(ID27) Diagnostic Application of Machine Learning Models for Predicting Feline Infectious Peritonitis

Background: Antemortem diagnosis of feline infectious peritonitis (FIP) often requires diagnostics that may be costly, invasive, and delay curative treatment. Affordable techniques enabling rapid and confident identification of affected cats are essential for timely initiation of antiviral therapy, enhancing cat welfare by minimizing treatment delays.

Objective: To develop machine learning models (MLM) trained on clinicopathologic data to predict a diagnosis of FIP.

Animals: 133 cats with histopathology and FCoV-immunohistochemistry confirmed FIP and 133 age-matched and effusion-matched control cats suspected of FIP with alternative diagnoses.

Methods: A multi-institutional, retrospective, case-control study was conducted using records from 2002-2024. Clinical features, normalized hematology and serum biochemistry were collected. 5-fold cross-validation was used to split, train, and test the data to evaluate logistic regression (LR), support vector machine (SVM), decision tree (DT), and random forest (RF) classification models. Model performance statistics on the test were evaluated and feature importance was determined where possible.

Results: MLMs demonstrated strong predictive performance for identifying cats with FIP, with area under the receiver operating curves of 0.809-0.908 (Table 1). The top-performing model, random forest, achieved a sensitivity of 83.46% (95% confidence interval [CI]; 74.81-88.85%), specificity of 83.46% (95% CI; 73.55-95.43%), and harmonic mean of precision and recall (F1 score) of 0.8355 for predicting FIP. Serum globulin concentration, eosinophil count, and serum bilirubin concentration were consistently the top contributing features for evaluated models.

Conclusion: MLMs trained on clinicopathologic data from cats with a high pretest probability of FIP can accurately predict a diagnosis and may aid in clinical decision-making.