Predicting age at loss of ambulation in Duchenne muscular dystrophy with deep phenotypic measures

• Published in: 2014 IEEE Global Conference on Signal and Information Processing (GlobalSIP) pp. 1344 - 1347
• Main Authors: Yinxue Wang, Bello, Luca, Yue Wang, McDonald, Craig M., Hoffman, Eric P., Guoqiang Yu
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.12.2014
• Subjects: Accuracy; Data models; Diseases; Drugs; Loss measurement; Numerical models; Predictive models
• DOI: 10.1109/GlobalSIP.2014.7032343

Although Duchenne muscular dystrophy (DMD), the most common single-gene lethal disorder, is caused by a homogeneous biochemical defect in all patients, substantial patient-patient variety in disease progression is observed. The loss of ambulation (LoA) is a functional milestone of DMD progression and the age at LoA is often used as an indication of disease severity. But as age at LoA is not always available, such as when patients remain ambulant at study end, its use has been limited. In this paper, we report machine learning approaches to predict age at LoA based on clinical measures of muscular strength and motor function, and validate the algorithms using the CINRG dataset. With extensive experiments and rigorous statistical analysis, we found that (1) the utilization of multiple clinical features yields better prediction than using any of the single measures, and (2) the prediction based on Lasso is more accurate than other multivariate analytical approaches such as ordinary least squares and ridge regression. To our knowledge, we are the first to provide point predictions for age at LoA in DMD using clinical phenotypic measures. Importantly, we find that not all clinical measures contribute to the prediction. Age at the last visit (before LoA), velocity of walking 10 meters, and velocity of climbing 4 steps are selected as important predictors by Lasso. The usefulness of the prediction model is illustrated with evidence that the association between a well-known modifier of DMD severity and age at LoA has better power when the predicted values are utilized.