ESTRO 2024 - Abstract Book

S4579

Physics - Machine learning models and clinical applications

ESTRO 2024

The Random Forest algorithm, combined with SM, outperforms other models with an AUC of 0.93. Most machine learning models, including the neural network, exhibit improved performance over existing metrics (98th percentile and 95th percentile). Notably, the Random Forest model achieves an 88% recall rate, while the neural network achieves a 100% recall rate. The most important feature of the Random Forest model is the previous fraction's 98th percentile SM score. Half of the top 20 features are SM scores derived from previous fractions. Figure 1: ROC curves for different machine learning algorithms. The SM 98% and SM 95% hazard thresholds are the mm thresholds that define the positive vs negative resimulation selection using that SM score. The red 98th percentile SM and blue 95th percentile SM threshold points are plotted on the relevant ROC curves, 98 column and 95 column, respectively.

Figure 2: Top 20 Random Forest Feature Importances. Feature importances are obtained from the random forest machine learning model and ordered and plotted with respect to their Gini importance scores (x-axis). PrevFxn Scores correspond to SM percentile scores from previous treatment fractions. PrevFxn Scores are scores from previous days. The DeliveryMaximumDose, TargetMaximumDose, TargetPrescriptionDose, and FinalCumulativeMetersetWeight are values obtained directly from the DICOM file and pertain to the proton radiation dose. The numerical features, such as 75, and 100, pertain to 75th percentile scores and 100th percentile scores for the maximum of the Surface Mapping scores for the tested fraction.