Alzheimer's disease (AD) is the most common type of dementia, characterized by a progressive loss of cognitive ability (Lane et al., 2018). The most prevalent signs of Alzheimer's disease include memory and orientation problems, as well as executive and motor dysfunctions (Guarino et al., 2019). At the molecular level, AD has been linked to synaptic impairment, as well as amyloid-beta and tau protein buildup (Bloom, 2014). As a result of this disease's progressive nature, loss of cognitive and motor function eventually leads to death.

Because there is currently no cure for AD, healthcare providers must rely on treatments aimed at delaying the illness's progression and lowering cognitive damage. The majority of these therapeutic methods rely on pharmacological substances (Cummings et al., 2019). Pharmacological medicines, on the other hand, are unable to provide an effective treatment for Alzheimer's disease, and, as a result of their extensive side-effect profile, they also cause other issues (Sharma, 2019). As a result, scientists have begun to investigate options other than pharmaceutical medications for treating this disease.

Transcranial magnetic stimulation (TMS), which has long been used successfully in the treatment of depression (Sonmez et al., 2019), is now being employed as a possible alternative treatment for other neurodegenerative diseases such as Alzheimer's disease (Freitas et al., 2011; Mimura et al., 2021), Parkinson's disease, and other dementias (Begemann et al., 2020; Hanoglu et al., 2021). TMS is a noninvasive approach that uses a magnetic field created outside the body to stimulate or inhibit specific areas of the brain by applying it to the scalp (Hamid et al., 2019). In some ways, TMS treatment appears to be superior than the pharmaceutical approach. TMS can be favored over pharmaceutical medicines due to its ease of use, low side effect profile, and excellent behavioral and cognitive benefits.

However, the effect of TMS on brain activity has not been fully understood (Sale et al., 2015). It is observed that some patients respond positively to treatment, while other gives no response (Dunlop et al., 2016). Unfortunately, this heterogeneity of treatment responses among patients has no obvious connection with the clinical measurements such as personality, demographic, sex, and age (Bailey et al., 2019). Hence, advanced computational methods are needed to analyze the connection between patients' certain properties and the TMS treatment to predict the patients' possible response to the treatment. That is necessary to generate efficient treatment strategies.

There are some studies with the purpose of predicting the TMS response of patients with psychiatric diseases. Chekroud et al. attempted to match depression patients with the treatment according to the patients' Quick Inventory of Depressive Symptomatology (QIDS) test rates (Chekroud et al., 2016). Another study analyzed structural magnetic resonance imaging results to predict the treatment outcomes of patients with schizophrenia (Koutsouleris et al., 2018). There are other studies focus on predicting the outcomes of treatments other than TMS, which are out of our focus on this study (Garnaat et al., 2019).

On the other hand, there are only a few studies on machine-learning analysis of EEG in predicting the effectiveness of rTMS. Corlier et al. (2019) generated a predictive model based on the functional connectivity on EEG, and as our purpose, they tried to obtain the most informative features for the rTMS benefit prediction. For the classification task, they used Elastiknet (Zou and Hastie, 2005), which randomly selects the subset of features and tries to maximize the classification accuracy over all random trials. Hasanzadeh et al. combined linear and non-linear features from EEG to distinguish TMS responders from non-responders, again on patients with depressive disorders. They also used the multi-source nature of the EEG data and tried to see the most significant electrode locations beside most significant EEG features. They used kNN classifier which is a primitive and low-capacity method that would fail on slightly complicated tasks (Hasanzadeh et al., 2019). Bailey et al. analyzed the electroencephalogram (EEG) signal of patients with depression to see whether it is possible to predict patients' response to rTMS treatment. They used band power and connectivity features with a linear SVM classifier to generate their prediction model (Bailey et al., 2019).

Besides these studies focusing on depressive disorders, to the best of knowledge, there is no significant attempt to generate a computational model for EEG analysis with the purpose of predicting the TMS response of patients with AD. Since AD is a neurodegenerative disease, the TMS therapy for patients with AD have more factors to be considered than depressive disorders. Besides, TMS effect is more variable in patients with AD than patients with depressive disorders (Iimori et al., 2019; Weiler et al., 2020). Hence, our task is more complicated, and the discussed methods above would not be applicable to the current task.

