Background A machine learning model (MLM)-generated surgical risk score (SRS) was integrated into the electronic health record to automate real-time predictions of inpatient postoperative complications (POC). We hypothesized that with the integration of the SRS, complication rates would decrease.

Methods Between 2018 and 2024, inpatients undergoing surgery during index hospitalizations were captured. Patients were dichotomized to “SRS-exposed” and “SRS-Non-exposed” based on whether the surgeon had access to the SRS. SRS-exposed patients were dichotomized to “high” or “low” risk SRS. The patient’s SRS was presented to the primary surgeon as a “Best Practice Advisory (BPA)” and InBasket message requiring acknowledgment. Patients were further split into pre- and post-implementation groups and analyzed based on type of surgery (open, laparoscopic, or robotic). Rates of severe POC (Clavien-Dindo Grade ≥3), return to the operating room, and mortality were recorded and compared using proportion tests.

Results 483 inpatients were deemed SRS-exposed. 260 were in the pre-implementation cohort and 223 in the post. Severe complication rates significantly decreased from 58.8% in the pre-implementation cohort to 46% for high-risk patients (p=0.029). In the high-risk open surgery group, complications significantly decreased from 61.5% to 46.6% (p=0.028). These trends were not observed in the SRS-Non-exposed cohort severe complication rates (37.2% to 39.4%, p=0.81). The rate of return to the operating room significantly decreased after implementation in the open high-risk group (17.6% v 5.5%, p=0.009). No significant differences were observed in minimally invasive surgery patients or mortality.

Conclusion Integration of a real-time, automated SRS resulted in a significant decrease in severe post-operative complications, especially in the high-risk group undergoing open operations. Utilization of the MLM SRS may improve clinical outcomes, optimization of medical care, and shared decision-making.