Machine learning model of the ultrasound index of masei enthesis and other variables of disease activity in patients with spondyloarthritis

Abstract

.Background: Spondyloarthritis (SpA) are a group of chronic inflammatory diseases with affectation, mainly of the axial skeleton, and also of peripheral joints. The enthesis is one of the target organs, and its inflammation known as enthesitis could be unnoticed. Machine Learning is a branch of artificial intelligence that studies the construction of a function y=f(x), from a finite set of observations D={x,y}, where y is an endogenous variable and x are explanatory variables. The objective of this method is to obtain a model that best fit to the data without overfitting, that could be useful to make predictions. Objectives: We try to find Machine Learning models that relate the MASEI index (Madrid Sonographic Enthesitis Index) in entheses depending on the activity of the disease (ASDAS, BASDAI and DAPSA) and other variables in patients with spondyloarthritis. Methods: Observational, descriptive and cross-sectional study. We have analyzed 24 patients with SpA who underwent musculoskeletal ultrasound using the MASEI index and who were treated in our clinics from May 2021 to September 2021 and under the approval of the CEICm of our center. First, we have done a feature selection of the variables most related to MASEI. To do so, we compute mutual information and chi-square test, using the scikit-learn (python) library and Matlab, respectively. Using Matlab Regression Learner package, we obtain the best Machine Learning model with the lowest RMSE for 5 fold cross-validation, which turned out to be a linear regression. Results: To obtain regressive models that explain TOTAL MASEI, the following variables have been chosen: Type of SpA, BASDAI-DAPSA-ASDAS activity, Arthritis, Enthesophytes, Corticosteroids and CRP because they present a high degree of mutual information with MASEI and a high chi-square index. (See Figures 1 and 2). With these variables we have obtained the model that presents a lower RSME error for validated data, which has turned out to be a linear regression, given by the formula: MASEI = -9.29-2,29* type of SpA + 10,42*ASDAS+4,08*Corticoids + 8,2*Arthritis/sinov. + 4,46*Enthesithis-8,6*CRP The basic statistical characteristics of the coefficients of this equation can be consulted in Figure 3. The characteristics of the model are specified in Figure 3. In Figure 4, we observe how the data fit the diagonal and in Figures 5 and 6, we have compute a prediction and confidence intervals of our model using ASDAS and MASEI as coordinate axes. Conclusion: We have obtained a linear model, which explains the MASEI variable as an explicit linear combination of the variables: type of SpA, ASDAS, Corticoids, Arthritis/sinov, Enthesithis and CRP. Our model, not only is simple, but it is also optimal, in the sense that for 5-fold cross validation, it obtains the lowest RSME error, compared to other Machine Learning methods, such as: neural networks, SVM, Gaussian regression processes, etc. Our model is useful to build confidence intervals, make predictions and to understand the relation between the variables mentioned above.