Machine-Learning-Based Prediction of Photobiomodulation Effects on Older Adults With Cognitive Decline Using Functional Near-Infrared Spectroscopy

Transcranial photobiomodulation (tPBM) has been widely studied for its potential to enhance cognitive functions of the elderly. However, its efficacy varies, with some individuals exhibiting no significant response to the treatment. Considering these inconsistencies, we introduce a machine learning approach aimed at distinguishing between individuals that respond and do not respond to tPBM treatment based on functional near-infrared spectroscopy (fNIRS) acquired before the treatment. We measured nine cognitive scores and recorded fNIRS data from 62 older adults with cognitive decline (43 experimental and 19 control subjects). The experimental group underwent tPBM intervention over a span of 12 weeks. Based on the comparison of the global cognitive score (GCS), merging the nine cognitive scores into a single representation, acquired before and after tPBM treatment, we classified all participants as responders or non-responders to tPBM with a threshold for the GCS change. The fNIRS data were recorded during the resting state, recognition memory task (RMT), Stroop task, and verbal fluency task. A regularized support vector machine was utilized to classify the responders and non-responders to tPBM. The most promising performance of our machine learning model was observed when using the fNIRS data collected during the RMT, which yielded an accuracy of 0.8537, an F1-score of 0.8421, sensitivity of 0.7619, and specificity of 0.95. To the best of our knowledge, this is the first study to demonstrate the feasibility of predicting the tPBM efficacy. Our approach is expected to contribute to more efficient treatment planning by excluding ineffective treatment options.

SECTION I. Introduction

An increasing number of older adults is experiencing neurological diseases accompanied by cognitive decline, highlighting the growing importance of maintaining cognitive health [1]. These diseases include subjective cognitive decline (SCD), a decrease in cognitive abilities despite lacking concrete evidence [1]; mild cognitive impairment (MCI), marked by perceptible cognitive deficits but retaining daily functionality and independence [2], [3]; and Alzheimer’s disease (AD), characterized by severe cognitive impairment and loss of daily independence [4]. According to the research framework of the National Institute on Aging and Alzheimer’s Association (NIA-AA), both the SCD and MCI are classified as symptomatic stages of AD [5]. This suggests that prompt and efficient interventions in the early stages of SCD or MCI could substantially slow down the progression to AD. Consequently, this enhances the cost-effectiveness of the treatment, thereby alleviating the burden on patients and their families [6], [7].

Transcranial photobiomodulation (tPBM), a recently introduced noninvasive neuromodulation modality that uses near-infrared light to modulate cortical excitability [8], demonstrated efficacy in enhancing cognitive function in individuals with SCD [9], MCI [10], [11], and AD [12]. Notably, this technique has fewer side effects [13] than conventional pharmacological treatments [14], [15] and can be readily implemented as a portable device [16], demonstrating its potential as an efficient early intervention tool for cognitive decline. However, its efficacy has not been consistent across all patients, similar to other noninvasive neuromodulation techniques [17]. Previous studies have reported non-responders to tPBM [18] and dosage-related inter-subjective variability influencing its treatment outcomes [19]. In this regard, the early prediction of the tPBM effectiveness could guide patients towards alternative treatment options, thereby increasing the cost-effectiveness of tPBM treatment in the healthcare and medical fields.

The effectiveness of tPBM has been evaluated using cognitive indicators [20] and neuroimaging techniques, such as electroencephalography (EEG) [21], [22], functional magnetic resonance imaging (fMRI) [9], and functional near-infrared spectroscopy (fNIRS) [10], [23], by comparing their variations before, during, and after treatment. Among these methodologies, fNIRS stands out as a promising assessment tool for tPBM because it uses near-infrared light in the same manner as tPBM, facilitating the integration of both techniques into a unified system [24], [25]. Using such a unified system, fNIRS can be easily recorded before the application of tPBM [26]. In addition, fNIRS provides several advantages compared to other neuroimaging techniques, such as relatively high spatial and temporal resolution, portability, low cost, and robustness to motion artifacts [27], [28]. Because of these benefits, fNIRS has been widely utilized in detecting changes in hemodynamic responses when participants perform tasks associated with motor and cognitive functions [29], [30].

