Predicting Disease Severity in Patients With SCN8A Gain of Function Variants: A Proposal for a 3-Phase Predictive Model Pipeline From Diagnosis to Treatment

Abstract

The SCN8A Channelopathy is a family of rare pediatric neurological diseases, consisting of two genetically distinct populations; loss of function (LOF) and gain of function (GOF). These two populations have different pathologies and previous work has successfully employed machine learning techniques to differentiate patients between these groups. The GOF population is heterogenous in disease severity and profile, making prognosis in clinical settings challenging, and slowing progress in understanding the severity spectrum of this disease family. With the goal of increasing our understanding of the outcomes of patients within the GOF population, data were collected on 180 patients in the International SCN8A Patient Registry and features were selected for use in training and testing a series of predictive machine learning models in a statistically-informed unsupervised approach and clinically-informed supervised approach. In the unsupervised approach, a stacked predictive model using random forest classifiers and an ordinal logistic regression meta-learner performed with >95% accuracy. In the supervised approach, an ordinal logistic regression model performed with 83% accuracy. These models are proposed to be used together to better inform clinicians on whether a patient will have a Mild, Moderate, or Severe disease outcome, and act as the second phase of a three phase machine learning pipeline that will provide clinicians with a patient’s likelihood to be 1) LOF or GOF, 2) Mild, Moderate, or Severe in the case of GOF, and 3) the five best treatment options for that patient given previous predictions. This pipeline seeks to provide personalized medicine given a small amount of input data collected early in a patient’s life, referencing both genetic and phenotypic data.