3128 - Clinical Utility of Spatial Mapping for High-Dimensional Genomic and Multi-Omic Data Classification in Precision Radiation Oncology

Purpose/Objective(s): As radiation oncology shifts toward personalized treatment based on multi-omic signatures, the lack of spatial structure in genomic and clinical tabular data limits the efficacy and interpretability of standard deep learning models. Drawing on medical physics principles, where raw signals are routinely transformed into structured physical representations (e.g., dose maps or sinograms) prior to inference, we evaluated a novel framework that reshapes abstract biomedical data into spatially organized images. The objective of this study was to determine if this spatial transformation provides superior diagnostic fidelity and resilience to the data constraints commonly found in clinical oncology.

Materials/Methods: We analyzed twenty-two classification datasets, including high-dimensional genetic cohorts critical for precision medicine. To ensure clinical relevance, we tested the models under two "real-world" stress tests: (1) Data Scarcity, by subsampling datasets (80% down to 20%) to simulate small institutional patient cohorts; and (2) Measurement Variability, by injecting 5–20% adaptive noise to simulate the artifacts and biological heterogeneity inherent in clinical biopsy and sequencing samples. We benchmarked the spatial mapping approach against a state-of-the-art transformer-based tabular foundation model.

Results: The spatial mapping framework achieved superior classification accuracy on high-dimensional genomic datasets, reaching up to 99.4%. In the presence of simulated measurement noise and biological variability, the spatial model demonstrated significantly higher stability than the transformer model. Furthermore, while the transformer model failed to process several full-scale genomic datasets due to memory limitations, the spatial mapping approach scaled efficiently, successfully handling large-scale clinical cohorts without computational failure.

Conclusion: Reframing high-dimensional genomic and clinical data as structured spatial fields provides a more reliable and scalable foundation for computational diagnostics in radiation oncology. This approach bridges the gap between tabular data and medical imaging, offering a robust tool for precision oncology that maintains high performance even in the presence of noise and data limitations typical of clinical practice.