eP446 - Can Machine Learning Models Predict Inadequate Bowel Preparation at Colonoscopy?

ePoster Presenter(s)

• 

  Jeremy A. Balch, MD

  Resident
  University of Florida
  University of Florida
  Gainesville, FL, United States

  All of the relevant financial relationships listed below have been mitigated.
  This individual has no financial relationships with ineligible companies.

Author(s)

• AR

  Ahmed A. Rashid, BA

  Medical Student
  University of Florida
  Gainesville, Florida, United States

  Disclosure information not submitted.
• 

  Pavel Mazirka, MD

  Resident
  University of Florida
  University of Florida
  Gainesville, FL, United States

  All relevant financial disclosures have been managed and will be listed below.
  This individual has no financial relationships with ineligible companies.
• KE

  Kathleen R. Ehresmann, MD

  Resident
  University of Florida
  Gainesville, FL, United States

  All of the relevant financial relationships listed below have been mitigated.
  This individual has no financial relationships with ineligible companies.
• LG

  Lindsey E. Goldstein, MD, FACS, FASCRS

  Chief, Surgical Service NF/SG Veterans Health System
  University of Florida
  Gainesville, Florida, United States

  Disclosure information not submitted.
• KT

  Krista P. Terracina, MD

  Assistant Professor
  University of Florida
  Gainesville, Florida, United States

  Disclosure information not submitted.
• 

  Thomas E. Read, MD

  Professor and Chief, Division of Gastrointestinal Surgery
  University of Florida
  University of Florida College of Medicine
  Gainesville, FL, United States

  All relevant financial disclosures have been managed and will be listed below.
  This individual has no financial relationships with ineligible companies.

Senior Author(s)

• JN

  Johan Nordenstam, MD, PhD

  Associate Professor
  University of Florida
  Gainesville, Florida, United States

  Disclosure information not submitted.

Purpose/Background: Several factors have been associated with inadequate bowel preparation for colonoscopy. We hypothesize that machine learning may better predict patients who have suboptimal preparation, and thus may be candidates for extended prep.Methods/Interventions: Data from a subset of colonoscopy procedures performed between January 2021-January 2022 were retrospectively collected and analyzed using a variety of machine learning algorithms. Sigmoidoscopy, pouchoscopy, and colonoscopy through a colostomy were excluded. Poor bowel preparation was defined as Boston Bowel Prep Score (BBPS) ≤5 as coded by the endoscopist in the procedure report (Provation®). Predictive variables (features) were obtained from review of the electronic health record (Epic®) and included: age, sex, body mass index, diabetes, smoking status, inflammatory bowel disease, constipation, use of laxatives, prior abdominal surgery, and prior hysterectomy. We compared model performance between logistic regression, decision trees, support vector machines (SVM), random forest, xgboost, and a simple neural network. Hyperparameter grid search was optimized for the precision-recall curve to improve performance on the minority class (poor bowel prep). To evaluate the impact of sample size and individual variable contribution to model performance, we generated learning curves and SHAP values.Results/Outcomes: Of the 600 colonoscopies sampled, 14 were excluded for missing BBPS. Of the remaining 586 colonoscopies, 75 (13%) had poor bowel prep. Due to class imbalance of the target variable, we performed up sampling with Synthetic Minority Oversampling TEchnique (SMOTE). Our best performing models failed to reliably predict poor bowel prep (Table 1), with accuracies ranging from 0.37 (neural network) to 0.68 (SVM). Decision trees had the overall best performance in terms of precision, recall, and F1 score. No model was better than chance. Performance was not improved by changing the prediction task from classification to regression. Learning curve analysis did not demonstrate a clear pattern towards improved performance with increasing data, implying that increasing sample size will not improve model performance. Variables with greatest impact on the best performing model were BMI, sex, and laxative use.Conclusion/

Discussion: Machine learning models were not able to adequately predict poor bowel prep for colonoscopy. Learning curve analysis suggested that a larger n may not improve model performance. Additional variables may be beneficial.Table 1Model Performance