Raksha Rajanna*
| Mullur Puttubuddhi Pushpalatha | Sunitha Rajanna | Impana Kamalamma Puttaraju
© 2026 The authors. This article is published by IIETA and is licensed under the CC BY 4.0 license (http://creativecommons.org/licenses/by/4.0/).
OPEN ACCESS
Depression among older adults is a growing public health concern, while early identification remains challenging due to the complex interactions between clinical, behavioral, and demographic factors. This study proposes an explainable Attention-based Long Short-Term Memory framework optimized by Particle Swarm Optimization (PSO-ALSTM) for early depression risk prediction using multimodal health data. The proposed framework integrates Attention-LSTM to capture temporal dependencies within longitudinal health records and employs PSO to optimize critical model parameters, including learning rate, hidden units, batch size, and dropout rate. In addition, SHapely Additive exPlanations (SHAP) analysis is incorporated to improve model interpretability by identifying the contribution of individual features to prediction outcomes. The framework is evaluated using multiple datasets, including National Health and Nutrition Examination Survey (NHANES), the Community Cognitive Health Dataset, the National Poll on Healthy Aging (NPHA), and a real-world dataset collected from 300 individuals. Experimental results demonstrate that the proposed PSO-ALSTM model achieved an accuracy of 92.8% and an Area Under the Curve (AUC) of 0.92, outperforming conventional machine learning algorithms and baseline deep learning models. The findings indicate that combining temporal deep learning, bio-inspired optimization, and explainable artificial intelligence will provide an effective approach for supporting early depression risk assessment in healthcare applications.
explainable artificial intelligence, Attention- Long Short-Term Memory, Particle Swarm Optimization, depression prediction, multimodal health data, longitudinal health analysis
The soaring global geriatric population has caused a corresponding increase in the number of older adults suffering from mental health disorders, including depression, anxiety, and dementia. Unfortunately, these conditions tend to be underdiagnosed because of the slow progression of symptoms, the presence of other medical conditions that complicate diagnosis, and difficulty obtaining ongoing care. Research has shown that mental health disorders have detrimental effects on quality of life, cognitive function, and mortality among older adults [1, 2]. In addition, the already strained healthcare system and lack of geriatric specialists point to the necessity for scalable and proactive means for the early detection and intervention of depression and other forms of mental illness in older adults [3].
Artificial intelligence (AI) is revolutionizing healthcare by enabling the prediction and monitoring of mental health. Some of the ways in which AI is being utilized in mental health include machine learning, natural language processing, and deep learning to analyze and understand large, heterogeneous datasets that include electronic medical records, voice recordings, and data captured by wearable devices [4, 5]. Current studies support the early identification of mental health disorders as well as predicting the progression of disease and providing personalized interventions through the utilization of various AI-based applications in aged populations [6, 7]. Furthermore, the ongoing monitoring of physical and behavioral data through AI-driven systems enables more accurate and timely assessment of mental health responses than traditional assessment methods [8].
The current body of literature has shown many promising advancements; however, several limitations remain. The outcomes of many studies lack generalizability across different populations of older adults due to the use of small or homogeneous datasets [9]. Several studies use a classification approach instead of focusing on early forecasting of mental health conditions. The majority of studies have little to no integration of multiple data sources (e.g., behavioral, clinical, and physiological), which are needed for accurate prediction. Additionally, a lack of explainability in AI models decreases trust in these models and, therefore, their adoption in real-life healthcare environments [10]. Finally, ongoing challenges related to data privacy, bias, and model reliability must be addressed as part of the ethical concerns surrounding AI-based mental health systems.
The proposed research will address the limitations discussed above with the following contributions:
• Designed a novel approach to predicting mental health conditions in older adults that uses Particle Swarm Optimization (PSO), attention mechanisms, and Long Short-Term Memory (LSTM) models in a combined manner to enhance the accuracy of predictions as well as abstract features that occur over time.
• Integrating the Attention-LSTM architecture captures time-dependent clinical and behavioral patterns from older people’s healthcare records and improves the ability to identify key indicators related to mental health.
• Utilized bio-inspired optimization through PSO to optimize hyperparameters (learning rates, number of hidden neurons, batch size, dropout rates) and enhance convergence while limiting overfitting.
