Longitudinal Risk Prediction for Pediatric Glioma with Temporal Deep Learning.

Area of Science:

• Artificial Intelligence in Oncology
• Medical Imaging Analysis
• Pediatric Neuro-oncology

Background:

• Pediatric glioma recurrence presents heterogeneous patterns, challenging prediction with current clinical and genomic markers.
• Frequent, long-term MRI surveillance is standard for pediatric glioma patients due to unpredictable recurrence.
• Limited data availability and existing machine learning methods have hindered progress in individualized recurrence prediction.

Purpose of the Study:

• To develop and validate a deep-learning approach for improved prediction of pediatric glioma recurrence using longitudinal MRI data.
• To enhance individualized risk assessment for pediatric glioma patients undergoing surveillance.
• To explore the adaptability of temporal learning for other cancers and chronic diseases.

Main Methods:

• Developed a self-supervised temporal deep-learning model for longitudinal medical imaging.
• Model encodes serial MRI scans, trained on chronological order classification (pretext task).
• Fine-tuned model for 1-year recurrence prediction in pediatric gliomas using historical surveillance scans from 715 patients (3994 scans).

Main Results:

• Temporal learning improved recurrence prediction performance (F1 score) by up to 58.5% compared to traditional methods.
• Performance gains were observed in both low- and high-grade pediatric gliomas, with AUC ranging from 75% to 89%.
• Prediction accuracy improved with more historical scans, plateauing between three and six scans.

Conclusions:

• Temporal deep learning offers high-performance longitudinal analysis for pediatric brain tumor surveillance and decision support.
• This approach shows potential for broad application in tracking and predicting risk for other cancers and chronic diseases.
• The AI model facilitates more precise, individualized patient management in neuro-oncology.