Schizophrenia is a severe and persistent debilitating psychiatric disorder with a prevalence of about 1% of the world population (McGrath et al., 2004). Although psychotic symptoms such as hallucinations and delusions are frequently present, impaired information processing is probably the most common symptom (Javitt et al., 1993). This deficit is reflected mainly by deficits in attention and working memory tasks when compared with healthy controls (Li et al., 2018). The diagnosis of schizophrenia is made by psychiatrists by ascertaining the presence of predefined symptoms (or their precursors) with personal interviews. However, in some cases this diagnosis is unclear, or patients are misdiagnosed with Schizophrenia (Coulter et al., 2019). Thus, finding biomarkers for the prediction of individuals with schizophrenia would be desirable to enable choosing the optimal treatment (pharmacologic or non-pharmacologic). Analysis of the electroencephalogram (EEG) during information processing tasks could provide objective complementary measures to support the subjective human-based decision process (Sabeti et al., 2009; Koukkou et al., 2018).

EEG is a non-invasive and low-cost technique used to measure electrical brain activity along with multiple scalp locations. EEG signals have been widely adopted to study mental disorders, such as dementia, epileptic seizures, cognitive dysfunction, among others, as well as schizophrenia (Loo et al., 2016; Olbrich et al., 2016; Horvath et al., 2018). Electrophysiological data reflects the spontaneous activity of myriad brain parcels, but also can include responses to afferent stimuli (Cong et al., 2015). Event-related potentials (ERPs) are electrical responses that are time-locked to a specific stimulus or event and can be used to assess brain dynamics during information processing in specific tasks (Woodman, 2010). When a subject is presented with a series of standard stimuli, interspersed with infrequent deviant stimuli, the Mismatch Negativity (MMN) (Lee et al., 2017) and the P300 (Li et al., 2018) components are generated. This task is known as the oddball paradigm and is used to study schizophrenia since consistent deficits in the P300 and MNN have been reported in this disease (Bramon et al., 2004; Javitt et al., 2017). Although MMN and P300 are usually produced by an infrequent unexpected event in a sequence of auditory stimuli, P300 can also be obtained with visual stimuli. The MMN is of shorter latency and does not require attention to the stimulus (Näätänen et al., 2004), whereas the P300 is of longer latency and requires attention to the stimulus (Huang et al., 2015).

Several studies have reported significant differences in the latency and amplitude of MMN and P300 between controls and patients, suggesting that these features are possible markers of the prodromal phase of schizophrenia (Atkinson et al., 2012; Loo et al., 2016) as well as potential endophenotypes for schizophrenia (Earls et al., 2016). Analysis of a large dataset of auditory P300 ERP (649 controls and 587 patients) confirmed the reliability of this reduced amplitude, with a large effect size (Turetsky et al., 2015). However, these findings of statistically significant differences in a group analysis do not imply that EEG is useful for the prediction of individual schizophrenia cases (Lo et al., 2015), which requires applying a prediction paradigm using Machine Learning.

Machine learning techniques have potential value for assisting the diagnosis of brain disorders (Burgos and Colliot, 2020). Recent works are based on EEG signals for the diagnosis of epilepsy (Tanu, 2018), Alzheimer's disease and dementia (Joshi and Nanavati, 2021), and Parkinson (Maitín et al., 2020), among other disorders. Particularly, ERP measures combined with machine learning techniques are being tested for the classification of schizophrenia. The most common features used are based on the amplitude and latency of different components [e.g., N100 and P300 (Neuhaus et al., 2013), P50 and N100 (Iyer et al., 2012; Neuhaus et al., 2014)], with several classifiers tested. Neuhaus et al. (2013) used visual and auditory oddball paradigms and a k-nearest neighbor (KNN) classifier and obtained a classification accuracy of 72.4 %. The same author with a bigger sample size and a Naive Bayes (NB) classifier achieved a 77.7% of accuracy (Iyer et al., 2012). Laton et al. (2014) evaluated the performance of several classifiers extracting features from auditory/visual P300 and MMN. The results using NB and Decision Tree (without and with AdaBoost) achieved accuracies of about 80%. Recently, Barros et al. (2021) published a critical review that summarizes machine learning-based classification studies to detect SZs based on EEG signals, conducted since 2016. These authors reported that Support Vector Machines (SVM) were the most commonly used classification algorithm, probably due to their computational efficiency. This kernel-based learning method also achieved the best performance in most studies. Nevertheless, to the best of our knowledge, none of the studies focused on ERP for SZ classification have used multiple kernels, employing instead only one specific kernel function.

The multiple kernel learning (MKL) method learns a weighted combination of different kernel functions and can benefit from information coming from multiple sources (Wani and Raza, 2018). A recent survey of artificial intelligence methods for the classification and detection of Schizophrenia (Lai et al., 2021), shows that MKL has been applied to both structural and functional Magnetic Resonance Images (MRI), increasing performance accuracy (Ulaş et al., 2012; Castro et al., 2014; Iwabuchi and Palaniyappan, 2017). Nevertheless, in this review MKL algorithms applied to electrophysiological data have been not reported, although a recent study used EEG dynamic functional connectivity networks to classify SZ based on MKL (Dimitriadis, 2019). To our knowledge, ERP data has not been used to classify SZ using MKL despite its use for other purposes such as brain-computer interfaces (Li et al., 2014; Yoon and Kim, 2017).

