EP754 - Leveraging AI to Enhance Perioperative Risk Stratification for High-Risk Geriatric Gastrointestinal Surgery Patients

Purpose/Background: Over one third of U.S. surgeries are performed on patients aged 65 and older, and these frail patients face higher risk of wound infection, kidney dysfunction, and deep venous thrombosis. Traditional tools, such as the Risk Assessment and Prediction Tool (RAPT) and Frailty Index (FI), require time-intensive chart review or interviews, limiting scalability. The Centers for Medicare and Medicaid Services TEAM program now incentivizes preoperative optimization, so efficient risk stratification is essential. We developed an automated AI system (the FLEX Score) using Electronic Health Records (EHR) data to predict risk for adverse surgical outcomes. Here, we describe how the AI-based FLEX Score was integrated into GI surgical workflows to stratify risk of poor outcomes, guide perioperative referrals, and streamline clinical decision-making.

Methods/Interventions: From May to September 2025, an EHR pipeline screened all scheduled GI surgeries at our institution. Weekly EHR queries extracted all ICD and CPT codes for each patient 30 days preoperatively. Using these, the FLEX Score model produced individualized risk predictions for four key outcomes: (1) predicted length of stay, (2) readmission within 30 days of discharge, (3) discharge to non-home, and (4) risk of 30-day mortality. Clinically validated thresholds flagged high-risk patients for review by a designated surgeon. These flagged patients were then referred to the Perioperative Optimization for Senior Health Clinic (POSH), where a multidisciplinary framework was used to develop individualized optimization plans (Figure 1).

Results/Outcomes: Over the subsequent five months, the FLEX Score screened over 2,300 surgical patients for all 20 GI surgeons which identified 25 high risk patients who were referred to the POSH clinic. Our rapid integration was contingent upon clinician engagement: surgeons were emailed weekly with their flagged patients and could opt-out of POSH referrals. Cutoffs were adjusted based on institutional data and reviewed by colleagues, promoting near-immediate buy-in. The FLEX Score substantially decreased the burden of chart review: one designated surgeon reviewed only flagged patients rather than entire surgical rosters.

Conclusion/Discussion: The integration of the AI FLEX Score model into GI surgical workflows facilitated individualized timely review of over 2,300 patients with minimal clinician time. The workflow was feasible and well-received when surgeons were active stakeholders. Key elements such as the designated surgeon reviewer and an opt-out referral structure lowered barriers to adoption. The success of our implementation underscores that the key to adopting AI tools in surgical care is not only dependent on the model's performance, but equally on clinician engagement and governance. Future work will quantify performance metrics in the GI surgery population and assess its impact on surgical outcomes.

AI-Based Automated FLEX Score Workflow
Figure 1: Workflow diagram of the Implementation of the FLEX Score into the GI surgical workflow