INNV-13. PREDICTION OF MGMT METHYLATION FOR GLIOBLASTOMAS USING AI MODELS. A COST-EFFECTIVE FRAMEWORK FOR LOW AND MIDDLE-INCOME COUNTRIES

Abstract

Abstract BACKGROUND MGMT methylation (MGMTmet) is prognostic and predictive for patients with glioblastoma (GBM). However, financial constraints limit access to molecular tests, especially in low and middle-income countries. Artificial intelligence (AI) tools have emerged as a potentially cost-effective solution for identifying MGMTmet. Therefore, we are investigating if an AI model could select those at higher risk to present positive results for molecular confirmation, thus making a better resource allocation. METHODS We have selected GBM patients from the public image bank, The Cancer Imaging Archive, and we included all patients with a molecular evaluation of MGMTmet. We defined our volume of interest based on the ESTRO-ACROP 2016 guideline for GBM radiotherapy contouring and then extracted the radiomic features. The collected features underwent structured analysis, and a prediction model for MGMTmet was developed. We used a Random Forest algorithm to identify the most relevant features. Subsequently, the dataset was divided into training and validation sets. RESULTS We examined 100 patients with histologically classified high-grade gliomas and wild-type IDH, Sixty-one were unmethylated, and 39 were methylated. In this cohort, over 39,000 features were extracted. To avoid noise and overfitting, the 13 most predictive features were selected. They belong to the following feature classes: demographic, first order, Gray Level Size Zone Matrix, and Gray Level Co-occurrence Matrix. The preliminary model achieved an accuracy of 80% for predicting MGMTmet. CONCLUSION our preliminary results show a promising role for AI models in screening patients for molecular testing. This initiative demonstrates the utilization of research to bridge the public health gap. By sharing experiences and fostering collaboration, physicians can work towards mitigating healthcare inequalities in low-middle-income countries. In our future perspectives, we plan to enhance the algorithm's efficacy by incorporating qualitative human-generated features and expanding our sample size.