AUTHOR=Wallert John , Westman Eric , Ulinder Johnny , Annerstedt Mathilde , Terzis Beata , Ekman Urban 

TITLE=Differentiating Patients at the Memory Clinic With Simple Reaction Time Variables: A Predictive Modeling Approach Using Support Vector Machines and Bayesian Optimization

JOURNAL=Frontiers in Aging Neuroscience

VOLUME=10

YEAR=2018

URL=https://www.frontiersin.org/journals/aging-neuroscience/articles/10.3389/fnagi.2018.00144

DOI=10.3389/fnagi.2018.00144

ISSN=1663-4365

ABSTRACT=<p><bold>Background:</bold> Mild Cognitive Impairment (MCI) and dementia differ in important ways yet share a future of increased prevalence. Separating these conditions from each other, and from Subjective Cognitive Impairment (SCI), is important for clinical prognoses and treatment, socio-legal interventions, and family adjustments. With costly clinical investigations and an aging population comes a need for more cost-efficient differential diagnostics.</p><p><bold>Methods:</bold> Using supervised machine learning, we investigated nine variables extracted from simple reaction time (SRT) data with respect to their single and conjoined ability to discriminate both MCI/dementia, and SCI/MCI/dementia, compared to—and together with—established psychometric tests. One-hundred-twenty elderly patients (age range = 65–95 years) were recruited when referred to full neuropsychological assessment at a specialized memory clinic in urban Sweden. A freely available SRT task served as index test and was administered and scored objectively by a computer before diagnosis of SCI (<italic>n</italic> = 17), MCI (<italic>n</italic> = 53), or dementia (<italic>n</italic> = 50). As reference standard, diagnosis was decided through the multidisciplinary memory clinic investigation. Bonferroni-Holm corrected <italic>P</italic>-values for constructed models against the null model are provided.</p><p><bold>Results:</bold> Algorithmic feature selection for the two final multivariable models was performed through recursive feature elimination with 3 × 10-fold cross-validation resampling. For both models, this procedure selected seven predictors of which five were SRT variables. When used as input for a soft-margin, radial-basis support vector machine model tuned via Bayesian optimization, the leave-one-out cross-validated accuracy of the final model for MCI/dementia classification was good (Accuracy = 0.806 [0.716, INS [0].877], <italic>P</italic> &lt; 0.001) and the final model for SCI/MCI/dementia classification held some merit (Accuracy = 0.650 [0.558, 0.735], <italic>P</italic> &lt; 0.001). These two models are implemented in a freely available application for research and educatory use.</p><p><bold>Conclusions:</bold> Simple reaction time variables hold some potential in conjunction with established psychometric tests for differentiating MCI/dementia, and SCI/MCI/dementia in these difficult-to-differentiate memory clinic patients. While external validation is needed, their implementation within diagnostic support systems is promising.</p>