A Bayesian Model for Patient-Reported Age at FSHD Symptoms Onset

Abstract

Introduction

Facioscapulohumeral muscular dystrophy (FSHD) is one of the most common muscular dystrophies and is usually associated with reduction of the D4Z4 repeat array on the 4q subtelomere (DRA). Because FSHD is hereditary, risk prediction is central to genetic counselling. However, prediction of disease occurrence and age at onset remains challenging mainly because of the retrospective nature of symptom onset reporting. Age at first symptom onset is often patient-reported and may be affected by recall error, rounding or heaping at preferred ages. Treating such reports as exact event times may bias risk estimation.

Objective

We developed an accelerated failure-time survival model to estimate the probability of FSHD onset and the expected age at disease onset in newborns of DRA-carriers, using only information from parents and accounting for censoring, familial clustering and uncertainty in reported onset age.

Methods

This study used the genetic, clinical, and sociodemographic data from the Italian National Registry for FSHD. Clinical information included symptom presence and severity, and standardized phenotype descriptions obtained through the Comprehensive Clinical Evaluation Form. Eligible families included at least one child, one DRA-carrier parent, and, when available, additional relatives.

The model developed was a Bayesian hierarchical accelerated failure-time survival model with an observation model for patient-reported onset age. The outcome was the unobserved age at first patient-reported FSHD symptom onset and was modelled through a Weibull distribution. A 3 levels hierarchical structure was introduced to model the correlation structure among subjects sharing the same family and the same parents. Weakly informative priors were used for all parameters involved. For affected individuals, the referred age at onset was treated as an error-prone report of the latent event time: non-rounded integer ages were assigned narrow Student-t reporting error distributions, whereas ages ending with 5 or 0 were assigned progressively wider error distributions to account for potential age heaping. In addition, clinically meaningful developmental codes (such as age 6 used to denote childhood onset) were modelled as interval-censored observations. Unaffected genetically at-risk individuals were considered censored at their last assessment age.

Predictive performance measures were computed using 10-fold family-level cross-validation, holding out entire family clusters to evaluate generalization to unseen families.

Results

Overall, 355 children of FSHD-affected subjects from 205 families were analysed. Across the analysed age ranges, the model showed moderate to good discrimination. Family-level cross-validation time-dependent Area under the ROC curve (AUC) exhibited values ranging from approximately 0.70 to 0.80, with mean equal to 0.75. Overall, the model achieved a mean balanced accuracy equal to 0.69, but its performances were highly dependent on parents’ characteristics, resulting in better results when analysing classical FSHD phenotype parents, and poorer results when applied to complex phenotype cases.

Conclusions

We developed a Bayesian model for analysing a patient-reported time-to-event outcome in FSHD. It was designed to account for the uncertainty surrounding the age at onset, that is a clinically important information but usually collected with imprecision. This method may represent also a useful tool to support clinicians during genetic counselling.