In previous studies that evaluated the efficacy of tPBM using fNIRS, differences were observed in the time series of the hemodynamic responses (temporal information) [10], [31] and functional connectivity (spatial information) [32]. Specifically, Chan et al. showed that among participants with MCI, the hemodynamic response during a visual memory span test was lowered only in those who underwent PBM treatment and not in those who were given sham stimulation [10]. Additionally, Urquhart et al. reported decreased functional connectivity between the right and left prefrontal cortices and within the right prefrontal area during tPBM; however, after stimulation, this connectivity was intensified [32]. Despite these valuable insights into the neurological effects of tPBM, to the best of our knowledge, no study has proposed a method to predict the responsiveness to tPBM before treatment.

In this study, we hypothesized that marked hemodynamic disparities exist between tPBM responders and non-responders. Thus, we introduced a machine-learning-based methodology using fNIRS data acquired before tPBM intervention to distinguish between responders and non-responders to tPBM, who exhibited cognitive decline. Initially, the study participants were categorized into responders and non-responders based on the change in the singular composite cognitive score acquired from multiple cognitive scores before and after the 12-week tPBM treatment. To construct a machine-learning-based predictive model of the tPBM efficacy, we derived both temporal and spatial characteristics from the fNIRS data acquired while participants engaged in four specific tasks, each linked to a separate cognitive domain. The tasks included a resting task, Stroop task, recognition memory task (RMT), and verbal fluency task (VFT).

SECTION II. Methods

A. Participants

A total of 81 participants (60 in the experimental group and 21 in the control group) were initially enrolled in the experiment conducted at Inje University Ilsan Paik Hospital, Korea. The experimental group underwent a 12-week tPBM intervention to determine the effectiveness of tPBM. The control group did not receive any intervention during the same period; it was recruited to determine whether the improvement in the cognitive function within the experimental group was caused by the tPBM intervention or merely by repeated exposure to cognitive tasks (i.e., the learning effect). The inclusion criteria for the participants were as follows: (i) aged 60 years or above and in hospital owing to cognitive impairment concerns; (ii) a score ranging from 20 to 28 on the Korean Mini-Mental State Examination (K-MMSE); and (iii) no prior history or current diagnosis of severe depression, dementia, or other neurological disorders. All the participants signed a written informed consent form approved by the Inje University Ilsan Paik Hospital Institutional Review Board (IRB no. 2020-11-015).

Of the 60 participants in the experimental group, 14 were excluded from the study because they withdrew from enrollment. Two additional participants were excluded because of incomplete cognitive test scores. In the control group, two of the 21 participants were excluded due to withdrawal from the experiment. Consequently, 43 participants from the experimental group and 19 from the control group were included in this study.

B. Experimental Paradigm

The participants underwent home-based tPBM treatment sessions, each lasting 15 min, during which they watched a silent film. The frequency of these treatment sessions over a 12-week period was at the discretion of the participants. All 43 members of the experimental group strictly followed the tPBM protocol and successfully completed more than 20 sessions. The tPBM device was engineered by adjusting the light-source power and wavelength of a commercial wearable fNIRS device (NIRSIT-LITE; OBELAB, Seoul, South Korea) equipped with five light sources and seven photodetectors separated by 3 cm. The device emitted light from its sources at a power density of 80 mW/cm², employing dual wavelengths of 810 and 850 nm. The infra-red light was delivered to the brain in the form of a square waveform oscillating at a frequency of 1 Hz with a 50% duty cycle. These parameters were determined based on those employed in previous studies that demonstrated significant effectiveness [33], [34].

