The application of functional magnetic resonance imaging (fMRI) in adolescent health risk assessment represents a cutting-edge advancement in this field. By detecting blood oxygen level-dependent (BOLD) signals, fMRI indirectly reflects brain activity, offering rich spatial and temporal information (Stiernman et al., 2021). This technology is extensively used to study neurodevelopmental processes in adolescents and to identify brain function characteristics related to mental health issues, such as depression, anxiety, and attention deficit hyperactivity disorder (ADHD) (Wang et al., 2023; McNorgan et al., 2020). For example, fMRI can help pinpoint brain activity patterns associated with these common adolescent mental health challenges.

However, despite the non-invasive nature and high spatial resolution of fMRI, there are certain limitations in its practical application. First, the high cost of fMRI data acquisition and analysis restricts its use in large-scale studies. Second, fMRI is sensitive to noise and individual differences, which necessitates careful interpretation of the results (Uyulan et al., 2023). Additionally, the fMRI scanning process may cause discomfort in some participants, especially adolescents, potentially affecting the accuracy of the data. Therefore, improving data quality while reducing participant discomfort, as well as simplifying the complexity of fMRI data processing, remain significant challenges in this area.

With the advancement of deep learning, researchers have increasingly integrated it with fMRI data to enhance the precision and efficiency of health risk assessment (Liu et al., 2022). Deep learning models, particularly convolutional neural networks (CNNs) and Long Short-Term Memory network (LSTM) (Saurabh and Gupta, 2024), have shown immense potential in processing and analyzing fMRI data. CNNs, with their hierarchical structure, can effectively extract complex spatial features from fMRI images (Chen et al., 2020), while AEs use unsupervised learning to achieve data dimensionality reduction and reconstruction, thereby alleviating the computational burden associated with high-dimensional data.

The application of deep learning in fMRI data analysis offers several significant advantages. For instance, CNNs can automatically extract brain activity features related to adolescent mental health risks without relying on traditional manual feature selection methods. This not only improves the efficiency of feature extraction but also captures a wider range of potential brain function patterns. Additionally, AEs excel in denoising and feature selection, making the models more robust and stable when handling fMRI data.

However, the application of deep learning in network magnetic resonance imaging also faces challenges. First, deep learning models typically require large amounts of labeled data for training, and the high cost of acquiring and labeling fMRI data limits the scale of training datasets. Second, the “black box” nature of deep learning models makes them difficult to interpret, which is particularly important in medical applications (Iravani et al., 2021). Moreover, issues such as overfitting and high model complexity may lead to suboptimal performance in practical applications. Therefore, balancing model complexity and interpretability, and training efficient models on small sample datasets, remain key areas of focus in this field.

Adolescent health risk assessment criteria are a critical application area for combining fMRI techniques with deep learning (Ernst et al., 2015). Existing health risk assessment standards are typically based on a variety of factors, including biomarkers, behavioral assessments, and psychological questionnaires, providing essential tools for identifying at-risk adolescents (Bjork et al., 2010). However, traditional assessment methods often rely on expert judgment, which can introduce bias and inconsistency.

In recent years, researchers have sought to develop health risk assessment criteria based on fMRI data and deep learning models. For example, some studies have utilized deep learning models to automatically analyze fMRI data, extracting features related to health risks and predicting individual mental health risks based on these features (Mueller et al., 2010). This approach not only improves the objectivity and accuracy of assessments but also enables early detection of potential health issues, providing critical information for intervention and treatment.

The integration of fMRI and deep learning technologies provides powerful tools for health risk prediction. Despite the significant potential demonstrated by current research and applications, challenges remain in terms of data acquisition, model interpretability, and the standardization of assessment criteria. Future research should focus on developing more efficient, accurate, and scalable health risk assessment models to better serve the health management needs of adolescent populations.

