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

Background: Gliomas account for approximately 25.5% of all primary brain and central nervous system (CNS) tumors and 80.8% of malignant brain and CNS tumors. The prognosis varies considerably; patients with low-grade gliomas (LGGs) have 5-year survival rates of up to 80%, while patients with higher-grade gliomas (HGGs) often experience rates below 5%. Recurrence is a common challenge, occurring in 52% to 62% of patients with LGGs and 90% of patients with HGGs, complicating clinical management and treatment planning. Currently, no widely available models exist for reliably predicting early glioma recurrence, which is critical for optimizing patient outcomes. Machine learning (ML) and deep learning (DL) techniques have shown promise in predicting recurrence for various cancers, with those utilizing multimodal data sources showing increasing promise.

Methods: We developed a DL-based predictive model with attention mechanisms, gLioma recUrreNce Attention-based classifieR (LUNAR), to predict early vs. late glioma recurrence using clinical, mutation, and mRNA-expression data from patients with primary grade II-IV gliomas from The Cancer Genome Atlas (TCGA) and, as an external validation set, the Glioma Longitudinal Analysis Consortium (GLASS).

Results: Our model outperforms traditional ML models and non-attention counterparts, achieving area under the receiver operating characteristic curve (AUROC) of 82.84% and 82.54% on the TCGA and GLASS datasets, respectively.

Conclusions: Our results demonstrate the potential of multimodal DL classifiers for predicting early glioma recurrence. By integrating clinical, mutational, and transcriptomic data from patients, LUNAR enables improved risk stratification. Its consistent performance across two independent datasets underscores its robustness.