The enhancement of the cognitive function owing to the 12-week home-based tPBM intervention was evaluated by assessing nine cognitive metrics before and after the intervention. These metrics comprised the Digit Span Backward and Forward Tests (DST-F and DST-B, respectively); Digit Symbol Coding (DSC); components of the Korean version of the Auditory Verbal Learning test (AVLT), including Immediate Recall, Delayed Recall, and Recognition (AVLT-IR, AVLT-DR, and AVLT-RC); RMT; Stroop task; and VFT. The cognitive function assessments conducted in this study encompassed multiple cognitive domains: attention was evaluated using the DST-F and DST-B; language and related functions were evaluated using the VFT; memory function was evaluated using the AVLT and RMT; and the executive function was evaluated using the Stroop task [35]. Among these cognitive metrics, the RMT (memory), Stroop task (executive function), and VFT (language and related function) were performed simultaneously with the acquisition of fNIRS data. Prior to the administration of these three assessments, 3-min resting-state fNIRS data were acquired. Descriptions of the nine cognitive tasks, which have been widely utilized to evaluate cognitive functions, are detailed in Appendix. An explanation of the paradigms of the RMT, Stroop Task, and VFT, all of which were conducted alongside fNIRS measurements, is provided in Fig. 1A.

Fig. 1.

A) Channel configuration of fNIRS system utilized in this study. B) Structure of cognitive task experiments conducted alongside fNIRS data acquisition.

C. Categorization of Responder and Non-Responder to tPBM

Cognitive metrics representing a range of cognitive domains were aggregated into a singular composite score, termed the global cognitive score (GCS), to categorize individuals as either responders or non-responders to the tPBM intervention. The initial stage involved validating the reliability of these metrics as valuable indicators of tPBM effectiveness. We chose to focus only on the cognitive metrics that showed significant improvement after tPBM treatment because it was thought that including indices that do not show significant changes might reduce the sensitivity of the GCS to detect meaningful cognitive improvements. Consequently, a paired-sample t-test was conducted within the experimental group to identify specific cognitive metrics that demonstrated notable enhancements after tPBM treatment. These identified metrics were subsequently used to compute the GCS score for all the participants, encompassing both the experimental and control groups. The GCS was determined by computing the average z-score normalized values of the cognitive metrics that exhibited significant improvements following tPBM treatment [36].

Within the experimental group, participants were classified as responders or non-responders to the tPBM intervention based on a $\Delta$ GCS, calculated as the difference between the post-intervention GCS and the pre-intervention GCS. The threshold value for the z-score normalized $\Delta$ GCS was set at 0.5, which was commonly used in previous studies as a marker of significant change [36], [37], [38], [39], [40]. For instance, Heseltine, et al. calculated global cognitive indicators by averaging 23 z-score normalized cognitive metrics and utilized a threshold of 0.5 to identify the effect of peptide T on HIV patients with cognitive impairments [36]. Additionally, in the study by Harvey et al. [40], a 0.5 threshold for each z-score normalized cognitive score was employed to identify meaningful changes due to ziprasidone and olanzapine. Therefore, in this study, participants with a $\Delta$ GCS of 0.5 or above were classified as responders to tPBM, while those with a $\Delta$ GCS below 0.5 were classified as non-responders.

D. FNIRS Recordings and Preprocessing

A wearable and wireless fNIRS device (NIRSIT-LITE; OBELAB, Seoul, South Korea) was used to acquire fNIRS signals. The device was composed of five dual-wavelength laser diodes (780/850 nm) and seven photodetectors with a separation of 3 cm between the laser and detector. The fNIRS signals were acquired from 15 channels at a sampling rate of 8.138 Hz. Detailed depictions of the placement of laser diodes (light sources), photodetectors, and their respective channels are shown in Fig. 1B.

