Deep learning has rapidly emerged as a transformative technology in oncology, offering new capabilities in treatment response prediction and personalized cancer care. This systematic review and meta-analysis aim to evaluate the predictive performance, methodological quality, and clinical implementation of deep learning models for cancer treatment outcomes. A comprehensive search across ten databases and preprint servers identified 158 eligible studies, with 89 included in the quantitative synthesis. Results revealed pooled AUCs of 0.823 (internal validation) and 0.787 (external validation), with superior performance observed in multimodal and Transformer-based models.
However, given the substantial heterogeneity (I2 > 70 %) across included studies, these pooled estimates should be interpreted as broad indicators of methodological feasibility rather than definitive performance benchmarks. Methodological inconsistencies, high risk of bias, and limited external validation were common, and only 9 % of models had been implemented clinically.
This study contributes to the literature by providing the first cross-cancer meta-analytic synthesis of deep learning in treatment prediction across cancer types and model architectures. Findings highlight both the promise and the current limitations of AI integration in oncology and emphasize the need for rigorous validation, transparent reporting, and translational research. The review encompassed studies on both solid tumors (breast, lung, colorectal, prostate, and others) and various treatment modalities including chemotherapy, immunotherapy, radiation therapy, targeted therapy, and surgical interventions. Outcome measures included treatment response prediction (measured via AUC), overall survival and progression-free survival (evaluated using C-index and hazard ratios), and clinical utility (assessed through net benefit and decision curve analyses).
Inicia sesión o regístrate para acceder al texto completo
¡Aún no hay comentarios. Sé el primero en comentar!