• The dataset used in this research study consists of four separate databases, including the National Health and Nutrition Examination Survey (NHANES), the Community Cognitive Health Study, the National Poll on Healthy Aging (NPHA), and a real-time dataset consisting of 300 people's data from rural and urban areas.
• Through significant testing and comparison with both traditional machine learning algorithms and baseline deep learning algorithms, it is demonstrated that the Particle Swarm Optimized Attention Long Short-Term Memory (PSO-ALSTM) outperformed all of these models based on various performance metrics.
The rest of this study is structured as follows: In Section 2, we present prior research on the use of advanced technologies to predict the mental well-being of individuals. Section 3 describes how we propose to collect, preprocess, extract features, integrate disparate sources, and establish our predictive framework. Section 4 covers the experimental design, performance evaluation criteria, and a comparative evaluation of our proposed method against existing methods. Finally, Section 5 summarizes the most important contributions, limitations, and avenues for future research.
In recent years, a growing number of scholars have focused on using machine learning and deep learning methods to predict adverse mental health states such as depression, anxiety, loneliness, cognitive decline, and other related conditions in older adults. These technologies have shown promise in detecting psychological risks from multiple types of data, including clinical, behavioral, physiological, imaging, and speech-based data. However, there are also major gaps in the literature regarding multimodal data integration, explainability, applicability in real-world settings, and the personalization of predictions for older adults. Castillo et al. [11] used prompts for dynamic conversations. The authors provided evidence that prompting could be used to find risk factors for depression among older individuals in low-resource settings.
Jia et al. [12] proposed a hybrid machine learning framework for estimating depression among homebound older adults in a 7-year longitudinal study. Their study highlights the importance of long-term tracking of health changes. Ahmad et al. [13] designed prediction models to identify the risk of depression among low-income seniors living in households, using Decision Trees (DTs), Logistic Regression (LR), neural networks (NNs), and Random Forest (RF). While these studies demonstrated acceptable predictive performance, they primarily relied on datasets derived from pre-structured questionnaires and socioeconomic measures. Limited integration of multiple types of health data that can help monitor depression in a more reliable and timely manner is a major area where researchers need further information.
Recently, deep learning models have been developed to improve prediction accuracy and enhance the ability to interpret predictions. For example, Jain and Bhakta [14] designed a depression risk prediction model using a CNN-BiLSTM-Attention framework, as well as an LSTM + SHapely Additive exPlanations (SHAP) model, using AI tools to identify feature contributions in their studied population. Similarly, Lukac [15] developed an explainable anxiety risk predictive model for older adults with abdominal obesity using machine learning and SHAP analysis. While prevailing studies have increased model transparency to date, their development has focused mostly on sampling individuals from one disease/condition at a time and from a single region of residence.
New technologies for detecting late-life depression are emerging, including advanced neuroimaging methods that use brain scans and brain wave signals to assess mental health status. In one study conducted by Han et al. [16], they demonstrated the ability to automatically diagnose late-life depression using resting-state functional Magnetic Resonance Imaging (fMRI) and 3D convolutional neural networks (3D CNNs). Unfortunately, most neuroimaging techniques require expensive equipment, specialized clinical environments, and high levels of computational power. Other researchers have begun to examine the social and behavioral aspects of the mental health of populations. Xu et al. [17], for example, used citizen science and machine learning methods to detect loneliness in older adults. This research highlighted the role of participatory and social awareness activities in mental health analytics. However, the majority of current methodologies for detecting loneliness do not account for physiological, emotional, and behavioral components, nor do they enable continuous monitoring of dynamic changes in older adults' emotional states.
The area of mental health analysis through speech and language has seen a significant increase in research activity. For instance, Huang et al. [18] performed linguistic and automatic classifier analyses of language use to identify mild cognitive impairment and dementia, while Nykoniuk et al. [19] used machine-learning analyses of speech features to identify individuals with dementia at risk of developing Alzheimer's disease. Tan et al. [20] developed an adversarial learning process to explore how speech pauses relate to dementia. Zhang et al. [21] and Zhou et al. [22] have conducted studies to create an emotional audio-textual corpus for predicting depression using speech features. The aforementioned work shows that many speech-based systems could provide a less invasive and more manageable way to identify at-risk individuals. The majority of these systems rely on limited training datasets, which could be improved by incorporating additional emotional and linguistic adaptations. In addition, cross-cultural speech differences and the multilingual population represent areas for further exploration in future research.