The fNIRS data were preprocessed using MATLAB 2022b (MathWorks, Natick, MA, USA) in conjunction with the functions available in the BBCI toolbox (https://github.com/bbci/bbci_public). The changes in the concentrations of oxygenated hemoglobin ($\Delta$ HbO) and deoxygenated hemoglobin ($\Delta$ HbR) for each channel were derived from the optical density data using the modified Beer-Lambert law (MBLL) [41]. The $\Delta$ HbO and $\Delta$ HbR were processed using wavelet coefficient-based filtering to mitigate motion artifacts [42], and a band-pass filter within the 0.01–0.2 Hz frequency range employing a 6th-order Butterworth zero-phase filter was utilized to attenuate physiological noises [43], [44]. The preprocessed $\Delta$ HbO and $\Delta$ HbR data were then segmented according to the task block periods. Among the four tasks accompanying the acquisition of the fNIRS data (rest, RMT, Stroop task, and VFT), the task block execution periods for the RMT and Stroop task varied among the participants owing to different reaction times. Therefore, the raw fNIRS data for these tasks were segmented with fixed durations of 234 and 47 s for the RMT and Stroop task, respectively, for all the participants. A single segment was obtained for both the resting state and RMT, whereas four and two segments were obtained for the Stroop task and VFT, respectively, as inter-session breaks were provided to the participants for these two tasks. Baseline correction was applied to all the segments by subtracting the mean temporal value calculated over the interval from −0.5 s to 0 s with respect to the onset of each segment.

E. Feature Extraction

Candidate features were extracted from $\Delta$ HbO and $\Delta$ HbR, and the change in the total hemoglobin ($\Delta$ HbT = the sum of $\Delta$ HbO and $\Delta$ HbR) considering both temporal and spatial characteristics. The candidate features included statistical features directly extracted from the time series data of each channel, such as the mean, variance, kurtosis, skewness, and range; the range was defined as the difference between the maximum and minimum values in $\Delta$ HbO, $\Delta$ HbR, and $\Delta$ HbT, as well as the connectivity features represented by the functional connectivity calculated using the Pearson correlation coefficients between the channel pairs based on a previous study [45]. For data from the VFT and Stroop task, consisting of multiple segments, all candidate features were averaged across these segments. Consequently, for all the tasks, the total number of candidate features was 540, comprising 225 statistical features (15 channels $\times 5$ metrics $\times 3$ hemodynamic representations) and 315 connectivity features (₁₅C$_{2} \times 3$ hemodynamic representations).

F. Feature Selection and Classification

A support vector machine (SVM) classifier with ridge regularization was employed, for which the function of “fitclinear” in MATLAB 2022b (MathWorks, Natick, MA, USA) was used. A leave-one-subject-out cross-validation (LOSO CV) strategy was used to evaluate the classification accuracy of the classifier model. In the LOSO CV process, all data except one subject’s data were used for training, and this division was repeated until all the subjects’ data were tested. The accuracy, F1-score, sensitivity, and specificity were calculated and averaged over all the LOSO CV iterations. Sensitivity was defined as the probability of accurately classifying responders, whereas specificity was defined as the probability of correctly classifying non-responders.

During each LOSO CV iteration, an optimal subset of features was selected from the training set. This selection was based on the rank order determined by the Fisher’s score, a widely recognized method for supervised feature selection [46]. The “N-feature accuracy” metric was assessed by averaging the outcomes of all the LOSO CV iterations for a specific number of features (N), with N varying from 1 to 15 [3]. For the classification of the responders and non-responders to tPBM, the accuracy was identified as the highest accuracy obtained among the N-feature accuracies.

SECTION III. Results

A. Statistical Analysis of Cognitive Indicators

In our study, a paired sample t-test was conducted to identify the cognitive metrics that demonstrated significant improvements following tPBM treatment. The results indicated that most cognitive metrics, with the exception of the DST-B and DSC, exhibited statistically significant improvements, as presented in Table I. Consequently, the scores from the VFT, AVLT-IR, AVLT-DR, AVLT-RC, DST-F, Stroop task, and RMT were utilized in the computation of the GCSs before and after the intervention. Using a $\Delta$ GCS threshold of 0.5, 22 responders and 21 non-responders were identified. The average frequencies of the tPBM intervention for the responder and non-responder groups were 59.1818 ± 17.0032 and 57.1429 ± 21.0364, respectively, with no statistically significant difference ($p =0.7336$ ).