Here, we explore the efficacy of MKL for the classification of schizophrenia based on ERP measures extracted from auditory and visual P300 and MMN. Using the same dataset provided by Laton et al. (2014), we extended the set of predictor variables beyond the latency and amplitude of the ERP components, by including additional morphological features (based on time) together with some features extracted from the frequency domain. Due to the huge number of features, the Boruta method (Kursa and Rudnicki, 2010) was applied, which is a wrapper Random Forest (RF) based feature selection algorithm, to estimate the impact of a subset of important and relevant feature variables in the classification accuracy.

In this work, we explored the use of MKL classification for distinguishing SZs from HCs based on ERP data. Using all features, the best classification accuracy (83%) was achieved when kernels were built by grouping features according to paradigms. Moreover, when MKL was combined with the Boruta features selection method, a classification accuracy of 86% was obtained. With this feature selection algorithm, the large number of predictor variables was reduced significantly (96%) with a lower computation time. Therefore, the training time of MKL was also reduced [0.18 (std 0.03) seconds per inner cross-validation loop], thus solving one of its main shortcomings: its high computational cost, especially when many features are used (de Carvalho, 2019).

The feature selection algorithm results showed that the variables that contributed most to the discrimination were related to the auditory P300 paradigm. This corresponds with the general finding that auditory P300 measures are more effective in differentiating SZs from HCs than those obtained from the visual stimuli (Park et al., 2005). The selected features were peak related, mainly related to amplitude, latency, and their combination. To a lesser extent, peak to peak related features were included in the selection. However, only three Signal related features were occasionally included. Thus, features from the frequency domain did not contribute much to the improvement of the classification.

Our results are in line with prior works. A previous study (Santos-Mayo et al., 2017) proposed a system to help diagnose schizophrenia by analyzing P300 signals during an auditory oddball task. The authors extracted time and frequency domain features similar to ours but using different collections of signals from electrodes in different regions of the scalp. Our results are comparable to theirs when the electrode groupings were used, but they obtained larger AUC values (more than 0.95) for their Left and Right hemisphere electrode groupings. We did not explore these locations. However, their dataset was unbalanced and small, which possibly limits the reliability of their findings. Other authors using also P300 amplitude and latency values as features (Shim et al., 2016) reported classification accuracies of 81% using an SVM classifier. When they combined these features with a selection of source-level density measures they increased their accuracy to 88%, a result similar to ours. Laton et al. (2014) also extracted latency and amplitude features of responses to three different odd-ball tasks and applied several classification algorithms. They achieved an average accuracy of 77% (std = 3.5). Their best result (about 85%), corresponded to an RF classifier, comparable to our results. Laton et al. (2014) also found that in a ranking of the 20 main variables contributing to the classification, 14 were extracted from the P300 auditory oddball paradigm. This suggests that, out of the three ERP paradigms used, the auditory P300 contributes most to the classification which is congruent with our findings.

One limitation of our study is that we did not use the spatial distributions of the ERPs over the scalp. Further research should include features using ERP scalp topographic maps (STM). This would take advantage of the differences in STM between schizophrenia and normal control groups reported by different authors (Morstyn et al., 1983; Frantseva et al., 2014). This is a pure image processing approach. Another track is to use independent component analysis (ICA) to split up the multi-channel ERP data into several independent spatiotemporal components. ICA separates the mixed signals into unmixed signals which are statistically independent. These approaches could generate features for a classifier. Another limitation of the present study is the small sample size which is usual in psychiatric cohorts from one site. We addressed this limitation using cross-validation strategies. However, training with larger data sets (possibly from multiple sites) would yield a more stable and reliable estimate of future performance and guarantee better generalization.

Using Multiple Kernel Learning (MKL) classifiers on features defined for ERP obtained in oddball paradigms, it was possible to distinguish SZs from HCs with a classification accuracy up to 86%. Accuracy improved when the Boruta feature selection was applied. The auditory P300 provided the most informative features. Future work should explore new ERP features including topographic information.

The data analyzed in this study is subject to third party restrictions, which were used under license for this study. Requests to access these datasets should be directed to Laton et al. (2014).

The provenance of data in the study (which involved human participants), and the adherence to adequate ethical standards, were reviewed and their use approved by Ethics Committee of Cuban Neuroscience Center. The patients/participants provided their written informed consent to participate in this study, as stated in the original article where they were described.

ES wrote and edited the manuscript, developed the theory, and carried out the experiments and the interpretation of the results. MO reviewed the manuscript. MV critically revised the manuscript. HS conceived the present idea and supervised it. All the authors discussed and agreed with the main focus and ideas of this article and contributed to the methodology and analysis.

This work was supported by the VLIR-UOS project A Cuban National School of Neurotechnology for Cognitive Aging (NSNCA), Grant number CU2017TEA436A103.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.