An autoencoder (AE) is a neural network model primarily used for unsupervised learning. It achieves dimensionality reduction and feature extraction by learning to encode input data. The basic structure consists of two parts: the encoder and the decoder. The encoder maps the input data to a hidden representation, while the decoder attempts to reconstruct the original input from this hidden representation. By minimizing the error between the input data and the reconstructed data, the autoencoder learns useful features of the data.

We utilize a convolutional autoencoder model with the proposed number of layers to extract features from the Adolescent Health Risk Behaviors (AHRB) dataset (Demidenko M. I. et al., 2024). Initially, we preprocessed the raw data using the PyReliMRI toolkit (Demidenko M. et al., 2024), which includes standardized steps such as head motion correction, slice timing correction, and spatial normalization.

The goal of the autoencoder is to encode the input x into a hidden representation h, and then reconstruct the input x from h. The encoding and decoding processes are shown in Equations 1, 2, respectively:

where W₁ and W₂ are the weight matrices of the encoder and decoder, respectively, b₁ and b₂ are the bias terms, and f and g are activation functions (typically nonlinear functions such as ReLU).

The loss function for the reconstruction error is typically the mean squared error (MSE), as shown in Equation 3:

In the analysis of functional magnetic resonance imaging (fMRI) data, convolutional autoencoders can be used to extract spatial and temporal features for assessing adolescent health risks. The convolutional autoencoder comprises an encoder and a decoder. The encoder consists of three convolutional layers, each followed by a ReLU activation function and a max-pooling layer. The decoder consists of four convolutional layers, the first three layers are the upper sampling layer, and the last layer generates an image with the same shape as the input, as shown in Figure 2 and Table 1 for details.

For the input connectivity matrix data, features are first extracted through convolutional layers. The final layer of the encoder provides the hidden representation, as shown in Equation 4:

The hidden representation h is fed into the decoder, which reconstructs the input data through convolution and upsampling layers, as shown in Equation 5:

Finally, the model optimizes its parameters by minimizing the reconstruction error, thereby learning useful features from the fMRI data for subsequent adolescent health risk assessment.

We map the voxels in three-dimensional spatial coordinates to specific brain functional regions using a brain partition map. The voxel values within each brain region are averaged to represent the overall fluctuation of blood oxygen concentration levels in that region, focusing our research on regions of interest (ROIs). This process is illustrated in Figure 3.

The state sequence mapping process converts the preprocessed numerical sequence data into state sequences. This study uses a dynamic threshold set by the rule of thumb (Figure 4), marking data above the threshold as active (1) and below as inactive (0). The threshold is defined as follows:

where μ(Hnk) denotes the mean and σ(Hnk) denotes the standard deviation of the time series Hnk, see Equations 7, 8 for details.

We define the mapping function f(mn,lk,η) to map each time slice η and mn,mk to a state u:

We extract discrete probability distributions of synchronous expression between brain regions. First, we define a function ϕ(·) to calculate the dynamic temporal relationship between two time series. The state sequence mapping step has successfully converted the fMRI data from a numerical sequence to a state sequence. Next, we use a measure function ψ(·) to evaluate the degree of synchronous expression between two time series, defined as in Equation 10:

where f(·) represents the mapping function under certain prior conditions, converting data from a numerical sequence to a state sequence, and η^k denotes the mapping parameter. We then statistically analyze the frequency information of synchronous expression between brain regions. The function ψ(·) calculates the concurrent activation of two brain regions given a time slice parameter. This concurrent activation could be coincidental or indicative of potential interactions between these regions. With a sufficient number of time slices, we can obtain the probability distribution of whole-brain synchronous activity, distinguishing between coincidental and genuinely related phenomena, as shown in Equation 11:

For a multivariate time series data H_k, the interaction between any two time series Hik and Hjk is defined as follows in Equation 12:

where X_ϕ(·)(i, j) represents the cumulative sum of function values for each time slice pair. Based on this, the interaction between any two time series Hik and Hjk is further defined with a time interval parameter It = [rt, st] as shown in Equations 13, 14:

This computes the interaction of custom asynchronous intervals with active-passive relationships. Note that Xϕ(·)k(i,j,Im)≠Xϕ(·)k(j,i,Im), indicating that the resulting interaction matrix X_ϕ(·) is asymmetric, as the interactions have active-passive relationships.