Advances in wearable sensing and passive monitoring technologies have greatly improved mental healthcare options for seniors. Bahadi et al. [23] evaluated depression assessment using wearables and sensors, while Mamidisetti and Reddy [24] evaluated wearable, environmental, and passive smartphone sensing technologies for mental health assessment. Tveter et al. [25] used sensor data to evaluate the potential of unsupervised machine learning to identify physiological stress levels. Collectively, these studies indicate that passive and longitudinal monitoring systems are becoming increasingly available for mental health assessment. However, many wearable technologies continue to face difficulties with data privacy, sensor reliability, energy efficiency, handling missing data, user acceptance, and other challenges. Furthermore, very few systems assess multiple forms of mental health; rather than providing a complete and integrated assessment framework of both cognitive and emotional function.
Partial research has also been done evaluating more general analytics related to seniors and healthcare data using interpretable machine learning, along with analysis of various biometric markers. Xie et al. [26] discovered biomarkers from blood tests to help determine if someone is older than 65, and whether they may be at risk of developing Alzheimer’s disease. In a study, Haraldsen et al. [27] used analytics and machine learning to better understand the factors that create barriers to outdoor physical activity for older adults. Although progress has been made in developing advanced machine-based technologies, there is still a need for all-inclusive models capable of combining multiple data sources and creating an integrated/understandable forecasting methodology for mental health status.
There has been considerable success in applying technologies such as machine learning, deep learning, wearables, neuroimaging, and speech analysis to predict mental health; however, significant knowledge gaps do still exist related to multimodal data fusion, explainability, longitudinal monitoring, cross-cultural validity, real-time analytics, and personalized solutions for consumers. Thus, there is a pressing requirement for the development of an integrated and explainable framework/model that considers physiological, behavioral, cognitive, and emotional components of health when assessing and diagnosing patients for depression and related mental health issues in the population, specifically aiding in both early detection of these issues as well as ongoing assessment.
This research presents a novel PSO-ALSTM framework for predicting and understanding mental health conditions among seniors by analyzing long-term health records. First, the Attention-LSTM will capture sequential dependencies among variations in individual patient data over time by processing each individual’s longitudinal data through the LSTM network. The attention mechanism produces adaptive weights for different time points in a patient’s longitudinal record, enabling the Attention-LSTM to identify the most important clinical and behavioral factors that drive significant variations in mental health outcomes. Second, the PSO optimization algorithm will address issues associated with manual hyperparameter tuning and enhance the convergence of the Attention-LSTM optimization process.
Third, the PSO optimization algorithm will optimize the relevant parameters of the Attention-LSTM model (i.e., learning rate, number of hidden units or neurons per layer, and dropout rate), where each iteration of the PSO will update the best-performing candidate from each of the best performers as well as from the global best-performing candidate from the entire population of candidate solutions during the optimization process. Once the optimal parameters have been evaluated, the optimized Attention-LSTM model will produce a contextually relevant feature representation for classification purposes. Finally, the classification of the mental health risks associated with the optimized Attention-LSTM model will be accomplished with the aid of SHAP for determining the contributions of each factor to the final outcome of the model. Collectively, the proposed workflow will improve the predictive power, robustness, and transparency of forecasting mental health conditions among older adults, thereby providing an opportunity for earlier and more accurate forecasting of mental health-related issues in older adult populations. An entire architecture of the proposed PSO-ALSTM framework with all these stages is shown in detail in Figure 1.
Figure 1. Proposed Particle Swarm Optimized Attention Long Short-Term Memory (PSO-ALSTM) model design diagram
This research outlines the development of a PSO-ALSTM framework for forecasting mental health conditions in the population. The proposed framework combines PSO, Attention-LSTM, and SHAP. The aim of the development of this framework is to create a model that increases the predictive power and efficiency of convergence with respect to explanations of how and why predictions are made.
3.1 Dataset representation
The dataset used to construct and evaluate the proposed model is defined in Eq. (1).
