Machine learning identification of EEG features predicting working memory performance in schizophrenia and healthy adults

Abstract

Background: With millisecond-level resolution, electroencephalographic (EEG) recording provides a sensitive tool to assay neural dynamics of human cognition. However, selection of EEG features used to answer experimental questions is typically determined a priori. The utility of machine learning was investigated as a computational framework for extracting the most relevant features from EEG data empirically.

Methods: Schizophrenia (SZ; n = 40) and healthy community (HC; n = 12) subjects completed a Sternberg Working Memory Task (SWMT) during EEG recording. EEG was analyzed to extract 5 frequency components (theta1, theta2, alpha, beta, gamma) at 4 processing stages (baseline, encoding, retention, retrieval) and 3 scalp sites (frontal-Fz, central-Cz, occipital-Oz) separately for correctly and incorrectly answered trials. The 1-norm support vector machine (SVM) method was used to build EEG classifiers of SWMT trial accuracy (correct vs. incorrect; Model 1) and diagnosis (HC vs. SZ; Model 2). External validity of SVM models was examined in relation to neuropsychological test performance and diagnostic classification using conventional regression-based analyses.

Results: SWMT performance was significantly reduced in SZ (p < .001). Model 1 correctly classified trial accuracy at 84 % in HC, and at 74 % when cross-validated in SZ data. Frontal gamma at encoding and central theta at retention provided highest weightings, accounting for 76 % of variance in SWMT scores and 42 % variance in neuropsychological test performance across samples. Model 2 identified frontal theta at baseline and frontal alpha during retrieval as primary classifiers of diagnosis, providing 87 % classification accuracy as a discriminant function.

Conclusions: EEG features derived by SVM are consistent with literature reports of gamma's role in memory encoding, engagement of theta during memory retention, and elevated resting low-frequency activity in schizophrenia. Tests of model performance and cross-validation support the stability and generalizability of results, and utility of SVM as an analytic approach for EEG feature selection.

Keywords: EEG; Gamma frequency; Machine learning; Schizophrenia; Sternberg task; Support vector machine (SVM); Working memory.

Fig. 1

Example of Sternberg Working Memory Task (SWMT) trial depicting span of 4 items and time spans of pre-stimulus baseline, encoding, retention, and retrieval stages. Span ranged from 4–8 items, with span width and items selected randomly on a trial by trial basis

Fig. 2

Scatterplot of Sternberg Working Memory Task (SWMT) performance (out of 90 trials possible) as predicted by SVM Model 1 across the full study sample (N = 52). Multiple regression explained 76 % of the variance in SWMT performance based on frontal gamma activity during encoding and central theta 1 activity during retention. Both correct and incorrect trials entered the model for each feature. SVM Model 1 score (x-axis) represents the residual difference between predicted (trend line) and observed value for SWMT performance

Fig. 3

Scatterplot of MCCB Working Memory (WM) Composite score (standardized; t-score) as predicted by SVM Model 1 across the full study sample (N = 52). Multiple regression explained 42 % of the variance in MCCB WM score based on frontal gamma activity during encoding and central theta 1 activity during retention, with only data from correct trials entered entering the model for each feature. SVM Model 1 score (x-axis) represents the residual difference between predicted (trend line) and observed value for MCCB WM score

Fig. 4

Scatterplot of Continuous Performance Test-Identical Pairs version (CPT-IP) score (standardized; t-score) as predicted by SVM Model 1 across the full study sample (N = 52). Multiple regression explained 39 % of the variance in CPT-IP score based on frontal gamma activity during encoding and central theta 1 activity during retention for correct trials and occipital gamma activity at retrieval for incorrect trials. SVM Model 1 score (x-axis) represents the residual difference between predicted (trend line) and observed value for CPT-IP score

Fig. 5

Overlay of group average data for correct and incorrect trials extracted in gamma band (31.48 – 49.16 Hz) during encoding stage. Gamma increased significantly preceding incorrect relative to correct trials for both groups (paired-samples t tests; HC, t(11) = 5.37, p < 0.0005; SZ, t(39) = 7.01, p < 0.0005) and interacted by group (Wilk's Λ = 0.86, F_{(1, 50)} = 8.46, p = 0.005), with HC evidencing significantly greater range of modulation by accuracy level. Data submitted to statistical analysis was extracted from 1000–7000 ms, essentially containing the period of 200 ms prior to onset of first stimuli in set to 200 ms following onset of the 5^th stimuli of the memory set (or 1400 ms following offset of the 4^th stimuli). Importantly, as depicted in the figure, differences in gamma activity by accuracy were present before onset of first stimuli of each trial (0–1200 ms) and, therefore, are not interpreted to represent a memory load effect in these data