad impacts for improving patient-centric healthcare delivery quality and patient safety.
CONTACT: Judy Jin; firstname.lastname@example.org; (734) 763-0519; Department of Industrial & Operations Engineering, University of Michigan, Ann Arbor, MI 48109-2117