The EEG forms a critically high-dimensional dataset whereas it is highly challenging to collect a large number of samples for neurological disorders. Hence, the EEG data analysis for cognitive neuroscience studies is highly challenging. Due to the limited number of samples, studies in the literature mostly reduce the EEG data to a very low dimension by ignoring the advantage of the multi-source nature of EEG data.

In this study, we examined the possibilities to maximize the efficiency of the TMS treatments by predicting patients' responses before the TMS application by focusing on preserving the advantage of the multi-source nature of the EEG data. For that purpose, we analyzed pre-TMS EEG signals, which will be named as EEG1 throughout the paper for simplicity. Besides that, we investigated whether a dose application of TMS would increase the correlation between EEG2 and rTMS. For that purpose, we analyzed the post-TMS EEG signals, which will be named as EEG2 in the paper for simplicity (Figure 1). Throughout our analysis besides generating a prediction model, we seek more insight on EEG and TMS relation in patients with AD. For that purpose, we analyzed EEG1 and EEG2 from various perspectives. We measured the sensitivity and stability of the correlations between EEG1/EE2 and rTMS outcomes with respect to the small sample number. We applied multiple dimension reduction and classification methods, compared their performances, and interpreted the results.

The rest of the paper is organized as follows: section Materials and Methods provides the details of the dataset collection, data preprocessing, and the brief explanation of applied feature selection and classification techniques. It explains the process of generating the ground-truth labels and optimizing the machine-learning algorithms. Section Results describes all the applied data analysis steps. It explains our motive for further data analysis for more insight on the effects of TMS treatment on patients' EEG signals, the sensitivity measure of the data analysis process, comparative results of feature selection, and classification algorithms. Section Discussion discusses the obtained machine-learning results. It explains the interpretation of these results from a clinical view and lists the future motivation for this study.

All patients had mild to moderate AD symptoms (CDR = 1–2). The mean cognitive scores and demographic features of the patients are shown in Table 2. The change in MMSE scores before and after the TMS application in TMS+ and TMS– patients can be seen in Figure 4. Between two groups (MMSE pre-TMS and MMSE post-TMS), one-way ANOVA was applied, and statistically significant results were observed in MMSE pre-TMS group (F = 5.153, p < 0.05). According to post hoc analysis, statistical difference has been obtained between TMS+ and TMS– (t = −2.27, p < 0.05).

For the feature selection process, we applied PCA, wavelet transform, and SWLDA, and by the lead of SVM and SWLDA weights, we used only theta band features. We compared the classification accuracies with these different feature selection techniques (Table 3) where the classification accuracy is calculated as the percentage of the number of correctly classified samples to the total number of samples. For the classification analysis, we applied SVM, EEGNet, Shallow ConvNet, and SWLDA and compared their performances (refer to Table 4; Figure 5).

In the PCA, the decay of eigenvalues showed us that most of the information was stored in 20 features only. Using the most informative 20 features selected by PCA, we got an average of 62% accuracy on leave-one-out validation of EEG2. In the wavelet transform application for the feature selection purpose, by optimizing the numerical analysis, we decided to keep the largest 10% of the wavelet coefficients. Through that, we got 67% on the leave-one-out validation of EEG2. In the SVM and SWLDA, we observed that the average weights of theta wavelength were remarkably higher than the others (Figure 6). We used this idea as a lead and measured the performance of the data classification task using features collected only from theta wavelength which yield the best classification performance. In EEG1 analysis, we observed that none of the wavelengths nor the locations were given significantly more information than others.

To measure the effect of using the multi-source form of EEG signals, we analyzed the dataset with the power of each wavelength. As a result, we got 42% of accuracy overall for leave-one-out validation of EEG2. Hence, keeping the signals collected from each location gives significant information for classification purposes.

We observed the highest classification performance as 93% with SVM on theta band features. As it is seen in Table 5, regardless of the classification or feature selection technique, we always got better performance with EEG2 than EEG1. The average performance of the leave-one-out validation of EEG2, without any further feature selection or dimension reduction method, was recorded as 78%. In this analysis, we have seen that most of the patients always give a classification accuracy >50%. Always, the same patients had poor classification accuracies. For example, one of the patient's average EEG2 classification rate was about 10% (Figure 5A), which means that this patient information does not fit on the general of the other patients. However, we observed that the majority of the patients form a coherent team and give more than 90% of accuracy each. On the other hand, in the EEG1 analysis, we recorded 68% as the average accuracy and the coherence between the samples was lower. Furthermore, the gain by the feature selection and dimension reduction methods was higher in EEG2 analysis. Through the feature selection method application, the average performance of the leave-one-out validation of EEG2 increased to 93%, while we obtained 79% of accuracy on validation of EEG1.