$D=\left\{\left(X_i, y_i\right)\right\}_{i=1}^N$ (1)
where, D is a collection of N samples and each sample in D represents an individual person. The feature matrix of the element i of D is represented by Xi, which contains all of the longitudinal health data for the person indexed by i. The mental health condition associated with person indexed by i is represented by yiand the indices of sample i are: i = 1, 2, 3, ..., N, where N is the total number of samples in the dataset. Using this structured and consistent representation of samples of Das, it allows for the creation of a supervised learning model that will create a mapping from the complex temporal patterns from health data to mental health conditions.
3.2 Feature space representation
The structure and dimensionality of input features and output labels are described in Eq. (2).
$X_i \in \mathbb{R}^{T \times F}, y_i \in\{0,1\}$ (2)
where, model uses a matrix with T rows and F columns to create an input X, which represents the temporal observations (T) and the characteristics being measured (F). The characteristics represented as R will contain continuous real numbers. The label associated with the input is typically binary y.
The input $X_i$ is modeled as a real-valued matrix of size $T \times F$, where $T$ represents the number of temporal observations (e.g., clinical visits, daily records), and $F$ denotes the number of features such as physiological measurements, behavioral indicators, or demographic attributes. The notation $\mathbb{R}$. indicates that the features take continuous real values. The label $y_i$ is binary, where 0 indicates the absence and 1 indicates the presence of a mental health condition (e.g., depression). This formulation explicitly captures the timedependent nature of health data.
The objective of the learning process is to predict mental health conditions based on input characteristics is defined in Eq. (3).
$\hat{y}=f\left(X ; \theta^*\right)$ (3)
where, f represents the non-linear mapping defined by the proposed Attention-LSTM model, $\widehat{y}$ is the predicted probability/class label, $\theta^*$ represents the optimal parameters generated by the learning process. These parameters are optimized using PSO leads to improved convergence and predictive performance.
3.3 Attention-based Long Short-Term Memory model
The preceding passage refers to two equations that provide a method for establishing the proportion of information retained versus discarded in memory, defined by Eq. (4).
$f_t=\sigma\left(W_f\left[h_{t-1}, x_t\right]+b_f\right)$ (4)
where, $f_t$ is the forget gate takes as input two components, $h_{t-1}$ is the last hidden state and the current input $x_t, W_f$ is the weights, $b_f$ is the bias, and the activation function $\theta(\cdot)$ acts to map the outputs into the range of 0 to 1 .
The function of this equation is also to determine the amount of new information to be added to the memory is defined in Eq. (5).
$i_t=\sigma\left(W_i\left[h_{t-1}, x_t\right]+b_i\right)$ (5)
where, the input-gate $i_t$ controls which new information flows into the cell state, $h_{t-1}$ is the last hidden state and the current input $x_t, W_f$ is the weights, $b_f$ is the bias. The sigmoidal activation of the input gate determines how much new information will become part of $C_t$. Using this selection mechanism allows the model to effectively learn and integrate relevant temporal patterns adaptively. The result of this equation is the generation of new candidates that can be inserted into the cell-state is defined as Eq. (6).
$\tilde{C}_t=\tan \mathrm{h}\left(W_C\left[h_{t-1}, x_t\right]+b_C\right)$ (6)
where, the LSTM allows for the inclusion of new possibilities, based on the new input and the previous hidden state, in the form of the candidate memory $\tilde{C}_t$. The tanh function will transform these values to be between −1 and 1, which will allow for the possibility to add both negative and positive contributions.
The combination of candidate (new) and retained (old) information in the memory of the LSTM is represented as Eq. (7).
$C_t=f_t \odot C_{t-1}+i_t \odot \tilde{C}_t$ (7)
where, a combination of the two previous components will occur: the retained cell memory, $f_t \odot C_{t-1}$, and the new candidate information, $i_t \odot \tilde{C}_t, \odot$ indicates element-wise multiplication. This allows us to have an LSTM that is able to hold information for many previous time periods while still adapting to new inputs.
The output of the LSTM will be represented as a combination of both retained (short-term) information and new (long-term) information, with respect to timing is defined in Eq. (8).
