A Clinician’s Guide to Machine Learning in Neuropsychological Research and Practice

Abstract

Machine learning (ML) techniques can help harness insights from data that complement and extend those that can be attained by traditional statistical methods. The current article introduces clinicians to concepts underlying ML and explores how it can be applied within the domain of neuropsychology. Specifically, we illustrate an application of ML to a dataset that includes a battery of standardized measures designed to provide diagnostic support for concussions, including standardized neurocognitive (CPT 3) and neurobehavioral (BESS, NIH 4 meter gait) measures, gait sensor data, and a CDC concussion symptom checklist. These variables were used to predict the decision-making of a pediatric neurologist evaluating a group of child/adolescent patients. With a sample of 111 cases, ML (using a general linear model and deep learning as illustrations) achieved accuracies of 91% and 84.8% and AUCs of 1.0 and .947, respectively, when predicting the neurologist’s binomial decision-making (safe/remove). In presenting the data and various considerations for interpretation, we attempt to balance both the promise and perils of ML.