Machine learning analysis of factors contributing to hypotension after lumbosacral epidural anaesthesia in dogs undergoing abdominal surgery

Abstract

The incidence of hypotension after a lumbosacral epidural in dogs depends on the volume of local anaesthetic administered. So far, there are no reports comparing both methods used to calculate this volume-body weight (BW) and occipito-coccygeal length (OCL)-in veterinary medicine. In this study, we evaluated the effect of these two common dosing strategies on risk of intraoperative hypotension in dogs undergoing abdominal surgery. Data were collected retrospectively from 210 dogs in a veterinary specialist hospital, with 110 receiving OCL-based dosing and 100 receiving BW-based dosing. Comparing logistic regression and ExtraTrees machine learning models (both demonstrating strong predictive performance, with AUROC 0.78 and 0.79, respectively, and supported by external validation on an independent dataset), we identified additional factors associated with post-epidural hypotension, including body temperature, minimum alveolar concentration, sex, and premedication. Our findings revealed that post-epidural hypotension occurred more frequently in the OCL group compared to the BW group (49.1% vs. 33%, respectively, p = 0.025). Despite both dosing strategies being considered effective, the OCL method resulted in a greater incidence of post-epidural hypotension. Our findings highlight the importance of careful individualized risk-benefit analyses for epidural dosing strategy.

Fig. 1

CONSORT-2010 diagram for retrospective observational data inclusion.

Fig. 2

Descriptive statistics: histograms showing the distributions for the continuous variables, coloured according to dogs from the OCL group (blue) and BW group (orange). The p-value comparing between groups is also shown, from Welch’s t-test for all comparisons except mean body temperature, which used Mann-Whitney U due to being normally distributed, according to the Shapiro-Wilk normality test.

Fig. 3

Partial dependence plots, showing the trends of each covariate towards the predicted probability of post-epidural hypotension, visualising the linear associations within the logistic regression model, and non-linear associations from the tree-based machine learning model.

Fig. 4

Shapley additive explanations (SHAP) value plot, showing the influence of each variable within the dataset towards the prediction of post-epidural hypotension under the ExtraTrees estimator. The further to the right, the greater influence towards predicting hypotension, and to the left, against predicting hypotension. The colour indicates the value of the variable involved in the respective prediction. This shows that OCL dosing strategy broadly influences the prediction of the model towards post-epidural hypotension.