TABLE I Cognitive Task Scores for the Experimental Group Recorded Before and After Undergoing the tPBM Intervention

B. Demographics of Responder, Non-Responder, and Control Groups

Demographic analysis of the responder, non-responder, and control groups was conducted using the Chi-square test to evaluate the gender differences and the Kruskal-Wallis test [47] to assess the differences in the MMSE scores, educational levels, and age. Additionally, the GCS recorded before the tPBM intervention was subjected to statistical analysis. This analysis aimed to ascertain the existence of a ceiling effect, which would indicate a lack of response to tPBM intervention among participants who had already exhibited relatively high cognitive scores. The comprehensive demographic information is presented in Table II. The analysis did not reveal any statistically significant differences among the three groups in terms of the gender composition, MMSE scores, educational level, age, and GCS recorded before the tPBM intervention. However, there were significant differences among the three groups in terms of the $\Delta$ GCS ($p \lt 0.001$ ); therefore, a post-hoc analysis was carried out to further evaluate these differences using the Wilcoxon rank-sum test [48].

TABLE II Demographic Information of the Responder, Non-Responder, and Control Groups

The $\Delta$ GCS measurements for each group, along with the results of the post-hoc analysis, are shown in Fig. 2. The $\Delta$ GCS for the responder group was found to be significantly higher than those of the non-responder and control groups, with Bonferroni-corrected p-values less than 0.001 in both comparisons. In contrast, no significant difference in the $\Delta$ GCS was observed between the non-responder and control groups (Bonferroni-corrected $p =0.2639$ ). These findings showed that the observed $\Delta$ GCS in the responder group could be attributed to the effects of the tPBM, rather than to the learning effect of cognitive tasks, especially when compared to the control group.

Fig. 2.

Box plots representing the $\mathsf { \Delta }$ GCS of the responder, non-responder, and control groups. The median value is indicated by the red line within each box plot. Significance levels are marked as follows: * indicates a Bonferroni-corrected p <0.05, $^{\ast \ast }$ denotes a Bonferroni-corrected p <0.001, and $^{\ast \ast \ast }$ indicates a Bonferroni-corrected p <0.0001.

C. Classification Results

Table III presents the performance metrics, including the accuracy, F1-score, sensitivity, and specificity, for the machine learning-based classification aimed at distinguishing between the responder and non-responder groups. These metrics varied with respect to the different cognitive tasks and selected feature subsets. Notably, the highest performance was reported for the RMT. For this task, when the candidate feature subset included both statistical and connectivity features, the model achieved an accuracy of 0.8537, an F1-score of 0.8421, sensitivity of 0.7619, and specificity of 0.95, with five selected features. Additionally, for the resting-state task, the highest recorded accuracy was 0.7073 when the candidate feature subset consisted exclusively of statistical features. In contrast, for the Stroop task and VFT, the highest accuracy of 0.7561 was achieved when the candidate feature subset comprised only connectivity features.

TABLE III Performance Metrics of the Machine Learning Model in Classifying Individuals Into Responder and Non-Responder Groups, Detailed According to the Types of Cognitive Tasks and Candidate Feature Subsets

D. Feature Analysis

To evaluate the characteristics of the hemodynamic responses of responders and non-responders to tPBM intervention from the fNIRS data acquired before the intervention, we analyzed the five most frequently selected features in the RMT paradigm that had the highest accuracies among the four paradigms. These features included the mean of $\Delta$ HbO in channel 7, $\Delta$ HbT in channel 11, connectivity of $\Delta$ HbR between channels 10 and 13, kurtosis of $\Delta$ HbR in channel 13, and skewness of $\Delta$ HbT in channel 7. The average values of these five features for the responders and non-responders are presented in Table IV. According to the Wilcoxon rank-sum test, all five features exhibited significant differences between the responder and non-responder groups. The mean values of $\Delta$ HbO in channel 7 and $\Delta$ HbT in channel 13, along with the kurtosis of $\Delta$ HbR in channel 13 and the skewness of $\Delta$ HbT in channel 7, were statistically higher in the responder group than in the non-responder group. Conversely, the connectivity of $\Delta$ HbR between channels 10 and 13 for the responder group was lower than that of the non-responder group.

