Machine learning applications in predicting pharmacological treatment outcomes and responses in esophageal cancer

This study evaluated machine learning (ML) models predicting esophageal cancer (EC) treatment outcomes, focusing on data modalities, feature engineering, model frameworks, and validation.

Following PRISMA guidelines (PROSPERO: CRD42024619947), six databases (2015-2024) were systematically searched. Two reviewers independently extracted data on model methodologies and performance. Study quality was assessed using a modified TRIPOD (Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis) + AI checklist.

Among 30 studies (14,342 patients), classical ML models were the most frequently employed approach (n = 43), followed by ensemble methods (n = 34), with deep learning being the least utilized (n = 11); however, the best-performing models across all studies demonstrated mean AUC values of 0.847 for deep learning, 0.835 for ensemble models, and 0.816 for classical approaches. Imaging and clinical data constituted the predominant both unimodal and multimodal modeling inputs, with supervised learning representing the dominant paradigm. Multimodal models achieved a significantly higher AUC (0.84 vs. 0.78) than single-modal models. Model validation primarily relied on k-fold cross-validation and external cohort approaches. Quality assessment showed moderate reporting completeness (64.79% median fulfillment).

While ML (particularly deep learning and multimodal approaches) demonstrated potential for EC treatment prediction, key limitations persisted, such as opaque computational methods, poorly justified predictor selection, and unaddressed population heterogeneity/class imbalance. Addressing these challenges would be critical to enhancing the reliability and clinical applicability of ML models in future research.