The comprehensive multivariate time series interactions, Xϕ(·)k∈RN×N×T, represent multi-scale interval asynchronous synchronous expression values. Here, Xϕ(·)k is a third-order tensor, where N denotes the number of time series, and T denotes the number of time slices of any time series. We convert the tensor Xϕ(·)k into a discrete probability distribution form Qϕ(·)k, defined as follows in Equation 15:

where qϕ(·)k(i,j,Im) denotes the discrete probability value between the i-th and j-th time series under interval parameter I_m and mapping function ϕ(·), defined as follows in Equation 16:

Finally, we use the extracted multi-scale asynchronous correlation information as features to train the model, resulting in an adolescent health risk prediction model.

The Adolescent Risk Behavior (AHRB) study dataset (Demidenko M. I. et al., 2024) recruited participants from diverse backgrounds to ensure a representative sample of the adolescent population. Each participant underwent a series of assessments, including neuroimaging, behavioral tests, and self-reported questionnaires. The primary focus of the study is to capture the dynamic changes in behavior and brain function as participants transition from late adolescence to early adulthood. The dataset includes two main cohorts from different years: Year 1 consists of approximately 108 participants aged 17–20, and Year 2 consists of approximately 66 participants aged 19–22. This study aims to track the developmental trajectory of risk behaviors and their underlying neural mechanisms.

The functional magnetic resonance imaging (fMRI) component of the AHRB study includes tasks designed to probe emotional and reward processing. Specifically, the study utilizes the Emotional Faces task and the Monetary Incentive Delay (MID) task. For our analysis, we use the raw Blood Oxygen Level-Dependent (BOLD) data from the MID task, which aligns with similar tasks used in the MLS and ABCD studies. The MID task requires participants to respond to cues indicating potential monetary rewards or losses. During the anticipatory phase, participants receive cues that indicate whether they can win or lose money based on their performance. The BOLD response during this phase is particularly interesting as it reflects the neural processes involved in anticipation, a critical component of reward-based decision-making. Understanding the neural basis of reward processing is crucial, as it is a key aspect of adolescent risk behavior.

The evaluation metrics used in this study include accuracy (Acc), Precision (Prec), Recall (Rec), and F1−score. These metrics are defined as follows in Equations 17–20:

Here, True Positive (TP) represents the correctly classified positive samples, True Negative (TN) represents the correctly classified negative samples, False Positive (FP) represents the incorrectly classified positive samples, and False Negative (FN) represents the incorrectly classified negative samples.

We first consider the threshold parameter η in Equation 6, which distinguishes the active or inhibitory states of the brain in fMRI imaging. A dynamic threshold converts a numerical sequence into a 0/1 sequence, where larger η values make the active state determination more stringent, resulting in fewer data points, while smaller values capture more data but may introduce noise. For all experiments, we set η = 1, as it balances avoiding overfitting while maintaining sufficient data points.

To improve model generalization given the limited data, we reduced the complexity of the 2DCNN-AE model by decreasing the number of layers and parameters in the convolutional layers. Additionally, we applied L2 regularization with a weight decay of 0.01 and used a Dropout rate of 0.5 in the fully connected layers to further prevent overfitting.

For model evaluation, we conducted a 10-fold cross-validation. The dataset was divided into 10 subsets, with the model trained on 9 subsets and validated on the remaining subset. This process was repeated 10 times, and the final performance metrics were averaged across all folds to ensure robust assessment of the model's stability and generalization. The experimental Settings are shown in Table 2.

