QS212 - Predicting High-Grade Anal Dysplasia Using a Novel Machine Learning Approach

Quick Shot Presenter(s)

• 

  JC Chen, MD (they/them/theirs)

  Resident
  Ohio State University
  Ohio State University
  Columbus, OH, United States

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

Author(s)

• AA

  Abdullah Altaf, MD

  Research Fellow
  The Ohio State University
  Columbus, Ohio, United States

  Disclosure information not submitted.
• 

  Evan D. Adams, MD (he/him/his)

  Fellow
  The Ohio State University
  Columbus, OH, United States

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

  Kent J. Peterson, MD

  Fellow
  Ohio State University
  Columbus, OH, United States

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

  Alan E. Harzman, MD

  Professor
  The Ohio State University
  Columbus, Ohio, United States

  Disclosure information not submitted.
• 

  Amber L. Traugott, MD (she/her/hers)

  Associate Professor
  The Ohio State University
  The Ohio State University Wexner Medical Center
  Columbus, Ohio, United States

  Disclosure information not submitted.
• 

  Lisa A. Cunningham, MD

  Assistant Professor
  The Ohio State University
  Columbus, Ohio, United States

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

  Emily Huang, MD

  Assistant Professor
  The Ohio State University
  Columbus, Ohio, United States

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

  Timothy Pawlik, MD PhD

  Department Chair
  Ohio State University
  Columbus, Ohio, United States

  Disclosure information not submitted.
• 

  Matthew F. Kalady, MD

  Professor, Division Director,
  The Ohio State University Wexner Medical Center and James Cancer Center
  Ohio State University
  Columbus, Ohio, United States

  Disclosure information not submitted.

Senior Author(s)

• 

  Syed Husain, MD

  Professor
  The Ohio State University
  Ohio State University
  Columbus, Ohio, United States

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

Purpose/Background: The incidence of anal cancer has doubled over the past few decades. Recent trials suggest prompt identification of patients at risk for high-grade anal dysplasia (HGD) is essential for timely intervention and prevention of progression to anal cancer. The current study sought to develop a novel machine learning (ML) model to identify individuals at risk of harboring HGD who would benefit from routine screening of HGD.

Methods/Interventions: Patients ≥18 years old who underwent high-resolution anoscopy (HRA) for abnormal anal cytology between 2017-2024 at our institution were identified. HGD was defined as biopsy-proven anal intraepithelial neoplasia (AIN)-2 or AIN-3. An eXtreme Gradient Boosting (XGBoost) ML model was trained and validated using nine clinicodemographic variables (age, sex, sexuality, marital status, smoking status, HIV status, other high-risk comorbidity status, high-risk HPV status, and HPV vaccination status) associated with HGD from recursive feature elimination. Other high-risk comorbidities included conditions requiring immunosuppression (history of transplant, autoimmune disease, active cancer) and history of genital or anal dysplasia.

Results/Outcomes: A total of 468 patients were randomized into training (n=328, 70.0%) and testing cohorts (n=140, 30.0%). HGD was seen in 53.4% (n=250) of the entire cohort, 51.5% (n=169) in the training and 57.9% (n=81) in the testing cohort (p>0.05). The XGBoost model demonstrated reasonable accuracy to predict HGD with AUC-ROC 0.761 (95% CI: 0.711-0.811) in the training cohort and 0.748 (95% CI: 0.665-0.830) in the testing cohort. Moreover, sensitivity and specificity of the model was 77.3% and 69.7%, with positive and negative predictive values of 74.4% and 73.0%, respectively (Figure 1). The SHapley Additive exPlanations (SHAP) algorithm suggests that in descending order, high-risk HPV status, presence of other high-risk comorbidities, smoking status, age, and HIV status were the most strongly predictive risk factors for harboring HGD (Figure 2). The XGBoost model was made available online for ease-of-use and wide clinical applicability (https://altaf-husain-dysplasiamodel.streamlit.app/).

Conclusion/Discussion: The current XGBoost model offers a robust and clinically applicable tool for identifying patients who may be at risk of harboring HGD. By leveraging readily available clinicodemographic data, the model offers valuable insight to enable clinicians to accurately identify high-risk individuals and facilitate timely, targeted interventions like HRA. Use of the model may enhance early detection and mitigate progression to invasive cancer, thereby improving patient outcomes. Further validation in external cohorts is needed to enhance generalizability and support widespread adoption of the model into clinical practice.

Figure 1. Receiver operating characteristic (ROC) curves of the XGBoost Model to predict risk of HGD within training and testing cohorts.

Figure 2. SHAP summary plots for the XGBoost Model demonstrating factor weightage to predict HGD.
Positive SHAP values signify factors that predict a higher risk of anal dysplasia, while negative values indicate factors that predict a lower risk. Blue represents lower factor values (e.g., negative high-risk HPV status or single marital status), and red denotes higher values (e.g., positive HIV infection or current smoker). The y-axis measures the overall importance of each factor.