ESTRO 2025 - Abstract Book

S3392

Physics - Machine learning models and clinical applications

ESTRO 2025

Results: Mice and Most-frequent techniques improved DT's weighted (balanced) accuracy from 88.6% to 93.6%, while HyperImpute boosted RF's from 92.4% to 96.7%. PATE-GAN achieved 98.5% accuracy for RF, whereas PD-GAN improved DT's to 94.8% (Figure 2). The analysis identified six important prioritization features, with high-indication features like IBD, family history, and positive FIT showing strong correlations with colorectal cancer (CRC), while changes in bowel movement showed minimal correlation, aligning with findings from clinical trials. These findings suggest potential updates to referral criteria.

Conclusion: Employing synthetic data and imputation techniques significantly enhances the performance of AI models on tabular datasets, thereby improving the prioritization of colonoscopy referrals. DT identified the most and least important features for prioritizing patients at higher risk of CRC.