Abstract A059: CP-Fuse: A Comprehensive Assessment of Clinico-Pathological Fusion in TCGA Survival Prediction
Bibliographic record
Abstract
Abstract Cancer remains a leading cause of global mortality. Multidisciplinary tumor boards play a central role in diagnosis and treatment planning but often face challenges related to growing case complexity, time constraints, and inconsistent expertise. Artificial intelligence models that incorporate diverse data sources such as histopathology and clinical records offer valuable decision-support tools, and multimodal approaches have been shown to outperform unimodal methods. This study introduces CP-Fuse, a novel multimodal framework that aims to improve progression-free survival (PFS) prediction in cancer patients by integrating digital pathology whole slide images (WSI) and clinical variables. Our approach has two main branches: the WSI branch (P-solo), using the Hierarchical Pyramid Transformer with DeepSurv, and the clinical branch (C-solo), using FTTransformer with DeepSurv. To combine the two modalities, we tested five fusion strategies: CP-fuse-marginal (simple concatenation), CP-fuse-crossattention (cross-attention between modalities), CP-fuse-trainableweights (late fusion with learnable scalar weights), CP-fuse-metalearning (meta-learner combining outputs), and CP-fuse-VAE (a variational autoencoder learning a shared representation). Performance is assessed on five TCGA cohorts (HNSC, BLCA, UCEC, LUAD, and BRCA), comprising a total of 2,867 patients. Rather than relying on a single train-validation split, we used 10-fold cross-validation repeated 100 times. This setup reduces the impact of random splits and provides a more reliable estimate of model performance. We report the average C-index across all runs. To obtain a robust estimate of model performance and mitigate the impact of individual fold splits, we conducted 100 runs of 10-fold cross-validation, each employing different fold configurations. Under these repeated evaluations, the clinical-only model (C-solo) consistently outperformed the WSI-only model (P-solo) across all cohorts, showing the largest C-index gap in HNSC (0.6859 vs. 0.5503) and the smallest in BRCA (0.7198 vs. 0.6063). Fusion models provided further gains in risk ranking and calibration, especially CP-fuse-VAE, which attained the highest C-index in four of five cohorts, exceeding C-solo by 0.0392 in LUAD (0.6771 vs. 0.6379) and by 0.0352 in BRCA (0.7550 vs. 0.7198). These findings underscore the potential of our multimodal approach. Notably, the CP-Fuse-VAE model outperformed the unimodal clinical model (C-solo), demonstrating that integrating WSI with clinical variables reveals predictive insights inaccessible to single modalities. Our results show that cross-modal interactions enhance relative risk ranking and capture essential prognostic signals. Furthermore, our use of an unprecedented 100-run cross-validation protocol to quantify fold-to-fold variability sets a new standard in methodological rigor, ensuring robust and reproducible. Prior survival studies have typically relied on a single split or standard k-fold validation without analyzing variability across folds, a gap our work addresses directly. Citation Format: Juan Felipe. Duran, Yujing Zou, Harry Glickman, Elliot Wadge, Khalil Sultanem, George Shenouda, Martin Vallières, Shirin Abbasinejad Enger. CP-Fuse: A Comprehensive Assessment of Clinico-Pathological Fusion in TCGA Survival Prediction [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 A059.
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.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.005 | 0.008 |
| Meta-epidemiology (narrow) | 0.002 | 0.000 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.002 | 0.001 |
| Science and technology studies | 0.001 | 0.000 |
| Scholarly communication | 0.001 | 0.001 |
| Open science | 0.001 | 0.002 |
| Research integrity | 0.002 | 0.001 |
| Insufficient payload (model declined to judge) | 0.002 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".