Prediction of gastrointestinal active arterial extravasation on computed tomographic angiography using multivariate clinical modeling.

Aims: To evaluate the ability of logistic regression and machine learning methods to predict active arterial extravasation on computed tomographic angiography (CTA) in patients with acute gastrointestinal hemorrhage using clinical variables obtained prior to image acquisition.

Materials And Methods: CT angiograms performed for the indication of gastrointestinal bleeding at a single institution were labeled retrospectively for the presence of arterial extravasation. Positive and negative cases were matched for age, gender, time period, and site using Propensity Score Matching. Clinical variables were collected including recent history of gastrointestinal bleeding, comorbidities, laboratory values, and vitals. Data were partitioned into training and testing datasets based on the hospital site. Logistic regression, XGBoost, Random Forest, and Support Vector Machine classifiers were trained and five-fold internal cross-validation was performed. The models were validated and evaluated with the area under the receiver operating characteristic curve.

Results: Two-hundred and thirty-one CTA studies with arterial gastrointestinal extravasation were 1:1 matched with 231 negative studies (N=462). After data preprocessing, 389 patients and 36 features were included in model development and analysis. Two hundred and fifty-five patients (65.6%) were selected for the training dataset. Validation was performed on the remaining 134 patients (34.4%); the area under the receiver operating characteristic curve for the logistic regression, XGBoost, Random Forest, and Support Vector Machine classifiers was 0.82, 0.68, 0.54, and 0.78, respectively.

Conclusion: Logistic regression and machine learning models can accurately predict presence of active arterial extravasation on CTA in patients with acute gastrointestinal bleeding using clinical variables.