The analysis of 21 selected publications highlights the evolving application of Spiking Neural Networks (SNNs) in neuroimaging over the past decade, as shown in Figure 2. The citations of all 21 publications were collectively analyzed, as presented in Table 2.

States reveal that, from 2015 to 2017, this area of study was in the early exploration phase, as only two publications from the selected studies were available each year. Moreover, the trend shows the same results for 2018, 2019 and 2020, limiting studies to just one article per year from the chosen dataset. The reason for this decline could be methodological challenges.

However, there was a noticeable rise in research in 2021 with four publications, suggesting increasing feasibility with advancements in deep learning. After a drop in 2022 to only one study (which could be due to the selected dataset), results showed a surge in 2023 in this research area, with five studies marking a significant shift toward SNN applications. The most probable reason for this growth could be the improvements in computational resources and model efficiency.

However, the publication trend declined to four in the year 2024, still reflecting that this research area is reaching stability and reflecting the growing integration of SNNs in neuroimaging. Overall, the research has progressively grown from initial exploration to more practical implementation and refinement, which reflects a growing confidence in the field.

Based on classifications from the selected databases, the chosen articles are linked with 10 distinctive research fields, with some spanning multiple areas. In total, 40 connections have been established between the 21 publications and these research domains. As depicted in Figure 3, many publications fall within “Computer Science” (n = 14), followed by “Neuroscience” (n = 8) and “Medicine” (n = 6).

From the perspective of the biomedical research field, a collective contribution is found to make it a multidisciplinary nature of studies, such as “Medicine” (n = 6), “Biochemistry, Genetics, and Molecular Biology” (n = 2), and “Immunology and Microbiology” (n = 1).

From an engineering and computational standpoint, “Computer Science” (n = 15) and “Engineering” (n = 3) reflect a major division of the associations.

Additionally, research contributions span other diverse fields, including “Mathematics” (n = 2), “Decision Sciences” (n = 1), “Multidisciplinary” (n = 1), and “Chemistry” (n = 1), further highlighting the interdisciplinary scope of the selected publications.

In terms of geographical trend, each publication has been linked to the countries affiliated with its authors. The study of 21 selected publications reveals contributions from multiple countries. A single paper can be linked to one or multiple countries, as each publication might have multiple authors, and each author can be affiliated with more than one country. However, if a paper has multiple authors from the same country, we only consider it once to avoid duplication.

At the point of analysis, 10 countries have contributed to at least one publication. Figure 4 below categorizes the publications into single-country papers and international collaboration papers, highlighting the involvement of different countries.

China is the most frequent among the contributing countries, participating in 8 publications from the selected studies. Afterward, the highest contribution is found from the authors from New Zealand and India, each contributing to 4 publications. The United States has been involved in 3 publications. Regarding international collaboration, contributions were found in this area of research from the United Kingdom, Iran, Australia, and New Zealand. Other contributing countries include Malaysia, Russia, Ghana, Spain, Bangladesh, South Korea, and Italy, all of whom have participated in at least one publication.

Below is the comparison between a country’s contributions vs. international collaboration: -

China is the most active in single-country and collaborative research, as signified by its leadership in this field from the selected studies. New Zealand and the United Kingdom are heavily engaged in international collaborations, reflecting their strong global research attitude and commitment to cross-border partnerships. India and China are leading in solely conducted research, showcasing their independent research contributions and advancements in the field. Notably, the United States appears only in collaborative research in our selected dataset. It suggests that the US primarily participates in multinational research projects rather than independent studies.

Healthcare professionals have used multiple types of medical data, including biomedical data such as electroencephalography (EEG), electromyographic signal (EMG), electrocardiogram (ECG), ultrasound, X-rays, computed tomography and magnetic resonance imaging (MRI), for a long time to judge patients’ diseases, diagnoses, and health conditions. Neuroimaging techniques are one of the various types for dealing with a wide range of conditions in the brain. MRI has recently been the most used technique, and it is said to be a vital technique for diagnosing brain tumors (Ahmadi et al., 2021; Sam et al., 2023) because it can generate images without damaging brain tissues (Dong et al., 2024). Furthermore, Kamal et al. (2023) used MRI to identify microbleeds in the brain, including the gene expression data, making it a multimodal analysis for determining the severity of Alzheimer’s disease.

