MAML-residual transformer for few-shot prediction of targeted therapy response in NSCLC

Non-small cell lung cancer (NSCLC) exhibits high heterogeneity, and the scarcity of annotated imaging data limits the robustness and generalization capability of conventional deep learning approaches, particularly in few-shot scenarios. Therefore, developing accurate and generalizable prediction models under limited sample conditions remains a critical challenge for precision oncology.

In this study, we proposed a Model-Agnostic Meta-Learning_RT (MAML_RT) framework for predicting targeted therapy response in NSCLC. The framework integrates model-agnostic meta-learning with residual transformers. Multi-task meta-training was employed to learn highly generalizable parameter initializations, enabling rapid adaptation to new cohorts with limited labeled samples. Meanwhile, the multi-head self-attention mechanism of the residual transformer was utilized to model global correlations among radiomics features and capture long-range dependencies between diverse tumor phenotypic characteristics.

In a cohort of 300 patients, MAML_RT achieved an accuracy of 0.91 and an area under the receiver operating characteristic curve (AUC) of 0.93 for predicting targeted therapy efficacy. Under an extremely small-sample scenario (n = 50), the model maintained an AUC of 0.76, significantly outperforming the comparison models. Furthermore, on an independent external validation cohort, MAML_RT achieved an AUC of 0.85 and an accuracy of 0.88, demonstrating its cross-center generalization capability.

The proposed MAML_RT framework provides an effective solution for targeted therapy response prediction in NSCLC under limited labeled data conditions. By integrating meta-learning with residual transformer-based feature modeling, this approach improves model adaptability and generalization, offering potential support for clinical decision-making and individualized treatment stratification.