ESTRO 2025 - Abstract Book

S3434

Physics - Machine learning models and clinical applications

ESTRO 2025

Results: The validation cohort showed a pCR rate of 35.6%. From the calibration plot (Fig.1a), we estimated: • calibration in-the-large = 0, indicating no systematic under/overestimation; • slope = 0.84, suggesting a slight overestimation at higher probabilities; • Integrated Calibration Index = 0.3622. The model’s discriminative ability was moderately high, with AUC=0.77 and PR -AUC=0.65 (compared to 0.356 for a random model). The agreement between clinical results and predicted pCR curves, stratified by ERI (Fig. 2) demonstrated consistent performance across four ERI categories: • highly responsive (ERI range 1 – 6.9), pCR=65%, • moderately responsive (ERI 6.9 – 13.1), pCR=55 %, • poorly responsive (13.1 – 36), pCR=14% (standard) and 35% (escalated protocol), • non-responsive (ERI>36), pCR=0% in both regimens. After recalibration [4], the model’s performance improved with AUC=0.79 and PR -AUC=0.69. The slope was 0.99 (Fig.1b), indicating that ERI had a more significant impact on pCR than initially modelled, while dose escalation provided modest improvements in pCR rates.