Operationalization of Artificial Intelligence to Assist in Surgical Discharge: A Feasibility Case Study

AbstractArtificial intelligence (AI) systems have the potential to support the identification of patients ready for safe postoperative discharge, but translating AI predictions into clinical practice requires careful evaluation of both technical and clinical factors. This study assessed the technical and clinical feasibility of operationalizing the predictive DESIRE (Discharge after Surgery Using Artificial Intelligence) model for postoperative day-2 discharge readiness. Over a 3-month silent trial at a single academic center, DESIRE generated discharge predictions for 132 adults undergoing gastrointestinal or oncologic surgery, with clinicians blinded to AI outputs. The model produced real-time predictions for 114 of 132 patients (86%). However, corrupted data related to surgery duration temporarily affected reliability. DESIRE achieved an area under the curve of 0.77 (95% confidence interval, 0.67 to 0.85), high specificity (98%), and positive predictive value (92%), but low sensitivity (14%), reflecting the intentional conservative approach to identifying safe discharges. In total, 12 patients deemed "safe" by the model required hospital-specific care that was not included in the AI's definition of safe discharge, including 11 patients with respiratory insufficiency. Overall, the study demonstrates that AI-driven discharge prediction is technically and operationally feasible but remains highly dependent on data quality, governance, and clinical oversight. Strengthening real-time data validation and establishing robust monitoring frameworks will be critical to ensuring AI's trustworthiness and reliability in clinical practice.