5PSQ-153 Predictive factors of response to ustekinumab and vedolizumab: development of machine learning algorithms for clinical response prediction in patients with inflammatory bowel disease

Abstract

<h3>Background and Importance</h3> The optimisation of biological therapy in inflammatory bowel disease (IBD) remains a challenge in clinical practice. Identifying predictive factors of therapeutic response could guide individualised treatment decisions and improve long-term outcomes. Artificial intelligence tools, such as machine learning (ML), offer new opportunities to integrate complex clinical and biochemical data for precision medicine in IBD. <h3>Aim and Objectives</h3> To determine clinical response rates at week 26 and response and remission rates at weeks 52 and 104 in IBD patients treated with ustekinumab or vedolizumab, to identify predictive factors of response, and to develop ML-based predictive models of clinical outcomes. <h3>Material and Methods</h3> An observational, single-centre, retrospective study was conducted in IBD patients treated with ustekinumab or vedolizumab at Virgen Macarena University Hospital between January 2017 and December 2021. Logistic regression analyses were used to identify variables associated with clinical response. ML algorithms were developed using routinely available clinical and laboratory data to build predictive models of treatment response. Model performance was evaluated using the F1-score metric. <h3>Results</h3> A total of 228 patients were included (136 ustekinumab, 92 vedolizumab). Clinical remission rates at weeks 52 and 104 were 51.5% and 45.8% for ustekinumab, and 58.7% and 55.4% for vedolizumab, respectively. In Crohn’s disease, prior infliximab use, male sex, and longer disease duration were associated with greater likelihood of response to ustekinumab, while smoking predicted poorer response. For vedolizumab, left-sided ulcerative colitis, colonic Crohn’s disease, and male sex were associated with response, whereas previous exposure to infliximab or corticosteroids predicted poorer outcomes. The presence of extraintestinal manifestations negatively impacted response in both treatment groups. ML models integrating 26 clinical and analytical variables achieved robust predictive performance (F1=0.86). Nutritional parameters positively influenced response, while markers of inflammation showed a negative association. <h3>Conclusion and Relevance</h3> Ustekinumab and vedolizumab are effective therapeutic options for IBD. Specific clinical and biochemical features act as predictive factors of biological therapy response. ML-based models demonstrated good predictive accuracy and may support personalised treatment strategies in routine hospital pharmacy and gastroenterology practice. <h3>Conflict of Interest</h3> No conflict of interest

Ähnliche Arbeiten