Development, external validation and integration into clinical workflow of machine learning models to support pre-operative assessment in the UK

Alwyn Kotzé ^1,2, Tom Lawton ^3,4, Simon J Howell ²

⁰Leeds Teaching Hospitals NHS Trust, Leeds, UK.

Anaesthesia +

|

September 15, 2025

View abstract on PubMed

Summary

This study integrated electronic health records with machine learning to improve pre-operative risk stratification for surgical patients. The developed models accurately identify lower-risk patients, aiding in surgical backlog reduction.

Area Of Science

Medical Informatics
Machine Learning in Healthcare
Surgical Risk Stratification

Background

Increasing surgical demand and patient complexity necessitate improved pre-operative assessment.
Current pre-operative workflows rely on manual tasks and lack real-time data, hindering efficiency and contributing to surgical backlogs.
NHS England mandates risk stratification and optimization for patients awaiting surgery.

Purpose Of The Study

To develop and validate machine learning models for pre-operative risk stratification using real-time electronic health record data.
To integrate these models into a live pre-operative assessment system (Smart PreOp) for direct clinical application.
To address limitations in current pre-operative workflows by enabling automated data retrieval and risk categorization.

Main Methods

Established certified electronic linkages between the Smart PreOp system and NHS England's GP Connect to access general practitioner data.
Developed machine learning models to predict American Society of Anesthesiologists (ASA) physical status and 30-day postoperative mortality risk.
Constrained variable selection to electronically available, real-time data for all UK surgical patients, including procedure, demographics, and medications.

Main Results

External validation on over 178,000 patients demonstrated high precision (0.95) and recall (0.69) for identifying lower-risk patients (ASA physical status 1 or 2).
Mortality prediction models showed good discrimination but required hospital-specific calibration, supporting the need for localized models.
The Smart PreOp system's architecture supports hospital-specific modeling and updates.

Conclusions

Integrated systems development and predictive modeling can produce accurate, implementable electronic health record (EHR) prediction models.
The developed models show promise for direct implementation into EHR systems to enhance pre-operative care.
Further prospective studies are needed to evaluate the clinical impact and acceptability of these integrated tools.

Keywords:

electronic health records machine learning pre‐operative care risk assessment