New Model for Estimation of the Age at Onset in Spinocerebellar Ataxia Type 3

Abstract

ObjectivesThe aim of this study was to develop an appropriate parametric survival model to predict patient's age at onset (AAO) for spinocerebellar ataxia type 3/Machado-Joseph disease (SCA3/MJD) populations from mainland China.MethodsWe compared the efficiency and performance of 6 parametric survival analysis methods (exponential, weibull, log-gaussian, gaussian, log-logistic, and logistic) based on cytosine-adenine-guanine (CAG) repeat length at ATXN3 to predict the probability of AAO in the largest cohort of patients with SCA3/MJD. A set of evaluation criteria, including −2 log-likelihood statistic, Akaike information criterion (AIC), bayesian information criterion (BIC), Nagelkerke R-squared (Nagelkerke R^2), and Cox-Snell residual plot, were used to identify the best model.ResultsAmong these 6 parametric survival models, the logistic model had the lowest −2 log-likelihood (6,560.12), AIC (6,566.12), and BIC (6,566.14) and the highest value of Nagelkerke R^2 (0.54), with the closest graph to the bisector Cox-Snell residual graph. Therefore, the logistic survival model was the best fit to the studied data. Using the optimal logistic survival model, we indicated the age-specific probability distribution of AAO according to the CAG repeat size and current age.ConclusionsWe first demonstrated that the logistic survival model provided the best fit for AAO prediction in patients with SCA3/MJD from mainland China. This optimal model can be valuable in clinical and research. However, the rigorous clinical testing and practice of other independent cohorts are needed for its clinical application. A unified model across multiethnic cohorts is worth further exploration by identifying regional differences and significant modifiers in AAO determination.

Clinically Meaningful Magnetic Resonance Endpoints Sensitive to Preataxic Spinocerebellar Ataxia Types 1 and 3

“…We further observed moderate‐to‐strong associations of cerebellar and pons MR metrics with estimated ataxia duration. Existing statistical models to estimate time to ataxia onset in preataxic individuals utilize CAG repeat length and current age 17,18 . Although they perform well at the group level, these estimates carry a large uncertainty in estimating onset age of individual patients.…”

“…Existing statistical models to estimate time to ataxia onset in preataxic individuals utilize CAG repeat length and current age. 17,18 Although they perform well at the group level, these estimates carry a large uncertainty in estimating onset age of individual patients. Noninvasive imaging metrics can enrich these parametric models and are expected to reduce the uncertainty of these estimates, which will be critical in subject stratification in trials.…”

Explainable artificial intelligence based on feature optimization for age at onset prediction of spinocerebellar ataxia type 3

“…To our knowledge, this is the first study attempting to apply XAI to the AAO prediction of SCA3/MJD. This study proposed an XAI model based on feature optimization, which achieved a better AAO prediction accuracy than previous studies (Peng et al, 2021a , b ) and can explain the impact of features and provide a personalized prediction.…”

“…The previous piecewise XGBoost model for AAO prediction of SCA3/MJD patients achieves metrics of MAE (5.56 and 4.78), RMSE (7.13 and 6.31), MedianAE (4.15 and 3.59), and proportion < 5 (55% and 65%) for CAGexp ≤ 68 and CAGexp > 68, respectively (Peng et al, 2021a ). Another survival analysis study proposed a parametric survival analysis method to predict the AAO with a reported R 2 of 0.54 (Peng et al, 2021b ).…”

“…Models including the maximum likelihood estimation model (França Jr et al, 2012 ), least-squares linear regression (Collin et al, 1993 ; Aylward et al, 1996 ; Peng et al, 2014 ; Bettencourt et al, 2016 ), linear regression based on the log-transformed AAO (Andrew et al, 1993 ; Lucotte et al, 1995 ; Chattopadhyay et al, 2003 ), piecewise regression (Andresen et al, 2007 ), quadratic regression (Tezenas Du Montcel et al, 2014 ; Chen et al, 2016a ), survival models (Brinkman et al, 1997 ; Langbehn et al, 2004 , 2010 ; Almaguer-Mederos et al, 2010 ; Du Montcel et al, 2014 ; De Mattos et al, 2019 ; Peng et al, 2021b ), and machine learning (ML) models (Peng et al, 2021a ) have been used for AAO fitting. Most statistical models attempt to investigate the relationship between the AAO and a few modifiers, but only a few studies have focused on AAO prediction, and the prediction accuracy still needs to be improved.…”