ESTRO 2024 - Abstract Book

S4476

Physics - Machine learning models and clinical applications

ESTRO 2024

1402

Digital Poster

an artificial intelligence-based cancer prediction and triage tool in GI endoscopy units

Laya Rafiee Sevyeri 1 , Myriam Martel 2 , Alan N Barkun 3 , Shirin Abbasi Nejad Enger 1

1 McGill University, Gerald Bronfman Department of Oncology, Montréal, Canada. 2 McGill University, Research Institute of the McGill University Health Center, Montréal, Canada. 3 McGill University, Division of Gastroenterology, Montréal, Canada

Purpose/Objective:

We prioritize patients for colonoscopy procedures using an AI-based model that relies on demographic, pathological, and past medical information. This project assesses the performance of a decision tree model for prioritizing colonoscopy referrals to address staff shortages and reduce the backlog of colonoscopies accumulated since the onset of the COVID-19 pandemic. Additionally, the model's performance in classifying patients into cancerous and non-cancerous categories is examined.

Material/Methods:

Data from 14,657 patients, comprising 7,469 females and 7,188 males, who were referred to two tertiary centers for colonoscopy from September 2018 to August 2022 in Québec, Canada, were included. The data includes findings of colorectal cancer in 127 patients, representing less than 1 percent of the total study population. Patients’ ages are 18 99 years, with a median of 60 years. The average waiting time is 126.5 days, with a maximum waiting time of 1,946 days. The dataset contains demographic patient information, waiting times (from the day of referral to the day of procedure), and detailed diagnostic findings for all patients. Fecal immunochemical test (FIT), blood test, and imaging results (CT scan, MRI, and Ultrasound) are available for only 2.6%, 28%, and 6% of patients, respectively. In Québec, patient referrals for colonoscopy are facilitated through a referral sheet known as the AH-702, which includes six validated priority levels (with corresponding indications within each group) based on screening, surveillance, or case finding indications specified by the referring physician. Wait time assignments are prioritized based on the most urgent indication, if multiple, for referral in a given patient. In this study, the highest and second-highest priority levels, identifying patients requiring immediate and urgent care, consist of only 20 and 308 samples, respectively. Approximately 34.5% of patients can wait up to 60 days for a semi elective procedure, while the remaining patients fall into either the elective category (> 60 days), screening, or are follow-up referrals. For this study, we utilized a well-known classical white-box machine learning model, a decision tree with a depth of 7. To evaluate the model's performance on this dataset, we trained it on 75% of the data with 5-fold cross-validation and tested it on the remaining 25% for two specific tasks: cancer prediction and prioritization. Missing values were substituted with -1 and imaging outputs were modified into different classes representing each diagnosis. Given the sparse nature of the dataset with numerous missing values for each feature, we employed a regularization technique