JMIR Publications

ABSTRACT

Background:

Patients with anti-melanoma differentiation-related gene 5 antibody positive dermatomyositis-associated interstitial (MDA5+DM-ILD) were prone to rapidly progressive interstitial lung disease (RP-ILD) and high mortality. Therefore, a validated prediction model for mortality evaluation using a concise and practicable web-based tool needs to be established.

Objective:

To develop and validate a risk prediction model of 3-month mortality using machine learning in a large multi-center cohort of patients with anti-MDA5+DM-ILD patients in China.

Methods:

Consecutive 609 patients with anti-MDA5+DM-ILD were enrolled retrospectively from 6 hospitals in China. Patient demographics and laboratory and clinical parameters were collected on admission. The primary endpoints were the 3-month mortality for all causes. Six machine learning algorithms including extreme gradient boosting, logistic regression, Light Gradient Boosting Machine, random forest, support vector machine and k-nearest neighbor algorithm were applied to construct the model and evaluate the model using k-fold cross validation method, ROC curve, calibration curve, decision curve analysis (DCA) and external validation.

Results:

A total of 509 patients from 6 institutions in China were finally included in the training cohort (n =203), an internal validation cohort (n=51) and an external validation cohort 1 (n=92) and cohort 2 (n=163). respectively. By machine learning, eight important variables including RP-ILD, erythrocyte sedimentation rate, serum albumin, age, C-reactive protein level, aspartate aminotransferase, lactate dehydrogenase, ratio of the absolute neutrophil count and the absolute lymphocyte count, were screened out and identified as for model construction. Compared with different machine learning approaches, logistic regression finally was selected as optimal model construction algorithm. The prediction model constructed by logistic regression achieved the best performance with an AUC, sensitivity, and specificity of 0.866,84.8%, and 84.4% on the validation set and 0.90, 85.0%, and 83.9% on the test set, respectively. The calibration curve and decision curve analyses showed satisfactory consistency and clinical utility of this prediction model. Furthermore, the optimal logistic model achieved encouraging predictive performance in the external validation cohort 1 (AUC 0.836, 95% CI, 0.754–0.916), and external validation cohort 2 (AUC 0.915, 95% CI, 0.871–0.959), indicating that the prediction model had better extrapolation.

Conclusions:

A robust clinical model and a web tool for predicting the risk of 3-month mortality of anti-MDA5+DM-ILD patients was successfully developed.