Automated Triage of Primary Care Referrals to Vascular Surgery using Machine Learning
No Thumbnail Available
Authors
BarrettDanes F.
Premaratne S.
Andrikopoulou, E.
Check for full-text access
Issue Date
2025
Type
Article
Language
Keywords
Alternative Title
Abstract
Objectives: We assessed the feasibility of utilising machine learning (ML) methods to automate triage of vascular referrals from primary care. Method(s): General practitioners (GP) provided hypothetical patient referral letters for peripheral vascular disease and varicose veins, which were augmented to include non-vascular referrals with realistic noise. Vascular surgeons triaged the letters to identify referral type (RT) and clinical urgency (CU). Support vector machine (SVM), k-nearest neighbours (kNN), Random Forest (RF), Multi-Layer Perceptron (MLP), with and without functional API, and transformers (BERT, ClinicalBERT) were used to perform multi-class classification identifying RT and CU. Performance assessed using precision, recall, F1-scores, and t-tests over different dataset configurations. Result(s): Sixty-three GP letters were augmented to 408 letters by synthetic letter generation. BERT achieved the highest combined F1 score (0.8776 +/- 0.0156) on the full dataset, followed by SVM (0.8345 +/- 0.0402), ClinicalBERT (0.8309 +/- 0.0204), MLP (0.8112 +/- 0.0434), RF (0.8043 +/- 0.0347), kNN (0.7821 +/- 0.0456), and Functional API (0.7670 +/- 0.0164). For CU classification, BERT demonstrated superior performance (0.7579) compared to SVM (0.6716), ClinicalBERT (0.6540), MLP (0.6325), RF (0.6086), kNN (0.5924), and Functional API (0.5617). While combined scores appeared similar, statistical analysis revealed significant differences between models. Excluding non-vascular cases maintained consistent model rankings demonstrating generalisability. Conclusion(s): Transformer-based and classical machine learning models demonstrate viability for automated vascular referral triage. Further, larger studies with real patient-level referral letters are recommended.
Description
Citation
Publisher
License
Journal
BMJ Heal care.inf.
Volume
Issue
1