Integration of machine learning and viscoelastic testing to improve survival prediction in horses experiencing acute abdominal pain at a veterinary teaching hospital

Background: Viscoelastic coagulation testing (VCT) identifies subclinical disruption of coagulation homeostasis and may improve prognostication, particularly for patients with severe systemic inflammation or shock. Machine learning (ML) algorithms may capture complex relationships between clinical variables better than linear regression (GLM). Objective: To evaluate the utility of ML models incorporating VCT and clinical data to predict survival outcomes in horses with acute abdominal pain. Study Design: Retrospective observational cohort study. Methods: VCT (VCM Vet™) was performed on 57 horses with acute abdominal pain at admission, with clinical data collected retrospectively. Coagulopathy was defined as ≥2 abnormal VCT parameters. GLM and random forest (RF) classification models were developed to predict short-term survival. A training cohort of 40 horses was used for model development, and model performance was determined using the remaining 17 horses. RF models were implemented in a web-based application to demonstrate clinical application. Results: There were 31 survivors and 26 non-survivors. The majority of cases were colitis (47.7%), with smaller proportions of impactions, strangulating obstructions and other causes of colic. Coagulopathy diagnosis alone performed poorly for survival prediction (sensitivity 81% [95% CI 64–94], specificity 31% [95% CI 15–50], AUC = 0.515). Final GLM included SIRS score (OR 0.37 [95% CI 0.071–1.68]; p = 0.2), L-lactate (OR 0.51 [0.25–0.82]; p = 0.02), clot time (CT; OR 1.0 [0.99–1.0], p = 0.39), and clot amplitude at 10 min (A10; OR 0.89 [0.74–1.02], p = 0.2). Final RF model included heart rate, PCV, L-lactate, white blood cell count, neutrophil count, clot amplitude at 20 min (A20) and CT. RF models improved sensitivity (RF_full 91% [95% CI 60–100]; RF_reduced 83% [95% CI 42–99]) and specificity (both 83% [95% CI 42–99]) compared to GLM (sensitivity 65% [95% CI 47–79], specificity 42% [95% CI 26–61]). Main Limitations: Small number of horses, convenience sampling. Model validation with an independent population is needed to support clinical applicability. Conclusions: L-lactate remains a key predictor of survival in horses with colic. The integration of VCT with clinical data in machine learning models may enhance prognostication.