In addition to traditional imaging techniques, other non-invasive methods such as functional magnetic resonance imaging (fMRI), diffusion tensor imaging (DTI), and electroencephalography (EEG) have been used for brain data collection in the recent past. These techniques played a dynamic role in helping us understand the functional and structural characteristics of the human brain in a better way (Sengupta et al., 2018). Capecci et al. (2015b) presented that EEG spatiotemporal data were used to study brain pathology and degeneration to analyze the functional changes in the brain activities of the controlled and Alzheimer’s disease (AD) groups. Furthermore, Kasabov N. K. et al. (2017) used fMRI in their research to develop a methodology using the NeuCube architecture of spiking neural network (SNN) for visualization, classification, and dynamic learning for spatiotemporal brain data. Guo et al. (2023) also applied fMRI like (Kasabov N. K. et al., 2017) for their study based on a fMRI based SNN for verifying anti-damage capabilities under random attacks.

In addition to the above-mentioned diseases, Sam et al. (2023) presented the use of EEG signal data for the quantitative assessment of depression levels by examining and categorizing EEG signals, and Francis and Al-Hababi (2023) utilized structural magnetic resonance images (sMRI) for early detection of suicidal ideation in youngsters and used depression as a biomarker in sMRI, respectively.

Most studies used a single or a few modalities to analyze different areas. However, combining predictive modeling with multimodal brain data has great potential; research in this area is still in its early stages due to a lack of advanced methods. The major challenge in combining brain data into a single model captured from various modalities is that each modality uses different temporal and spatial characteristics. To address this, the recent advancement in Spiking neural network systems (SSNs) makes it possible to incorporate multidimensional data within a single model (Sengupta et al., 2018).

Spiking neural networks are one of the most reliable techniques that computer simulations and computational models use to study the brain, e.g., brain-inspired machine learning. These techniques can reveal and learn frequency, time, and space information usually hidden in spatiotemporal brain data (STBD) Capecci et al. (2015a), Sengupta et al. (2018) presented an approach integrating multimodal information with a spiking neural network framework, creating a personalized SNNc-based architecture using the NeuCube. This experiment was run to represent the algorithm’s capabilities (oiSTDP) to capture discriminative join information from the data through its connection strengths. Utilizing this framework, the study showcases the integration of DTI and fMRI data of individuals who are beginning antipsychotic treatment to develop a personalized classifier for the prediction of treatment response in schizophrenia. Analysis with the SNNc network uncovered improved connectivity in the cerebellar region, which suggests that the captured activity of this area can be an aid as a potential biomarker that would help in treatment response in individuals with schizophrenia.

Kamal et al. (2023) stated that machine learning techniques are practical for analyzing Alzheimer’s disease datasets. Machine learning techniques can predict the disease by detecting microbleeds in the brain. The study used MRI images and gene expression data, making it multimodal for identifying microbleeds using SNN and decision trees. Afterward, pixel density analysis (PDA) was used to locate microbleed areas in MRI images. PDA and probabilistic graphical model (PGM) were also used to explain and decide on the diagnosis of microbleeds, and the severity of AD. Figure 5 represents the framework used in this research for identifying cerebral microbleeds and Alzheimer’s using multimodal data.

Furthermore, the study compared SNN with other state-of-the-art methods for performance evaluation using recall, accuracy, precision, and F-score. Table 3 below shows that SNN yielded the highest performance with 96.13%, 97.02%, 97.13%, and 96.84% for recall, accuracy, precision, and F-score, respectively.

Whereas Turkson et al. (2021) proposed a hybrid model, a spiking deep convolutional neural network architecture to classify MRI images to detect Alzheimer’s disease. Figure 6 shows the general architecture proposed in this study. The Alzheimer’s disease neuroimaging initiative (ADNI) dataset (450 scans) was used to perform three binary classification tasks (AD vs. NC, AD vs. MCI, and NC vs. MCI). The pre-processed images were used to extract AD key features using an unsupervised spiking neural network. The pre-trained spikes were then classified using a supervised deep convolutional neural network (CNN). The SNN was tested with two models, firstly with a model that was pre-trained with a spike, and secondly with a model without a pre-trained spike, and the results were more accurate with a pre-trained model for three binary classification tasks. The study stated that SNN improved performance, demonstrating the potential of spiking networks for reliable neuroimaging analysis.

Moreover, Doborjeh et al. (2021) used longitudinal neuroimaging data to utilize the advancements in deep learning models through brain-inspired spiking neural networks (SNNs), which can model structural brain data across time and space. The study introduced a methodology and an SNN-based computational framework for developing personalized, predictive models from longitudinal brain data to detect, interpret, and predict changes in an individual’s functional brain state. The approach consists of several steps, such as clustering similar data, incorporating missing values, and training a 3D brain-template SNN for classifying and predicting outcomes and visualizing structural brain changes. All this information is then used to interpret results and identify predictive markers at both individual and group levels. As a result, the model successfully classified and predicted cognitive decline, for example, dementia and mild cognitive impairment (MCI), 2 years ahead with 91% and 95% accuracy.

