The aim of the present study was to develop competing mathematical models of within-person response time patterns before and after errors within a cognitive paradigm (flankers task) and evaluate their utility for schizophrenia classification (N=524). A secondary aim was to apply these models to EEG time series, focusing on areas in which event-related potentials (ERPs) connected to unconscious and conscious error monitoring and response (i.e., frontocentral and midline regions) are recorded. Models were chosen based on alternative mathematical constructions intended to capture the shape, magnitude, and timing of error response and monitoring. Candidates included linear, exponential decay, and damped oscillator models with post-error-varying parameters or shock components.
Person-average neural (ERPs) and behavioral (post-error slowing, accuracy, mean RT) measures failed to reach acceptable schizophrenia classification specificity and sensitivity, respectively, when used separately, but achieved around 70% sensitivity and specificity when used together. Parameters from piecewise linear and damped oscillator models of error/correct negativity signal and linear shock and oscillator models of error/correct positivity signal, when used alongside
person-average behavioral measures, achieved 76-88% sensitivity and
71-76% specificity. Piecewise linear and damped oscillator RT models led
to 72-74% sensitivity and specificity without using neural information,
which would be of high practical utility given the minimal training and
equipment required by the task. Notably, in contrast to the vast
majority of past studies, this performance was achieved using
human-interpretable machine learning in a heterogeneous sample including
individuals with various forms of psychosis, their siblings, and
individuals with other psychiatric syndromes. Model parameters thus
facilitated ruling out not only genetically and diagnostically typical
individuals, but also those who, despite not having schizophrenia
specifically, may share genetic and/or environmental liabilities and
even symptom overlap with those with schizophrenia. Examinations of
parameter associations (e.g., with post-error accuracy) provided further
evidence that models capture individual differences in functioning and
signal promise for the development of neural-behavioral link functions.