engleski

How useful is machine learning in predicting childhood IgA vasculitis relapses?

Introduction: IgA vasculitis (IgAV) is the most common systemic vasculitis in children. The mandatory clinical feature of the disease is purpuric rash, which predominantly affects the lower extremities, accompanied by diffuse abdominal pain, joint involvement, nephritis and/or IgA deposition in biopsy specimen (skin, intestinal tract, kidney). In most cases, IgAV is a self-limiting disease with favorable outcomes, however, relapses are not uncommon. Objectives: The aim of this study is to evaluate the usefulness of supervised machine learning (ML) algorithms in the identification of patients which could develop IgAV relapses. Methods: A large set of predictive variables related to demographic variables, clinical history, symptoms, laboratory values, and medications were used in the initial data collection for developing a predictive ML model. After preparing the dataset, handling missing values and data imbalances, a random forest (RF) decision tree and support vector machine (SVM) model with polynomial kernel were trained, crossvalidated and tested. Results: This pilot study included 539 IgAV patients (260 males, and 279 females), with a median age of 6.17 (4.42 to 8.75) years. Among them, 78.11% had joint involvement, 44.53% had gastrointestinal involvement, and in 18.92% nephritis has developed. Atypically distributed rash (not affecting lower limbs as well as rash that had been generalized from the onset of the disease) was found in 5.19% IgAV patients, while 8.91% had persistent purpura for a month or more. The incidence of IgAV relapses was 10.2%. The RF model produced an overall accuracy of 95%, with a sensitivity of 100%, and specificity of 94.48%. The SVM model produced an accuracy of 87.58%, with a sensitivity of 100%, and specificity of 86.21%. The most useful predictor variable was the presence of persistence rash in both models. Other common predictors in both models included the presence of rash on atypical locations, age, and nephritis. Conclusion: The results of this pilot study show promising applications of ML as an useful aid in predicting vulnerable patients who developed IgAV. Support: Croatian Science Foundation project IP- 2019-04-8822.

IgA vasculitis ; relapse ; learning machine

nije evidentirano