Introduction
In Japan, the Ministry of Health, Labor, and Welfare (MHLW) registered all patients with ulcerative colitis (UC) from 2003 to 2014, and real-world data (RWD) are available in a national database. Deep learning models have been used to analyse RWD, but visualising the relationship between outcomes and explanatory variables is difficult.
Aims & Methods
Therefore, in this study, we used explainable artificial intelligence (AI) to calculate the regression coefficient for each parameter to determine the explanatory variables involved in steroid-resistant UC and constructed a model to predict steroid-resistant UC at the time of a patient's first visit. We obtained data on all patients with UC newly registered in the MHLW database between 2003 and 2011 (before the spread of biologics in Japan). Of the 79,096 patients identified, 4,003 who had three years of data, a Mayo score of 3 or higher at the first visit, and who had been using steroids from the time of the first visit were included in the analysis. For explainable AI, a point-wise linear model (PWL) was used to generate a custom-made logistic regression model for each case. Clustering analysis was performed based on the similarity of the coefficients (weights) of these logistic regression models. Finally, the association with outcomes was evaluated by analysing the weights of factors with a significant difference (p<0.01) between the clusters.
Results
The area under the curve (AUC) of the PWL model was 0.628 for the training set and 0.627 for the test set. The likelihood ratio was 1.36. By clustering, the analysed population was classified into two groups (S group, predicted to be in remission after three years: 2,040 cases; R group, predicted not to be in remission: 1,963 cases). In both groups, female sex (R group: median weight 0.0027), positive stool culture (R group: median weight 0.0041), and colonic mucosal abnormalities (R group: median weight 0.0113) had a strong relationship with steroid resistance. The R group tended to have significantly higher frequencies than the S group. In the R group, pseudopolyposis (median weight = 0.0351), caecal lesions (median weight = 0.024), CRP (median weight = 0.0163), ascending colon lesions (median weight = 0.0133), and transverse colon lesions (median weight = 0.0107) were observed. Conversely, the white blood cell count (S -0.0021 vs R -0.0037), frequency of defecation (S -0.0053 vs R -0.0055), bloody stool (S -0.0021 vs R -0.0016), stool form (S -0.0347 vs R -0.0367), and abdominal pain (S -0.0291 vs R -0.0383) were not significantly related to steroid resistance in both groups.
Conclusion
Explainable AI in a national database with three years of records allowed us to build a predictive model for refractory UC. Female sex, positive faecal bacterial cultures, colonic mucosal abnormalities, and lesions extending into the right colon were important predictors of steroid resistance. Understanding the predictors of steroid resistance and dependence at the time of steroid introduction, based on national data before the spread of biologics, can facilitate prompt and timely treatment plans for patients with UC.
References
![]()
Kanatani et al. Biology 11:951, 2022.
Disclosure
The authors, Hidekazu Suzuki, received the reserach grant from Biofermin Pharm. Co., and the lecture fee from Astellas Co., Otsuka Pharm. Co., EA Pharma Co., and Takeda Pharm. Inc.
