A Dynamic Machine Learning Approach to Complement Nurse-Led Clinics in Identifying High-Risk Patients with...

Area of Science:

• Vascular Medicine
• Artificial Intelligence
• Predictive Analytics

Background:

• Intermittent claudication (IC) management requires accurate prediction of adverse outcomes.
• Timely intervention and surveillance are crucial for patients with IC to prevent severe complications.

Purpose of the Study:

• To develop and validate dynamic machine learning (ML) models for predicting 6-month adverse events in IC patients.
• To enable risk-stratified prioritization for enhanced surveillance and prompt revascularisation.

Main Methods:

• Utilized a prospectively maintained nurse-led IC clinic registry (2020-2024) with 629 patients.
• Developed ensemble ML models (LASSO, gradient boosting, random forest, XGBoost, stacking) using diverse predictors.
• Assessed model performance via discrimination (AUC) and calibration (Brier score).

Main Results:

• Ensemble models achieved strong predictive performance: AUCs of 0.712 (revascularisation), 0.754 (CLTI), and 0.828 (mortality).
• Key predictors included ankle-brachial pressure index (ABPI), walking distances, age, comorbidities (renal disease, COPD), smoking, and frailty.
• Models provided calibrated 6-month risk estimates, outperforming earlier neural network approaches.

Conclusions:

• Ensemble ML models accurately estimate 6-month risks for revascularisation, CLTI, and mortality in IC patients.
• These models can support prioritization for surveillance and early escalation of care in nurse-led clinics.
• External validation and prospective impact studies are recommended next steps.