Abstract 1143: Reliable skeletal muscle area quantification and clinical data integration for predicting cancer cachexia

Background: Cancer cachexia, a metabolic disorder marked by severe muscle loss and common among certain types of cancers, significantly impacts prognosis and quality of life in patients. A core feature of cachexia is skeletal muscle mass reduction, which can be efficiently monitored by assessing the skeletal muscle area (SMA) in computed tomography (CT) scans routinely performed in cancer care. However, manual annotation of SMA on CT scans is labor-intensive and time-consuming, while existing automated tools often suffer from variability in accuracy and lack full automation, limiting their clinical utility. To address these challenges, we developed a reliable and fully automated pipeline to deliver consistent, accurate measurements. By integrating SMA and skeletal muscle index (SMI) with clinical data, our approach facilitates multimodal survival analysis and cachexia prediction, generating insights that empower clinicians to enhance patient care. Method: We trained two deep learning (DL) models (nnU-Net 2D) using annotated CT images of the mid and end-third lumbar (L3) level from 50 gastroesophageal and 15 pancreatic cancer patients. Models were trained using 5-fold cross-validation to ensure robustness. To ensure reliability at inference, uncertainty estimation methods, model ensembling, dropout, and post-hoc calibration, were employed. Metrics such as variance, entropy, and coefficient of variation quantified model uncertainty, and a threshold-based approach flagged high-error SMA estimations for expert review. Our pipeline processes axial CT series, identifies L3 slices, annotates skeletal muscle, generates uncertainty maps, and SMA/SMI estimates. We combined the radiology derived SMA/SMI with clinical features to form multimodal data, which was used for survival analysis and prediction of cachexia at the time of cancer diagnosis using a multi-layer perceptron (MLP) model. Results: On the gastroesophageal dataset, the DL model achieved an average Dice score of 97.80% ± 0.93% for skeletal muscle segmentation. Across the pancreatic, colorectal, and ovarian datasets, the DL model's SMA estimates differed with a median of 2.48% compared to manual expert segmentation. Uncertainty metrics demonstrated strong correlations with SMA estimation differences, with correlation coefficients of 0.83 (variance), 0.76 (entropy), and 0.73 (coefficient of variation). The multimodal MLP model for cachexia prediction achieved 70% accuracy and an F1 score of 76.92%. The multimodal survival analysis showed significant improvements in concordance index compared to using clinical data alone, with increases of 5.6% on the pancreatic, 5.9% on the colorectal, and 1.5% on the ovarian dataset. Conclusion: Our automated, uncertainty-aware tool offers a robust and reliable solution for monitoring skeletal muscle changes, enabling diagnosis and intervention in managing cancer cachexia. Citation Format: Sabeen Ahmed, Nathan Parker, Margaret Park, Daniel Jeong, Lauren Peres, Evan W. Davis, Jennifer B. Permuth, Erin Siegel, Matthew B. Schabath, Yasin Yilmaz, Ghulam Rasool. Reliable skeletal muscle area quantification and clinical data integration for predicting cancer cachexia [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular s); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1): nr 1143.