Presented by Marco Rotonda: Advances in the field of automated de-artifacting algorithms for electroencephalographic (EEG) signals have provided powerful tools to significantly speed-up data processing (Makeig et al., 1996; Delorme et al., 2007; Nolan, 2010; Mullen et al., 2015; Blum, 2019). Nonetheless, the efficacy of such, sometimes pre-implemented algorithms in marketed solutions for qEEG analysis still needs to be addressed. Here, we directly compared the performance of automatic algorithms implemented in different normative qEEG DataBases (DBs) currently available on the market (Neuroguide, qEEG-Pro, and iMediSync) with a manual artifact-rejection approach.
MATERIALS & METHODS (1): Raw, eyes closed resting-state EEG data from four exemplar subjects were fed into the pre-implemented, automated artifact-detection, and rejection or correction pipelines of the three DBs. Additionally, manual artifact-detection and rejection was performed separately. Thereafter, the absolute power values for five different frequency-bands (FBs; Delta, Theta, Alpha, Beta1, Beta2) were computed at each electrode position (19 channels, 10/20 layout) within each DB. The resulting values were assessed in a mixed-design ANOVA (within-subject factors: 2 (automated cleaning vs. manual rejection) x 3 (DBs) x 5 (FBs); between-subject factor: electrode).
RESULTS (1): Results suggested significant effects between DBs, which were driven by the amplitude of the absolute power values computed. Specifically, we found a main effect (ME) for DB and FB. Additionally, we found significative interactions between DB x Cleaning, Cleaning x FBs and DB x Cleaning x FBs.
INTERIM CONCLUSION (1): Taken alone, these results provide striking evidence that artifact-detections algorithm implemented in the selected DBs hugely impact data interpretation and can hence eventually diverge diagnostic interpretations when absolute power values are taken into consideration.
MATERIALS & METHODS (2): In order to correct for normalization biases across DB outputs in the current analyses, absolute power measures were transformed into Z-Score values across the whole dataset. Hereby, we sought to eliminate false positives in our statistical evaluation due to the different implementation of the power computation algorithms between DBs (White, 2003).
RESULTS (2): The mixed-design ANOVA suggested significant differences as a function of Electrode (DB x Electrode; FBs x Electrode).
INTERIM CONCLUSION (2): these results suggest that normalization processes can effectively reduce erroneous statistical differences due to differential implementation amplitude computations. Nonetheless, significant Electrode x DB interactions persist.
GENERAL DISCUSSION: Taken together, these preliminary findings underline the existent pitfalls regarding de-artifacting algorithms across DBs. Specifically, differential implementation of absolute power value computation greatly diverges between the present DBs investigated (similar: White, 2003; contrary: Thatcher & Lubar, 2009 and Keizer, 2018). Moreover, normalization of the data still produced significant differences across DBs.
CONCLUSION, OPEN QUESTIONS & FUTURE DIRECTIONS:
Several open questions remain to be addressed. First, the impact of differential cohorts employed between DBs for normalization processes has not been directly assessed in the present study. Second, whether automatic de-artifacting algorithms consistently out-perform manual data cleaning, needs to be further investigated. Third, the qualitative differences of the automatic de-artifacting algorithms employed by different DBs need further comparisons. We are aware of the limited sample size and aim at increasing the sample size.
