Abstract A047: Building clinically relevant and interpretable multi-modal machine learning algorithms to predict glioblastoma disease progression

Abstract Introduction: Machine learning (ML) algorithms have demonstrated high predictive performance in clinical diagnostics and decision making. However, these applications often suffer from low clinical adoption due to poor interpretability and unimodal workflows (e.g., trained only on imaging) that do not mirror real-world clinical reasoning. Predicting disease progression is essential for early management and improved outcomes in glioblastoma (GBM), a highly recurrent primary brain tumor with poor prognosis. We develop a novel multi-modal ML algorithm for GBM progression prediction using computer vision (CV) and natural language processing (NLP) algorithms to extract interpretable, clinically relevant features for further downstream outcome prediction. Methods: In a cohort of 92 pathology-proven GBM patients, we analyzed patient data including free text notes, MRI imaging, radiotherapy (RT) dosimetry, medications, and clinical variables. The target prediction was 24-month progression-free survival (PFS), with ground truth manually obtained via expert application of Response Assessment in Neuro-Oncology (RANO) clinical criteria (median 15.3 mo). Our custom local spaCy NLP model extracted RANO terms relating to disease progression and stability, while deepMedic tumor segmentation CV algorithms quantified tumor volume changes in brain MRI scans. A random forest model was trained on term frequencies, tumor volume changes, and scaled clinical data using an 80/20 train-test split with stratification to handle class imbalance. Results: Our multi-modal algorithm with clinically interpretable features yielded a 95% training accuracy and test performance metrics of 90% accuracy, 94% F1 score, and 91% AUROC. Feature importance obtained via mean decrease in impurity indicated that the top five features influencing model performance included max RT dose delivered to the optic chiasm and brain stem, frequency of stability-related radiology report terms, age at diagnosis, and 6-month change in contrast-enhancing tumor volume. These results highlight the potential utility of mirroring clinical progression criteria via intermediate NLP- and CV-derived features and using RT dosimetry parameters, which are underemployed in cancer ML applications. Conclusion: This study reports a novel multi-modal ML approach to GBM progression prediction using interpretable intermediate clinical features extracted from raw imaging, RT dosimetry, and free text data. Further user studies with clinicians and larger patient cohorts can validate and fine-tune this interpretable multi-modal approach to approximating real-world clinical reasoning and contrast results with current black box, unimodal algorithms. Citation Format: Shreya Chappidi, Erdal Tasci, Longze Zhang, Kevin Camphausen, Andra V. Krauze. Building clinically relevant and interpretable multi-modal machine learning algorithms to predict glioblastoma disease progression [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Artificial Intelligence and Machine Learning; 2025 Jul 10-12; Montreal, QC, Canada. Philadelphia (PA): AACR; Clin Cancer Res 2025;31(13_Suppl):Abstract nr A047.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.