The primary goal of this study is to explore asynchronous functional connectivity between different regions of the adolescent brain. We extracted discrete probability distributions of synchronous expression at varying time intervals between brain regions. The interval set I was defined as [0, 0], [1, 1], [2, 2], [3, 6], [7, 12], with smaller intervals capturing short time delay interactions and larger intervals representing longer delays. This setup helps balance the sensitivity to synchronous information while minimizing the risk of overfitting due to excessively large delay parameters. The pseudocode for our algorithm is shown in Algorithm 1:

First, the experiment utilized a two-sample t-test as a feature selection method for dimensionality reduction. Next, the significance level parameter p-value was set to 0.05, 0.01, 0.005, and 0.001 respectively, with the results shown in Table 3.

The results show that as the significance level parameter decreases, the number of extracted features also decreases. Initially, the experimental results improve with fewer features, but when the number of features is reduced too much (e.g., p = 0.001), the performance drops sharply. The best classification results were obtained with a significance level parameter of p = 0.005, achieving an accuracy, precision, recall, and F1-score of 70.142%, 66.276%, 71.946%, and 68.995%, respectively.

After selecting features with a significance level of p = 0.005, we further validated the parameter for the dynamic threshold μ in the state sequence transition. The experimental parameter μ was tuned within the range [0, 2] via grid search, with the results shown in Table 4.

As shown in Table 4, using the full temporal mean value as the high activity threshold (μ = 0) yields poor results. This might be due to the low threshold being too broad, defining half of the time points as active, which introduces a lot of noise. As the threshold increases, the results improve. The best classification accuracy is achieved at μ = 1.0 and μ = 1.2, as a tighter definition of “active state” can effectively distinguish important activities. Beyond μ>1.2, the performance declines as the threshold becomes too high, leaving few points defined as active. The optimal default choice for the dynamic threshold parameter μ is 1.0. The experimental results for different parameters p are shown in Figure 5.

Next, we compare our model with methods using Pearson correlation coefficient, higher-order statistics (Wee et al., 2016), and dynamic functional connectivity models (Harlalka et al., 2019). Other popular methods include those by Zhang and Wang (2020), Brown et al. (2019), Abraham et al. (2017), and Yang et al. (2020). The results on the AHRB dataset are summarized in Table 5.

As shown in Table 5, our proposed method achieves precision, recall, and F1-score results of 83.116%, 84.784%, and 83.942%, respectively. Compared to the aforementioned methods, our method ranks highly in precision, recall, and F1-score. Although the accuracy is slightly lower than the method by Abraham et al. (2017) (83.366%), our method excels in the other three metrics. The experimental results demonstrate that our approach, incorporating 2D convolutional autoencoders and multi-sequence, multi-scale asynchronous information extraction, uncovers more asynchronous correlation information, yielding good classification accuracy in adolescent health risk assessment applications.

To validate our proposed improvements, we conducted three groups of ablation experiments. We added complex convolution and channel attention mechanisms to the 2DCNN network, along with phase smoothness and coil sensitivity smoothness as physical priors. The ablation experiments are summarized in Table 6, where coil sensitivity smoothness is denoted as S, phase smoothness as P, complex convolution as C, and channel attention mechanism as A. The baseline model is 2DCNN.

As shown in Table 6, each proposed improvement brought about performance enhancements. Individually adding coil sensitivity prior and phase prior led to considerable improvements, while the combination of channel attention mechanism and complex convolution resulted in significant gains. Combining all improvements achieved the best results.

These results indicate that the introduced 2D convolutional autoencoder and multi-sequence, multi-scale asynchronous information extraction methods effectively capture asynchronous correlation information, enhancing model performance in adolescent health risk assessment applications. The proposed modifications lead to significant improvements, as evidenced by the comprehensive ablation studies.

In summary, our method demonstrates superior performance in most metrics compared to existing methods, highlighting its potential in adolescent health risk assessment based on rs-fMRI data.