ESTRO 2024 - Abstract Book

S4477

Physics - Machine learning models and clinical applications

ESTRO 2024

with higher weights for minority classes and an upsampling technique called SMOTE[1] for priority level and cancer prediction tasks, respectively.

Results:

Among all the features, the decision tree primarily relies on patients' medical history, including pathology information and the presence of inflammatory bowel disease (IBD), and demographic factors, such as BMI and age, to predict cancer/non-cancer status and categorize priority levels. While previous studies mainly relied on the FIT value in addition to full blood test results[2, 3], the unavailability of the FIT score in Québec province (where FIT>175 defines a positive result and <175 define a negative result) and the missing information for the blood test results in this study posed challenges for the model's prediction and classification. Despite the challenges related to missing data, the decision tree achieved a test accuracy of 99.3% and 92.9% besides a balanced accuracy of 94.7% and 88.6% (considering the imbalanced nature of the dataset) for cancer prediction and prioritization, respectively. For the task of cancer prediction, given the fact that only 10 test samples have cancer, the model achieved a recall of 99.3% for the negative class and 90% for the positive class. The recall metric quantifies the proportion of positive/negative samples correctly identified among all the positive/negative samples. Regarding the task of prioritization, priority classes with a minimum of 5 test samples and a maximum of 1,803 test samples, defining the immediate and semi-elective patients, achieved a recall of 100% and 89.6%, respectively (see Figure 1). Additionally, a random forest model, an ensemble of 50 decision trees, achieved a balanced accuracy of 95.0% and 92.4% for the cancer prediction and prioritization tasks, respectively.

Figure 1: Precision and Recall for each priority class on the test data.

Conclusion: