P299 Prediction model for steroid-free clinical remission with vedolizumab at week 22 in patients with Ulcerative Colitis: Machine learning using clinical data at baseline
Miyoshi, J.(1);Maeda, T.(2);Matsuoka, K.(3);Saito, D.(1);Morikubo, H.(1);Matsuura, M.(1);Tamura, S.(2);Hisamatsu, T.(1);
(1)Kyorin University School of Medicine, Department of Gastroenterology and Hepatology, Mitaka-shi, Japan;(2)Gifu University, Department of Electrical- Electronic & Computer Engineering- Faculty of Engineering, Gifu-shi, Japan;(3)Toho University Sakura Medical Center, Division of Gastroenterology and Hepatology- Department of Internal Medicine, Sakura-shi, Japan;
Background
Predicting the response of patients with ulcerative colitis (UC) to a molecular targeted drug before administration is an unmet need for optimizing individual patient care. We hypothesized that a new machine-learning approach could identify predictors of therapeutic efficacy that have not been reported as statistically significant factors using the conventional statistical approaches. As a proof-of-concept study, we explored predictors and a prediction model for the efficacy of vedolizumab (VDZ) for UC.
Methods
The clinical data at baseline (week 0) of 34 UC patients who started VDZ at Kyorin University Hospital was collected (Cohort 1; training cohort). The collected data contained 49 clinical features; patient background (e.g. sex, age, disease phenotype, etc.), treatment history, clinical/endoscopic activity, and blood examination items. Random forest (RF) was employed to investigate clinical features that affect steroid-free clinical remission (SFCR) with VDZ at week 22 in UC. A prediction model was developed using logistic regression based on the results of RF. The prediction ability of the prediction model was evaluated with the data of 35 patients who started VDZ at Toho University Sakura Medical Center (Cohort 2; extra-facility test cohort). The prediction ability of the VDZ prediction model for ustekinumab (UST) was examined using the data of 22 patients who started UST at Kyorin University Hospital (Cohort 3).
Results
RF computed the contribution of each clinical feature to achievement/no-achievement of SFCR at week 22. When the top 8 features (partial Mayo score, MCH, BMI, BUN, concomitant use of AZA, lymphocyte fraction, height, and CRP) were employed for logistic regression, the prediction accuracy of the prediction model was 100% and 68.6% for Cohorts 1 and 2, respectively. The sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) in Cohort 2 were 92.3%, 54.5%, 54.5%, and 92.3%, respectively. Meanwhile, in Cohort 3 with UST, sensitivity, specificity, PPV and NPV were 70.0%, 45.5%, 53.4%, and 62.5%, respectively.
Conclusion
Our machine-learning approach detected clinical features at baseline that impact the achievement of SFCR at week 22 in patients who received VDZ for UC and developed a prediction model that appears to be useful to screen non-responders to VDZ. Given the low prediction ability in Cohort 3, our findings support the notion that a prediction model for each molecular targeted drug is needed for optimizing the drug selection for individual patients. This study provides the proof-of-concept that machine learning on clinical data could open a new era of personalized medicine for UC.