Major depression disorder (MDD) is one of the most prevalent mental disorders. According to the 2016 National Survey on Drug Use and Health (Substance Abuse and Mental Health Services Administration, 2017), 6.7% of all U.S. adults had at least one major depressive episode in the past year. For many patients, major depression is a huge burden that can cause severe impairments to their life. Diagnosing depression in the early stage is considered to be crucial to prevention and treatment process (Willner et al., 2013). However, since the current diagnosis of depression is mainly based on the psychiatric interview, the subjectivity in its nature may make an early diagnosis challenging. In a meta-analysis study, Mitchell et al. (2009) found that among all depression cases, general practitioners only identified 47.3% of them (95% CI 41.7–53.0%). In order to facilitate diagnosis, many researchers started to focus on physiological data to explore biomarkers of MDD.

Previous studies have found that MDD patients have different neurophysiological characteristics compared to healthy controls; for example, reduced volume of orbitofrontal cortex (Bremner et al., 2002, Lai et al., 2000), altered connectivity in cingulate cortex networks (Fox et al., 2012, Pizzagalli, 2011), increased EEG absolute alpha power during rest (Knott et al., 2001, Zoon et al., 2013), elevated slow wave activity (Kwon et al., 1996, Roemer et al., 1992), and lower or flat levels of galvanic skin response (GSR; Greenfield et al., 1963, Vahey and Becerra, 2015), etc. Specifically, frontal lobe dysfunction which is associated with impaired cognition is often regarded as a marker for depressive disorders (Barabassy et al., 2010). Ahmadlou et al. (2012) investigated EEGs of the frontal brain of MDD patients and found that MDD patients had higher levels of fractality of left, right and overall frontal lobes in beta and gamma sub-bands compared with healthy controls. The changes of EEGs in the prefrontal area have also been proved to be associated with clinical response to SSRI antidepressants and to venlafaxine (Iosifescu et al., 2009). Moreover, MDD patients also showed different behavioral patterns. For instance, eye tracking studies have found that depressed individuals tend to spend more time viewing negative images and less time viewing positive images (Duque and Vázquez, 2015, Kellough et al., 2008).

Based on those findings, recently, machine learning approaches have been used widely to mine biomarkers from neuroimaging, EEG and behavioral data to establish computer-aided diagnosis of depression (e.g., Acharya et al., 2015, Hosseinifard et al., 2013, Hu et al., 2015, December, Kang et al., 2016, January, Kim and Na, 2018, Gotlib, 1998, Sato et al., 2015). Classification algorithms including Support Vector Machine (SVM; Hearst et al., 1998), Logistic Regression (Fan et al., 2008), Random Forest (Liaw and Wiener, 2002) were often applied to build classification and predictions. The research results are promising. For example, Liao et al. (2017) collected EEG from 12 MDD patients and 12 healthy controls and extracted features via a spectral-spatial extractor. The classification model using SVM achieved an average classification accuracy of 81.23%. Similarly, Mumtaz et al. (2017) recruited 34 MDD patients and 30 healthy controls in their study and used EEG-derived synchronization likelihood features as input to train a machine learning classifier. The accuracy of classification was over 90%.

However, the previous studies are limited in two perspectives. Firstly, most of the studies that measured neurophysiological data needed assistance from professions. And the cost of the measurement was relatively high. A low-cost, easy implementation and non-invasiveness measure is needed in order to make machine learning approach more available and accessible. Secondly, many studies examined the accuracy of the machine learning models in a quite small sample. There is a need to further replicate the findings with a larger sample.

In the present study, EEGs of the frontal lobes, eye-tracking information and GSR data were recorded via portable, low-cost devices. There are two reasons for just monitoring EEGs of the frontal lobes. Firstly, previous studies have indicated that frontal lobes are linked to affect expression and regulation (Davidson et al., 1990, Dawson et al., 1992), and EEGs in the frontal lobe dysfunction is often regarded as a marker for depressive disorders (Barabassy et al., 2010). Secondly, compared to other areas of the brain, it is more accessible and possible to use portable, low-cost devices to measure EEGs in the frontal lobes since convenient dry electrodes can be utilized in this situation. Features that extracted from the three modalities were used as input to a multimodal machine learning approach to classify depression patients and healthy controls. Three machine learning algorithms including Random Forests, Logistic Regression and Support Vector Machine were trained to build a dichotomous classification model based on a relatively larger sample.