Development of a Multiclass Predictive Baseline for Diverticulum Ultrasound Imaging Types Using Nonlinear Machine Learning

<title>Abstract</title> <p>Context: Diverticulum anomalies exhibit a range of serious clinical manifestations of a similar nature requiring accurate imaging classification to implement appropriate and targeted management. However, clinical data are often characterized by severe class imbalances and convoluted decision boundaries that prevent traditional diagnostic heuristics from being applied effectively. Objective To apply structured clinical variables to create an interpretable and reproducible machine learning framework to serve as a baseline for multiclass predictions related to the imaging types of Meckel's diverticulum. Methods The retrospective observational data used in this study comprised a sample population of 559 patients where each patient had an imaging type label associated with them. The independent predictors used in the models were a set of continuous demographic variables, a set of continuous lesion-specific variables, and a set of categorical variables, which included patient clinical symptoms and hospitalisation history, and ectopic-tissue distribution. Following a rigorous data imputation approach, the dataset was split into a training dataset and a second, independent test dataset using the stratified random-split method. Logistic regression, LASSO, Support Vector Classification, and Random Forests were the four classifiers developed and evaluated using an iterative approach. Results The Random Forest algorithm performed best overall in prediction accuracy for the testing set with an accuracy of 0.4762, balanced accuracy of 0.4567, macro F1 score of 0.3587 and macro AUC of 0.7991. Although the Random Forest showed excellent distinction for the upper-end of high frequency head categories, it was unable to provide similar levels of performance with sparse categories at the lower end of the tail, due to the inherent structural class imbalance within the data. Multivariate logistic regression confirmed that intestinal obstruction by band type retained an independent, statistically significant association with small effusions as the focus category (OR = 0.1456, p = 0.0338). Decision curve analysis confirmed that using Random Forest model outputs produced a consistently higher net clinical benefit than baseline strategies at either extreme. Conclusion The Random Forest algorithm supports a stable, reliable basis upon which multiclass diverticulum imaging types can be classified. Despite the challenges posed by a long-tailed data distribution, the proposed framework is proven practical and viable for use in real-world clinical decision support scenarios.</p>