Automated classification of treatment state of patients with prostate cancer from structured medical record data to help enhance care.

427 Background: The number of US prostate cancer (PCa) cases is projected to expand exponentially while the oncology workforce stagnates. The increase in clinic demand requires innovative technologic approaches for patient management. We are developing a process for automatically tracking patients across the care continuum which can flag patients overdue for follow-up, send alerts when patients recur, and thereby optimize clinic resources. In order to place patients into the correct care pathway, their disease status must be correctly classified, which is the focus of the current study. Methods: Clinician specialist input was used to define five mutually exclusive care states in PCa (newly diagnosed, active surveillance, active treatment, post-treatment surveillance, and metastatic). Using a variable set defined to classify each care state, decision tree modeling was applied to abstracted, de-identified data from 150 patients seen in the Dana-Farber/Brigham Cancer Center radiation oncology clinic between 5/14 and 8/14/2024 to validate model care state classification. Model output and a case summarization were presented in an interactive web-interface designed using user experience principles. Through a secure login, a provider accessed the data for each case and recorded if the care state assignment and case summary were accurate. Results: The concordance between the model classification and provider assignment was 100%. The model correctly sorted the care states of each of the patients, of which the distribution was 0% newly diagnosed, 1% active surveillance, 12% active treatment, 73% post-treatment surveillance, and 13% metastatic The interactive web interface included graphical and table views of PSA test values with automatic calculation of doubling time and velocity. In addition, treatment events (prostatectomy, radiation, hormone therapy) were overlaid on the PSA graph for easy clinical analysis. A treatment summary table provided a chronologic view for each patient. Conclusions: The care state assignment model was able to correctly classify the care state of all 150 patients in this cohort for which structured, provider abstracted data was available. The next steps in clinical implementation will be EMR integration, automatic record abstraction, and end-user testing. We have already begun work to automatically ingest and parse PCa records using natural language processing and expand the user interface to include a patient care journey map, generation of a pre-visit clinical note, and care optimization protocols. The goal is to provide a tool that will decrease time for case review and appointment preparation, give case prioritization alerts based on PSA testing results, and automatically monitor surveillance care compliance thus optimizing clinic resources while ensuring no patient is lost to follow-up.