Machine learning for differential diagnosis of white matter lesions in Fabry Disease patients based on gait and cardiac characteristics

Abstract

Brain manifestations in FD include progressive white matter lesions (WMLs).This research aims to identify a set of gait and cardiac characteristics to discriminate FD patients with WMLs from FD patients without WMLs. For the gait study, 76 subjects walked through a predefined circuit using wearable sensors that continuously acquired different stride features. All strides of 16 gait variables were normalized using multiple regression models. The mean and the variability of each gait time series were calculated, resulting in 32 gait measures. Using the 32 gait measures, a feature selection algorithm were applied. Then, five different classifiers (LR, SVM Linear and RBF kernel, RF, and KNN) based on different selected set features were evaluated. CNN and LSTM algorithms were implemented using as input the gait time series. For FD patients with WMLs vs controls the highest accuracy of 71.50% was obtained using RF. For FD patients without WMLs vs controls, the best performance was observed using KNN with an accuracy of 86.67%. For FD patients with vs without WMLs the best models were obtained using the CNN algorithm and using LR algorithm with an accuracy of 81.43% and 80.76%, respectively. Regarding the cardiac data, was used the data from two exams: an electrocardiogram (ECG) and an echocardiogram. A total of 114 FD patients were evaluated with the ECG (61 have WMLs). For FD patients with vs without WMLs the highest accuracy of 79.72%. With the simultaneously use of both gait and ECG features, two models were evaluated on a test group of nine patients. The best result was 80% accuracy with LR. Finally, logistic regression analyses were also performed on the 22 features of the echocardiogram of ninety three FD patients (forty-nine with WMLs). The results confirmed that age is significantly associated with the presence of WMLs. These findings are the first step to demonstrate the potential of machine learning techniques based on gait and cardiac characteristics as a complementary tool to understand the role of WMLs in FD patients.

As manifestações cerebrais da doença de Fabry (FD) incluem lesões na matéria branca (WMLs). O objetivo deste estudo é identificar quais as caracteristicas da marcha e cardíacas que permitem diferenciar os pacientes com FD com WMLS de pacientes com FD sem WMLs. Para o estudo da marcha, foram avaliados 76 sujeitos através de sensores vestíveis. Os valores da série temporal das 16 variáveis da marcha foram normalizados usando modelos de regressão múltipla. Usando as 32 medidas de marcha (média e variabilidade), foi aplicado um algoritmo de feature selection seguido de cinco classificadores diferentes (LR, SVM Linear e kernel RBF, RF e KNN). Os algoritmos CNN e LSTM foram implementados utilizando como input os conjuntos de séries temporais da marcha. Para pacientes com FD e com WMLs vs controlos, a maior exatidão de 71,50% foi obtida usando RF. Para pacientes com FD e sem WMLs vs controlos, o melhor desempenho foi observado usando KNN com uma exatidão de 86,67%. Para pacientes com FD com vs sem WMLs, os melhores modelos foram obtidos usando o algoritmo CNN e usando LR com base com uma exatidão de 81,43% e 80,76%, respetivamente. Em relação aos dados cardíacos, foram utilizados os dados de dois exames: o eletrocardiograma (ECG) e o ecocardiograma. Um total de 114 pacientes com FD (61 deste com WMLs) foram avaliados com o exame de ECG. Para pacientes com FD com vs sem WMLs, a maior exatidão foi de 79,72%. Com o uso simultâneo de marcha e ECG, dois modelos foram avaliados com um grupo de teste de nove pacientes. O melhor resultado foi a exatidão de 80% com o algoritmo LR. Finalmente, análises de regressão logística também foram realizadas nas 22 características do ecocardiograma de 93 pacientes com FD (49 com WMLs). Os resultados confirmaram que a idade está significativamente associada à presença de WMLs. Esses resultados demonstram o potencial das técnicas de machine learning baseadas na marcha e nas características cardíacas para entender o papel dos WMLs em pacientes com FD.