Predicting Gastrointestinal Bleeding on CT Angiography: A Comparison of Logistic Regression and Machine Learning Models

Sunday, March 5, 2023

3:36 PM – 3:45 PM

Location: Phoenix Convention Center, 224AB

Presenting Author(s)

Ishan Sinha, MD

PGY3 IR/DR Resident
Mount Sinai Hospital

Disclosure(s): Pumas AI: Consultant ()

Author/Co-author(s)

Brett Marinelli, MD

Assistant Attending
Memorial Sloan Kettering Cancer Center
JG

Jason Gruener, MD

Physician
Mount Sinai Hospital
Ariana C. Mills, n/a

Medical Student
Icahn School of Medicine at Mount Sinai
Samuel Z. Maron, n/a

Medical Student
Icahn School of Medicine at Mount Sinai
SH

Shreyas Havaldar, n/a

Data Analyst
Hasso Plattner Institute for Digital Health at the Icahn School of Medicine at Mount Sinai
Vivian Bishay, MD

Associate Professor
Icahn School of Medicine at Mount Sinai Hospital
BG

Benjamin Glicksberg, PhD

Assistant Professor of Genetics and Genomic Sciences
Hasso Plattner Institute for Digital Health at the Icahn School of Medicine at Mount Sinai
RL

Robert A. Lookstein, MD

Professor of Radiology and Surgery
Icahn School of Medicine at Mount Sinai

Disclosure(s):

Ishan Sinha, MD: Pumas AI: Consultant ()

Robert A. Lookstein, MD: No relevant disclosure to display

Purpose: As CTA has become the modality of choice to screen patients with acute gastrointestinal (GI) bleeding prior to transcatheter embolization, a better understanding of the clinical factors that contribute to a positive exam is needed. The purpose of this study was to evaluate the ability of logistic regression and XGBoost models to accurately predict arterial extravasation on CTAs and compare the clinical features that contribute to a positive prediction.

Materials and Methods: Positive and negative CTAs performed for acute GI bleeding at a single institution were retrospectively 1:1 matched using Propensity Score Matching. 36 clinical variables were included in model development including recent history, comorbidities, vitals, laboratory values, and medications. CTAs performed at the main institutional hospital were used for training; studies performed at other sites were reserved for testing. Logistic regression and XGBoost models were developed in Python using Scikit-learn {1}. Model performance was evaluated with the area under the receiver operating characteristic curve (AUROC) on the testing dataset. The five most important variables contributing to a positive prediction were determined by odds ratio and associated p-values for logistic regression and mean impurity decrease for XGBoost.

Results: After matching and data preprocessing, a total of 389 cases were included in the analysis (193 CTAs with active arterial GI extravasation and 196 controls). 255 cases (65.6%) were selected for training, and the remaining 134 (34.4%) cases were held out for external validation. AUROC was 0.82 for the logistic regression and 0.68 for the XGBoost model. The most important variables for predicting arterial extravasation for each model are shown in the attached table. Overlapping variables for both models were history of diverticulitis, anticoagulation, and transfusion within 24 hours.

Conclusion: Logistic regression outperforms XGBoost in predicting arterial extravasation on CTA for acute GI bleeding. Both models may help elucidate factors that contribute to a positive study, which could guide ordering physicians, aid in the triage of studies, and allow for a more efficient use of interventional resources.