6ER-004 Use of artificial intelligence to improve medical care in patients with inflammatory bowel disease. Inflammatory bowel disease. Preliminary models

<h3>Background and Importance</h3> Inflammatory bowel disease (IBD) is a chronic disorder that severely impact patients’ quality of life. Current treatments, which are often immunosuppressive and biologics, have significant side effects and are not always effective. The variability in response to these treatments makes it essential to personalise therapeutic management to reduce costs, minimise risks and improve outcomes. <h3>Aim and Objectives</h3> The aim of the project is to develop a model to predict if patients will respond to treatments, helping to make more informed clinical decisions and improving medium-term outcomes. <h3>Material and Methods</h3> To explore the development of a 26-week response prediction model, clinical, analytical, and treatment data from the 3 months prior to the initiation of infliximab/adalimumab/vedolizumab/ustekinumab were used in patients over 18 years of age with a confirmed diagnosis of IBD. Patients without availability/access to electronic health record data, with 50% of study variables not recorded or with a loss to follow-up prior to week26 were excluded. The proportion of patients intended for training and internal validation of the model was 90% and 10%, respectively. The selected centre was a tertiary-level hospital and the period analysed was from January 2012 to January 2024. <h3>Results</h3> A total of 1068 patients with a mean(±SD) age of 42.82(±21.08) years at the start of treatment and a mean (±SD) time of IBD evolution of 8.8(10.7) years. The proportion of male was 51.8%. Patients with CD accounted for 72.9%, and 24.25% had perianal disease. The drug response rate at week 26 was 82%.A neural network model was sketched. The proportion of missing values among the selected variables was 17%. A model with an F1 score of 0.802, an accuracy of 0.79 (rate of true positives vs. total true and false positives) and a recall of 0.811 (rate of true positives vs. total true and false positives). In the patients used to test the model, 66.40% classified as non-responders did not respond and 92.3% classified as responders did respond. <h3>Conclusion and Relevance</h3> The history, clinical and laboratory variables generated during the months prior to starting treatment could be used effectively to identify non-responders and optimise their therapy. <h3>References and/or Acknowledgements</h3> <h3>Conflict of Interest</h3> No conflict of interest