Previous sections primarily focused on the types of modalities used by several researchers, along with the integration of machine learning models using SNN to analyze neuroimaging data (Doborjeh et al., 2019; Saeedinia et al., 2021; Ghazali et al., 2020). Moreover, multiple studies highlight practical applications using neuroimaging and SNNs for several diseases, such as the detection of AD and the prediction of schizophrenia. However, neuroimaging, multimodal data analysis, and integration of deep learning with spiking neural networks are not only vital for AD but also for many other clinical areas, such as brain tumor classification (Kalpana et al., 2024) or segmentation (Ahmadi et al., 2021), suicide ideation assessment (Francis and Al-Hababi, 2023), Depression identification (Sam et al., 2023), Parkinson’s disease detection (Das et al., 2024), and so on.

Kalpana et al. (2024) proposed a hybrid model integrating SNNs with convolutional neural networks (CNNs) for brain tumor classification. As per the results, the model achieved a testing accuracy of 97.50 % and a training accuracy of 97.79% with a slight loss value of 0.38%. Also stated, these results present the model’s strength in capturing complex data in medical imaging, which is promising for clinical applications within the biomedical field. On the other hand, Ahmadi et al. (2021) utilized MRI for tumor area segmentation using a deep spiking neural network. The process for achieving the results consists of preprocessing and segmentation using DSNN. Table 4 shows the accuracy percentages compared in this study for the proposed method and previous approaches.

Furthermore, Francis and Al-Hababi (2023) also incorporated a hybrid version like Kalpana et al. (2024). However, they combined an attention mechanism (AM) with SNN for suicide ideation assessment using sMRI of healthy controls and depressive individuals who did not show any suicide ideation (SI). The hybrid model completed the classification tasks using stratified 5-fold cross-validation and achieved a test accuracy of 94%, sensitivity of 100%, specificity of 92%, and an area under the curve (AUC) of 0.96. The proposed algorithm provides an objective tool that can support clinical assessments in identifying early signs of SI risk among depressed patients who are currently not showing any suicidal thoughts.

Sam et al. (2023) performed a similar study to Francis and Al-Hababi (2023) by combining long short-term memory (LSTM) and SNN using EGG signals for depression identification. These models were used for the first time to analyze and categorize EEG signals according to the level of depression (minimum, mild, moderate, and severe). The model utilized spatial data mapping using SNN, leading to unsupervised learning and visualization of spiking patterns and unique perceptions about brain mechanisms. Comparative analysis of time-space brain data showed that SNN is more advantageous than other deep learning models. The study achieved higher accuracy in classifying samples from various groups and disclosed different patterns of brain activity, helping to understand the severity of depression. The findings in the study have the potential for early prediction and, thus, preventing depression by using the brain data acquired from different depression levels. The methodology used in this study showcases excellent potential for application in neuroimaging and clinical longitudinal data.

Apart from the above-mentioned clinical applications, SNN and neuroimaging are also helpful for Parkinson’s disease (PD) detection, as presented by Das et al. (2024). This study investigated the SEFRON (time-varying synaptic efficacy function based leaky integrate and fire neuron model) model and compared it with other previously used neural network models such as RBF, RNN, LSTM, and MLP. Results showed a higher accuracy for this model over others, making it reliable for clinical trials. Moreover, because of its performance, this model can help develop an automated PD detection device that physicians can utilize to diagnose PD in its early stages. However, this study has a limitation in that the sample size of the dataset used is small.

Another study proposed by Gowsikraja et al. (2025) uses MRI for AD classification using a model named SBERO_Deep SNN (skill Al-Biruni Earth Radius Optimization-enabled Deep Spiking Neural Network), encompassing the benefits of SNN. Segmentation was performed using a hybrid algorithm using UNeXt, combining the Skill Optimization Algorithm (SOA) and the Al-Biruni Earth Radius (BER). In the next step, statistical features were extracted and classified using a Deep SNN trained with SBERO. The study achieved the highest accuracy rate compared to other AD classification techniques. It also states that the proposed methodology is reliable for AD identification, allowing timely intervention to slow the disease progression and improve patient quality of life. Table 5 represents the overview of selected studies below.