Dynamic predictions of patients' length of hospital stay: multilevel modeling of electronic medical record data to augment clinician recommendations"
Ronald L. Fong, Director, Family Medicine Residency Network, Department of Family & Community Medicine.
Keywords: image analysis, visualization, background/noise correction, Matlab, microscopy multilevel modeling, bayesian approaches, natural language processing, deep learning
With the advancement and intersection of modern medicine and electronic medical records, clinicians have access to a suite of laboratory testing results and practitioner assessments to make informed treatment and patient discharge plans. However, this deluge of dynamic data poses a processing challenge in the fast paced hospital environment. We seek to unify these multifaceted, real time data (vital signs, laboratory values, imaging, and physician, nurse and other clinician notes) into a predictive model of patients’ estimated length of stay in the hospital. This model will provide physicians with a tool to navigate through patient data in a timely manner. Extended days of hospital stays compromise patient health and increase costs, and may be preventable with better data science applications. We propose to use Bayesian approaches and/or a convolutional neural network approach, among other methods, to examine records from patients admitted with a diagnosis of community-acquired pneumonia, a common and treatable condition with a FDA definition and standardized instruments predicting mortality risk. We are just beginning this research collaboration and seek feedback on our modeling approaches.