Abstract B027: Multimodal prediction of pathological complete response in esophageal cancer using automated machine learning and variational autoencoder-based synthetic data augmentation

Abstract Background: Pathological complete response (pCR) after neoadjuvant chemoradiotherapy in esophageal cancer is a clinically meaningful endpoint associated with improved survival. However, predicting pCR remains difficult due to small sample sizes and the high dimensionality of imaging and treatment-related data. This study presents a multimodal framework that integrates variational autoencoder (VAE)-based data augmentation with automated machine learning (AutoML) to improve the robustness and interpretability of pCR prediction. Methods: We retrospectively analyzed 89 esophageal cancer patients (44 pCR-positive, 45 pCR-negative) treated with curative-intent chemoradiotherapy between 2021 and 2024. Features were extracted from three sources: (1) clinical variables (e.g., age, gender, nodal stage, tumor location, treatment protocol), (2) radiomic and dosiomic features from planning CT and dose distributions within planning target volumes (PTVs), and (3) dose-volume parameters from tumor and elective field contours. All features were preprocessed using standard scaling and one-hot encoding.To address data scarcity and class imbalance, class-conditional VAEs were trained separately on pCR-positive and pCR-negative groups to generate 20 synthetic patients per class, expanding the dataset to 129 samples. We assessed distributional similarity between real and synthetic data using Uniform Manifold Approximation and Projection (UMAP) for visualization and Kolmogorov–Smirnov (KS) tests for statistical comparison. The Tree-based Pipeline Optimization Tool (TPOT), a genetic programming–based AutoML framework, was used for model selection and hyperparameter tuning. Feature importance was evaluated using SHAP (SHapley Additive exPlanations) values and LinearSVC coefficients. Results: On real data, TPOT identified XGBoost as the best-performing model (AUC = 0.61). After VAE-based augmentation, the optimal model changed to LinearSVC, which achieved the same AUC (0.61) using only five features, improving model interpretability. UMAP analysis showed partial embedding overlap between real and synthetic data, supporting the realism of the generated samples. KS analysis demonstrated significantly higher distributional similarity for radiomic and dosiomic features (mean KS ≈ 0.31, p < 0.05) compared to clinical and dose-volume variables (mean KS = 0.62, p < 0.001), indicating greater reproducibility of imaging-based biomarkers. The radiomic feature LowGrayLevelZoneEmphasis remained significantly associated with pCR in the real cohort ( p = 0.0072), suggesting translational relevance. Conclusion: This study highlights a novel approach combining synthetic data generation and AutoML for pCR prediction in esophageal cancer. The findings support the utility of radiomic and dosiomic features and demonstrate that VAE-augmented modeling may provide a scalable solution for predictive modeling in data-limited oncology settings. Citation Format: Şefika Dinçer, Taylan Tuğrul, Mahmut Kara, Mehmet Naci. Aldemir. Multimodal prediction of pathological complete response in esophageal cancer using automated machine learning and variational autoencoder-based synthetic data augmentation [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 B027.

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.