AI and High-Grade Glioma for Diagnosis and Outcome Prediction: Do All Machine Learning Models Perform Equally Well?

• DOI: 10.3389/fonc.2021.601425
• Publication Date: 2021-11-30T12:10:52Z
• Authors: Luca Pasquini; Antonio Napolitano; Martina Lucignani; Emanuela Tagliente; Francesco Dellepiane; Maria Camilla Ros...; Matteo Ritrovato; Antonello Vidiri; Veronica Villani; Giulio Ranazzi; Antonella Stoppac...; Andrea Romano; Alberto Di Napoli; Alessandro Bozzao
• Keywords: Isocitrate dehydrogenase; Boosting; AdaBoost

Radiomic models outperform clinical data for outcome prediction in high-grade gliomas (HGG). However, lack of parameter standardization limits applications. Many machine learning (ML) radiomic employ single classifiers rather than ensemble learning, which is known to boost performance, and comparative analyses are lacking the literature. We aimed compare ML predict clinically relevant tasks HGG: overall survival (OS), isocitrate dehydrogenase (IDH) mutation, O-6-methylguanine-DNA-methyltransferase (MGMT) promoter methylation, epidermal growth factor receptor vIII (EGFR) amplification, Ki-67 expression, based on features from conventional advanced magnetic resonance imaging (MRI). Our objective was identify best algorithm each task. One hundred fifty-six adult patients with pathologic diagnosis HGG were included. Three tumoral regions manually segmented: contrast-enhancing tumor, necrosis, non-enhancing tumor. extracted a custom version Pyradiomics selected through Boruta algorithm. A Grid Search applied when computing ten times K-fold cross-validation (K=10) get highest mean lowest spread accuracy. Model performance assessed as AUC-ROC curve values 95% confidence intervals (CI). Extreme Gradient Boosting (xGB) obtained accuracy OS (74,5%), Adaboost (AB) IDH mutation (87.5%), MGMT methylation (70,8%), expression (86%), EGFR amplification (81%). Ensemble showed across tasks. High-scoring shed light possible correlations between MRI tumor histology.