A deep learning model to predict glioma recurrence using integrated genomic and clinical data

Area of Science:

• Oncology
• Computational Biology
• Genomics

Background:

• Gliomas represent a significant portion of primary brain tumors, with survival rates highly dependent on grade.
• Recurrence is frequent in gliomas, impacting patient prognosis and complicating treatment.
• Predictive models for early glioma recurrence are lacking, hindering optimal patient management.

Purpose of the Study:

• To develop and validate a deep learning (DL) model for predicting early versus late glioma recurrence.
• To assess the efficacy of a multimodal data approach incorporating clinical, mutation, and mRNA-expression data.

Main Methods:

• Developed gLioma recUrreNce Attention-based classifieR (LUNAR), a DL model with attention mechanisms.
• Utilized clinical, mutation, and mRNA-expression data from TCGA and GLASS datasets for training and validation.
• Compared LUNAR performance against traditional ML and non-attention DL models.

Main Results:

• LUNAR achieved an AUROC of 82.84% on the TCGA dataset and 82.54% on the GLASS dataset.
• The DL model with attention mechanisms outperformed traditional ML and non-attention DL models.
• Consistent performance across two independent datasets demonstrates model robustness.

Conclusions:

• Multimodal DL classifiers show significant potential for predicting early glioma recurrence.
• LUNAR enhances risk stratification by integrating diverse patient data.
• The model's robust performance suggests its utility in clinical decision-making for glioma patients.