Bayesian Network Approach to Modeling Spinal Muscular Atrophy Populations
Jason H. Williams (1), Bhuvaneswari Jayaraman (1), Richard S. Finkel (2), Jeffrey S. Barrett (1)
(1) Laboratory for Applied PK/PD, Division of Clinical Pharmacology and Therapeutics, The Joseph Stokes Jr. Research Institute, Children’s Hospital of Philadelphia, Philadelphia, Pennsylvania, USA; (2) Division of Neurology, Children’s Hospital of Philadelphia, Philadelphia, Pennsylvania, USA
Background & Objectives: Proximal recessive Spinal Muscular Atrophy (SMA) is the most frequent heritable lethal disease in infants and a debilitating disease afflicting both pediatric and adult populations. Wide ranging severities are associated with a variety of distinct symptoms thus complicating enrollment and conducting of clinical trials. Furthermore, factors such as the presence of genetic modifier genes, intercurrent complications and surgical interventions result in markedly different disease course trajectories among patients and are not included in general prognostic indicators such as age-of-onset and highest motor-milestone achieved. Currently, available observation data is sparse and not optimal for modeling this complex disease. Standardization of clinical care [1] and progressing observational studies [2] aimed at describing the natural history of disease progression will likely result in richer clinical data available for future modeling efforts. The overall objective of this project was to develop a model for quantifying disease progression in populations with SMA which can be further refined once longitudinal studies have been completed.
Methods: A Bayesian Network was constructed based on expert knowledge and data from observational studies described in the SMA literature and consisted of a hierarchical nodal structure with directional links to define key dependencies between model parameters. Case-learning of parameter conditional probability tables was performed using the expectation-maximization algorithm implemented in Netica [3]. We explored several models using clinical questionnaire data (N=1700) from the SMA patient registry, extracted from medical records database and patient charts (N=55) within The Children's Hospital of Philadelphia network. Sensitivity analysis was performed across all model parameters and model performance was determined using receiver operating characteristic curves.
Results & Conclusions: Bayesian Networks allowed for calculation of posterior probability densities simultaneously across all model parameters and to combine empirical data from various sources with expert knowledge into the same modeling structure. Overall, the Bayesian Network approach seemed to be appropriate for the type of sparse longitudinal data that may arise from clinical observational studies and can be easily implemented in Netica for visual interpretation by clinicians involved in disease progression modeling efforts.