After observing satisfactory results in the leave-one-out validation, we continued with the classification model generation. We obtained 82.5% of accuracy on the EEG2 test data as an average by the Monte Carlo technique, and 64% of accuracy on the classification of EEG1, by the same procedure. During our classification model generation, we observed that the optimum training sample ratio was about 65% (9 samples out of 14). As the mean of 50 random trials, we got 82.5% of accuracy while the standard deviation was 12.5%. So, the minimum accuracy was about 75%, which means that we got 3 correct predictions out of 4 on average by EEG2 analysis.

As a conclusion of our analysis, we observed that it is possible to predict the response of patients with AD to the TMS treatment by both EEG1 and EEG2 analyses. However, EEG2 is more informative than EEG1 with an average accuracy of 93 % while we obtained 79% on EEG1. Besides, due to our limited sample number, we focused on the data analysis and investigated the challenges of the EEG data analysis on patients with AD. We measured the feasibility of generating a personal TMS treatment route for the patients with AD and showed the possibility is higher if we use the lead of EEG signals after rTMS application. In addition to that, this type of algorithm, which can predict which patient will benefit from the TMS application by using the EEG data, will not only prevent patients from losing time with TMS treatment, but also will save on health expenditures. Thus, it will be possible to use TMS more effectively in neurodegenerative diseases such as Alzheimer's disease.

Due to its non-invasive nature, relative ease of administration, and low cost compared to other neuroimaging methods, EEG represents a promising tool for developing a personalized medicine approach for rTMS treatment (Garnaat et al., 2019). However, there is still a very limited number of studies in this area.

One of the previous studies that tried to distinguish responders from non-responders to rTMS treatment in depression using resting-state EEG analysis differentiated these groups and retrospectively identified several pretreatment variables such as higher anterior individual alpha frequency values, lower power in the frontocentral theta frequency band, and increased prefrontal delta and beta values (Arns et al., 2012). However, a repeat study to confirm these promising results found no significant differences between responders and non-responders in terms of individual alpha frequency, frontal theta, even after controlling for sex or age (Krepel et al., 2018; Garnaat et al., 2019).

There are only a few studies on machine-learning analysis of EEG in predicting the effectiveness of rTMS. All these attempts focus on depressive disorders. Although sophisticated methods are introduced, since the TMS therapy for patients with AD have more factors to be considered than depressive disorders, the given methods are weak to handle the current task.

In prospective studies, there were only a few possible predictors of the outcome for rTMS treatment for depression patients, and further research will be required to translate any of the findings described above into guidance for the implementation of treatment (Garnaat et al., 2019).

Our findings are particularly parallel to the findings by Bailey et al., He also reported that alpha and theta bands are useful to predict benefiting from rTMS therapy for depression (Bailey et al., 2019). However, in another study, they conducted with a larger and different sample group, and they could not replicate these findings for alpha and theta (Bailey et al., 2021). As it is known, the main element of the changes in EEG in Alzheimer's disease is the changes in alpha and theta (Bhattacharya et al., 2011; Güntekin et al., 2020). The results of our study are compatible with this general framework. There are also some findings showing that EEG event-related theta responses may be sensitive to neuromodulation interventions (Cespon et al., 2019; Güntekin et al., 2020). However, the impaired dynamics associated with EEG in AD are not unique only to theta and alpha. Again, it is clear that the localization of rTMS applications can cause responses at different frequencies by many different mechanisms. Indeed, Koch et al., showed that rTMS intervention on the precuneus for 2 weeks modulated only beta responses in AD (Koch et al., 2018). For this reason, there is a need for larger sample groups and further research on patients with other frequencies.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

The studies involving human participants were reviewed and approved by the Ethical Committee of Istanbul Medipol University the number of Ethical Report: 10840098–604.01.01–E.45425. The patients/participants provided their written informed consent to participate in this study.

LH designed and supervised all the experiments. HV conceived the experiments. CK designed the figures and analyzed and performed the machine-learning analysis. All authors performed the experiments and/or analyzed the data. All authors contributed to the interpretation of the results, writing the article, and approved the submitted version.

Istanbul Medipol University support for researchers for submission expenses.

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.