TABLE IV Values of the Five Most Frequently Selected Features in the RMT Paradigm for Both the Responder and Non-Responder Groups

SECTION IV. Discussion

In this study, we categorized older adults with cognitive decline into responders and non-responders to a 12-week tPBM treatment based on the scores of various cognitive tasks and evaluated the possibility of classifying these groups using fNIRS data collected prior to the tPBM intervention by applying a machine learning-based classification strategy. Notably, by employing fNIRS data acquired during the RMT, a task associated with the working memory, the machine learning model exhibited a remarkably high prediction accuracy of 85.37%. To the best of our knowledge, this is the first attempt to distinguish between responders and non-responders to tPBM using fNIRS data recorded before the intervention.

Our methodology has the potential to significantly enhance the cost-effectiveness of tPBM therapy. Predicting the treatment outcomes of patients before tPBM administration facilitates the implementation of a more targeted therapeutic approach. Specifically, if treatment is predicted to be ineffective in certain individuals, alternative treatment modalities should be considered. This proactive approach not only provides the opportunity to optimize the use of medical resources, but also ensures that patients are provided with the most suitable treatment from the beginning. This strategy will be instrumental in the improvement of the overall efficacy of treatment regimens for the elderly with cognitive decline.

In this study, we hypothesized that fNIRS data recorded prior to the tPBM session might provide valuable insights into responsiveness to tPBM, as the NIRS system also utilizes near-infrared light, which can induce an instantaneous PBM effect [24], [25]. Previous studies have shown that participants exposed to fNIRS light exhibited improvements in both cognitive task performance and task-related hemodynamic responses compared to control groups who were not exposed to the light. Therefore, it seems reasonable to predict responsiveness to tPBM in advance using fNIRS, as the fNIRS measurements themselves may reflect the modulatory effects of near-infrared light.

To comprehensively identify the impact of tPBM across multiple cognitive domains and the feasibility of the proposed categorization methodology, we analyzed the changes in the cognitive scores used to evaluate the GCS, for the responders, non-responders, and control subjects. The Wilcoxon signed-rank test was employed to evaluate the differences in each cognitive score before and after the tPBM intervention [49]. The cognitive scores for each group, along with the associated statistical results, are presented in Fig. 3. In the responder group, significant enhancement was observed in all cognitive tasks after tPBM treatment. This was in contrast with the non-responder group, in which no significant differences were detected in any of the cognitive scores before and after the intervention. Additionally, the impact of tPBM became more evident compared to the control group. Although the control group showed a statistically significant increase in only three cognitive scores related to memory tasks (AVLT-RC, RMT, and AVLT-IR), the improvements in the responder group were greater. These comparisons not only highlighted the potential of tPBM in enhancing various cognitive functions for specific individuals, especially when compared to natural changes, such as learning effects observed in control groups, but also demonstrated the feasibility of the proposed categorization methodology.

Fig. 3.

Comparative presentation of cognitive scores for the responder, non-responder, and control groups, both before and after the tPBM intervention. The y-axis denotes the scores for each cognitive task. Significance levels are marked as follows: * indicates p <0.05, ** denotes p <0.001, and *** indicates p <0.0001.

