#OA4-AM25-MN-19 - Prospective Validation of an Artificial Intelligence Clinical Decision Support System for Preoperative Blood Orders

Preoperative preparation for transfusion is important for timely transfusion during surgery; however, excessive preparation is common, costly, and contributes to blood waste. Accurate estimation of surgical transfusion risk is critical for perioperative planning and resource stewardship. However, current methods for risk stratification – i.e., Maximum Surgical Blood Ordering Schedule (MSBOS) – don’t consider patient factors that contribute to surgical transfusion risk.

A personalized machine learning model, referred to as S-PATH, was previously developed to predict red cell transfusion during surgery. The purpose of this study was to implement S-PATH as a clinical decision support (CDS) tool embedded within the electronic health record (EHR), and to prospectively evaluate the tool’s recommendations for preoperative type and screen (T&S) orders compared with usual care.

Study

Design/Methods:

This study was conducted at Barnes-Jewish Hospital, a large academic medical center, and approved by the Institutional Review Board (#202206170). Preoperative blood orders are typically placed by clinicians working in the preoperative assessment clinic, guided by an MSBOS.

The S-PATH model was integrated within the EHR using the EHR vendor’s cloud computing platform (Cognitive Compute, Epic Systems, Verona, WI). Model predictions were automatically generated at 2AM for all patients scheduled for an in-person preoperative assessment clinic visits on that day. Model predictions were silently stored in a flowsheet row and were not available for clinicians to view. Model predictions were retrieved from the EHR reporting database and compared with observed T&S orders placed during usual care. Data was collected for all patients with surgical encounters occurring between November 11, 2024 and March 31, 2025. Analyses were performed using R 4.1.2. Statistical significance was evaluated using the Chi-square test.

Results/Findings:

This study included 5309 patients. Model predictions were available for 73% (3887/5309); model predictions were not available for walk-in appointments and for patients having rare procedures.

S-PATH recommended T&S orders for 43% (1801/3887) patients. Usual care resulted in 71% (2765/3887) patients having an active T&S by the start of surgery, an absolute difference of 25 percentage points (p < 0.001).

Nine percent (364/3887) of patients required red cell transfusion during surgery. Usual care missed 18% (66/364) of these patients, meaning they did not having an active T&S by start of surgery; S-PATH identified all 66 these patients as needing a T&S. In contrast, S-PATH missed 5% (19/364), consistent with the model’s intended design.

Conclusions:

An artificial intelligence-driven CDS tool for preoperative blood orders performed well in prospective validation, resulting in safer and more efficient use of preoperative type and screen orders compared with usual care.