Advertisement

PGI27 Applying Machine Learning to Large Databases to Predict Nonresponse to Conventional Treatment in Patients with Ulcerative Colitis

      Objectives

      Ulcerative colitis (UC) treatment patterns can vary across patients and clinicians without formal clinical guidance. This research aims to assess the UC patients at the risk of not responding to conventional therapies (CT).

      Methods

      Patients with at least two distinct UC diagnoses and continuous health coverage for one year prior and one year after CT initiation, were identified from Optum data (2010-2019). Patients less than 18 years of age or with autoimmune diseases were excluded. The CT included 5-aminosalicylates, corticosteroids, and immunomodulators. The non-response to CT was defined as either cessation of CT for any reason or initiation of biologic therapy or prolonged corticosteroid use or UC related inpatient/emergency visit or UC related surgery during follow up. Over 53,010 baseline features including patient clinical care, specialist visits, diagnoses, medications, procedures, lab orders, and patient demographics for each patient were analyzed. A machine learning approach, regularized logistic regression, was trained to predict patient non-response to CT. The model was trained on data in the one year before the first CT. Cross-validation was performed across patients to train, validate hyperparameters, and compute held-out performance.

      Results

      In total, 9,752 UC patients (median age 56.8 years, 46% female) who initiated CT were included, among them 7.53% were non-responders. Treatment nonresponse was predicted with a held-out area under the receiver operator curve of 0.635. Features most predictive of patient nonresponse to CT included conditions such as hematochezia, diarrhea, and drugs including atropine/diphenoxylate, zolpidem tartrate and montelukast, and stool culture test among others.

      Conclusions

      Treatment nonresponse could be predicted with moderate performance from insurance claims. The results are encouraging for early detection of patients who may not respond to CT in UC. Future research including advanced methods like neural network or random forest may better identify nonresponse to CT and guide future clinical intervention.