A point-based laminitis risk scoring system and machine learning framework for prediction of laminitis in horses.

Overview

• This study developed a point-based risk scoring system combined with machine learning models to predict laminitis, a painful hoof disease in horses, using routine clinical data.
• The approach aims to improve early detection and support veterinary decision-making through interpretable risk scores and predictive models.

Introduction to Laminitis and Research Objective

• Laminitis is a painful and potentially life-threatening inflammatory condition affecting the hoof of horses.
• Early detection is difficult, especially in the subclinical (before obvious symptoms) stage.
• Existing knowledge has identified several risk factors but routine clinical data have been insufficient for accurate early prediction.
• The study’s goal was to develop an interpretable machine learning framework to predict the risk of laminitis using clinical, physiological, and conformational horse data.

Methodology

• Point-based Risk Scoring System:
  • Used multivariate logistic regression to quantify the contribution of various clinical variables to laminitis risk.
  • This system assigns risk points based on factors measured during routine examinations.
• Machine Learning Models:
  • Combined the statistical risk scores with multiple machine learning classifiers to improve prediction accuracy.
  • Evaluated nine different classifiers including SVM (Support Vector Machines), TabPFN, and Random Forest.
• Explainability Techniques:
  • Applied tools like ELI5 and feature importance analysis to each model to make predictions transparent and interpretable.
  • This helps clinicians understand which factors influence the models’ assessments and trust their outputs.

Key Results

• Performance Metrics:
  • SVM achieved the highest F1-score of 0.858 ± 0.152 and Matthews Correlation Coefficient (MCC) of 0.811 ± 0.211, indicating a strong balance between precision and recall.
  • TabPFN model achieved the highest Area Under the Curve (AUC) with 0.932 ± 0.094, signaling excellent discrimination between laminitis presence and absence.
  • Random Forest demonstrated the highest precision at 0.917 ± 0.180, reflecting fewer false positives.
• Risk Score Outcomes:
  • Risk scores reached as high as 0.97, indicating close to certain risk in some cases.
  • Most influential predictors across all models were:
    • Lameness examination right fore (LERF)
    • Hoof testers right hind (HTRH)
    • Digital pulses
    • Rectal temperature
    • Age of the horse

Conclusions and Implications

• The integration of logistic regression–based risk scoring with interpretable machine learning methods creates an accurate and clinically meaningful tool for laminitis risk prediction.
• This framework offers practical decision support for veterinarians, potentially aiding in preventive management strategies to improve equine welfare.
• The transparency provided by explainability tools can increase clinical acceptance and facilitate better understanding of laminitis risk factors.
• The approach is supported by open-source code, encouraging further research and adoption within the veterinary community.

Additional Notes

• The availability of source code on GitHub aids reproducibility and customization by other researchers or practitioners.
• Future research may explore expanding datasets, real-time clinical integration, or improving models with additional biological markers.