To identify whether our prediction approach is robust across different machine learning models, we employed shrinkage linear discriminant analysis (sLDA), which has been widely used in fNIRS classification problems due to its high classification performance [50]. The prediction performance of sLDA for responsiveness to tPBM was described in Table S1 of the Appendix. Although the overall performance of sLDA was lower than that of SVM, sLDA demonstrated the highest performance in the RMT paradigm, achieving an accuracy of 0.8293, which was on par with the performance of SVM. This indicates that while SVM generally outperformed sLDA across various paradigms, sLDA still showed strong results, highlighting its utility and the robustness of our approach across different machine learning models.

This study demonstrated that the optimal feature subsets varied depending on the fNIRS data acquired during each specific cognitive task. Although previous studies have shown that hemodynamic responses differ depending on the cognitive task or domain under examination, there is notable variability in these responses, even within similar cognitive domains [51], [52]. Additionally, there are no studies that have proposed biomarkers or specific characteristics to distinguish between responders and non-responders to tPBM intervention based on the hemodynamic responses. The absence of prior research in this field poses a significant challenge in comprehensively interpreting the results obtained from our machine-learning paradigms.

To determine the potential fNIRS-based biomarkers for distinguishing between the responder and non-responder groups, we evaluated the most frequently selected features in the RMT paradigm, which also showed statistically significant differences between the groups. Among these identified features, we conducted an in-depth analysis of the mean values of $\Delta$ HbO in channel 7, motivated by its alignment with findings from a previous study [10], which reported a decrease in the $\Delta$ HbO during a visual memory span test exclusively in subjects who had undergone tPBM treatment. In line with this previous research, our study found that the average value of $\Delta$ HbO in channel 7 decreased after tPBM intervention in the responder group, whereas no such change was observed in the non-responder group, as depicted in Figure S1 in Appendix. Consequently, these findings led to the hypothesis that individuals with a relatively higher baseline $\Delta$ HbO were more likely to benefit from tPBM.

This hypothesis offers the rationale that the mean value of $\Delta$ HbT in channel 11 was selected as one of the dominant features to construct our machine-learning model, given that it showed statistically higher values in the responder group than in the non-responder group. Additionally, the connectivity of $\Delta$ HbR for channels 10 and 13 serving as a potential biomarker could be inferred with the inverse application of our hypothesis. Based on a previous study demonstrating that functional connectivity in the prefrontal area increases after tPBM intervention [32], tPBM may exhibit positive effects in participants with relatively low baseline functional connectivity. In line with this, in this study, the connectivity of $\Delta$ HbR for channels 10 and 13 was statistically lower in the responder group than in the non-responder group before tPBM intervention, and increased in the responder group from 0.2357 ± 0.3889 to 0.3309 ± 0.3513 after the intervention (no statistically significant difference). Although the kurtosis of $\Delta$ HbR in channel 13 and the skewness of $\Delta$ HbT in channel 7 present a challenge in directly correlating with our hypothesis because of the lack of related previous studies, it is notable that these features are derived from channels corresponding to the features mentioned earlier. Therefore, it is speculated that the differences in the baseline hemodynamic activation and functional connectivity in specific channels could yield a high prediction accuracy for tPBM efficacy in the machine learning paradigm, incorporating additional hemodynamic characteristics from these channels.

Despite our efforts to determine the distinct differences between the responder and non-responder groups, there was a limitation owing to the relatively small fNIRS dataset to generalize our findings. To develop more robust prediction model and physiologically comprehensive fNIRS-based biomarkers capable of differentiating between tPBM responders and non-responders, it may be necessary to collect additional fNIRS data from a more extensive participant cohort, both before and after tPBM intervention. Therefore, in future studies with an expanded dataset, we plan to propose comprehensive physiological insights into the effectiveness of tPBM and enhance the reliability and generalizability of our machine-learning-based methodology for the prediction of the response to tPBM. Furthermore, we intend to develop a personalized tPBM intervention approach. This will involve evaluating the associations between the biomarkers derived from our insights and various tPBM parameters, such as the control of fluence or irradiance, to optimize treatment strategy for individual patients.

Appendix

Detail paradigms of cognitive tasks utilized in this study, Table S1 and Figure S1 are included in this Appendix section.