The predictive value of machine learning for the response to anti-TNF-α therapy in patients with Crohn’s disease: a systematic review and meta-analysis

Abstract

Background This meta-analysis aimed to evaluate the performance of machine learning (ML) in predicting the response to anti-tumor necrosis factor-α (anti-TNF-α) therapy in individuals with Crohn’s disease (CD), providing an evidence-based basis for the construction or update of prediction tools. Methods PubMed, Cochrane Library, Embase, and Web of Science were comprehensively searched for studies published up to November 18, 2025 on ML predictive models for the therapeutic response to anti-TNF treatment in individuals with CD. The Prediction model Risk Of Bias Assessment Tool (PROBAST) was used to assess the risk of bias in the eligible articles. Model performance was assessed using the concordance index (C-index), and a meta-analysis was conducted. Results A total of 29 studies involving 6,697 patients with CD were included. Of these studies, 22 focused on the response to anti–TNF-α therapy, 6 on mucosal healing, and 1 on other outcomes. The meta-analysis demonstrated that for predicting the treatment response to anti–TNF-α treatment, in the validation dataset, the C-indices were 0.762 for ML models based on clinical features (CFs), 0.852 for models based on radiomic features, and 0.854 for models based on CFs combined with radiomic features. In internal validation, the C-indices for models based on CFs, radiomics, and CFs combined with radiomics were 0.773 (95% CI: 0.711–0.841), 0.840 (95% CI: 0.781–0.904), and 0.830 (95% CI: 0.780–0.883), respectively. In the external validation set, models based on radiomics (0.865, 95% CI: 0.783–0.955) and models based on CFs combined with radiomics (0.883, 95% CI: 0.816–0.955) outperformed models based on CFs (0.733, 95% CI: 0.645–0.833). Furthermore, the C-index of models based on endoscopic mucosal response was 0.794. Conclusions hile ML shows promise in predicting the therapeutic effects of anti-TNF-α treatment in individuals with CD, there are some methodological limitations. Future research should integrate multi-center, geographically diverse datasets to enhance the generalizability of models. Improving model robustness will help identify non-responders early, ultimately promoting the development of personalized therapeutic strategies and precision medicine. Systematic review registration https://www.crd.york.ac.uk/PROSPERO/, identifier CRD420251276470.