Artificial Intelligence–Driven Prediction of Anti-TNF Therapy Response in Ulcerative Colitis

Background: Anti-TNF therapies are frequently used in moderate to severe ulcerative colitis (UC); however, 30-40% of patients become medically refractory. Integrating artificial intelligence (AI) into a multimodal approach can improve the prediction of response to biologic therapy.

Design: A retrospective study of 482 patients with moderate to severe UC who received treatment with either infliximab or adalimumab from 2021 to 2025. Baseline data encompassed patient demographic details, disease duration, history of biologic therapy, CRP levels, fecal calprotectin levels, Mayo endoscopic score, and whole-scan image (WSI; 40x) of pretreatment biopsies. WSI images showed neutrophil infiltration, basal plasmacytosis, crypt distortion, and epithelial injury. We used gradient boosting with multimodal late fusion to predict the response to combined therapy after week 14. The data split ratio was 70% for the training set and 30% for the test set.

Results: Among 482 patients (mean age 38.6 years; 47.1% female; 29.3% previous biologic treatment), a response rate of 56% (n=270) was achieved. The multimodal model AUROC was 0.84 (0.80-0.88), whereas the clinical model AUROC was 0.74 (95% CI 0.69-0.79). The multimodal model demonstrated a Brier score of 0.17, positive predictive value of 81.6%, negative predictive value of 73.5%, sensitivity of 78.2%, and specificity of 77.4%. The calibration slope was 0.96, and the AUROC was 0.81 in internal validation. Endoscopic severity (OR 3.12, p<0.001), neutrophil predominant crypt damage (OR 2.48, p<0.001), basal plasma cell count (OR 2.14, p=0.002), and fecal calprotectin>1500 (OR 0.46, p=004) were indicators of response.

Conclusion: Our AI-driven multimodal model shows significant improvement for predicting early anti-TNF response in ulcerative colitis, with an internal validation AUROC of 0.81 and a threefold response gradient (31% vs. 82%). A more comprehensive, external, and prospective investigation to validate our results are necessary

Learning Objectives:

1. Describe limitations of clinical prediction of anti-TNF response in ulcerative colitis.
2. Explain how multimodal AI models integrate clinical and histologic data to predict outcomes.
3. Discuss how AI-derived response probabilities could guide biologic selection and trial enrichment.