$h_t=o_t \odot \tanh \left(C_t\right)$ (8)
where, $O_t$ is computed in a similar fashion to the other gates; tanh function keeps the output value relatively constrained. The outputs from the internal memory become available through the output gate according to their degree of filtering. The final hidden state $h_t$ will then be used as the features passed on to the next computational layer.
3.4 Attention mechanism
The Attention Score Function, which calculates how important each time step is when generating the sequence, is defined as Eq. (9).
$e_t=v^T \tan \mathrm{~h}\left(W_h h_t+b_h\right)$ (9)
where, the attention score $e_t$ shows how relevant to the output at time step $t$. To transform $h_t$ to $e_t$ uses three learnable parameters: $W_h$ (weights), $b_h$ (biases), and $v$ (weights) to change $h_t$ into a scalar $e_t$. This process allows for model evaluation of which of the time steps contributed most to the prediction.
The score to be converted to a probabilistic framework: each attention score can be expressed as a probability reflecting the relative importance of that time step is defined as Eq. (10).
$\alpha_t=\frac{\exp \left(e_t\right)}{\sum_{k=1}^T \exp \left(e_k\right)}$ (10)
where, $\alpha_t$ is the attention weight for the current time step t is accomplished via a softmax function that guarantees that the sum of all attention weights $\sum \alpha_t=1$ across time steps. The performance of the Softmax will amplify the differences between attention scores via the exponential function exp(). Therefore, this allows the model to put significantly more emphasis on the important time steps than on the other time steps.
One representation of the weighted temporal features (generated via the respective attention weights) for the purpose of prediction is defined as Eq. (11).
$Z=\sum_{t=1}^T \alpha_t h_t$ (11)
where, $Z$ is the context vector at the current time step, $h_t$ is the hidden states and their associated attention weights $\alpha_t$ at the current time step. The resulting context vector Z provides a compact representation of the important temporal patterns that will help explain the mental health condition.
3.5 Particle Swarm Optimization
To find an optimal solution for the hyper parameters of Attention-LSTM model, we have used PSO. The PSO will automatically find an optimal combination of hyper parameters, such as the learning rate, the number of hidden units, the batch size, and the dropout rate, thereby speeding up convergence and improving predictive performance without the need for user intervention.
Each particle in the swarm represents a potential solution within the search space is defined as Eq. (12).
$\mathrm{x}_i=\left[\theta_1, \theta_2, \ldots, \theta_d\right]$ (12)
where, $x_i$ is the position of the $i^{\text {th}}$ particle, and $\theta_1, \theta_2, \ldots . \theta_d$ is the set of d hyperparameters, which we are optimizing. Each particle has its velocity and position updated in an iterative process as it searches for a solution within the solution space. Each particle's movement is controlled so that there is a balance between exploration and exploitation of the search space is defined as Eq. (13).
$v_i^{t+1}=w v_i^t+c_1 r_1\left(p_i-x_i^t\right)+c_2 r_2\left(g-x_i^t\right)$ (13)
where, $v_i^t$ and $v_i^{t+1}$ are the velocity of the $i^{t h}$ particle for iteration $t$ and $t+1$, respectively. $w$ is the inertia weight that determines the effect of the previous velocity on the new velocity; $c_1$ and $c_2$ are the acceleration coefficients that regulate the cognitive (self) and social (swarm) elements of search; $r_1$ and $r_2$ are independent random numbers that create a stochastic component to the search; $p_i$ is the best position ever experienced by the $i^{\text {th}}$ particle; and $g$ is the best position (solution) ever experienced by the entire swarm.
Each particle uses the above information to update its candidate position (solution) within the search space is defined as Eq. (14).
$x_i^{t+1}=x_i^t+v_i^{t+1}$ (14)
The current and latest positions (time t or t + 1) for particle i are $x_i^t$ and $x_i^{t+1}$ respectively. As a result of these updates particles will improve their position in terms of optimizing regions of the hyper parameter space.
3.6 Objective function
The purpose of the objective function is to evaluate the quality of each particle (i.e., the hyper parameter configuration) using classification error with regard to the best performing model. The binary cross-entropy loss is defined as Eq. (15).
$\mathcal{L}=-\frac{1}{N} \sum_{i=1}^N\left[y_i \log \left(\hat{y}_i\right)+\left(1-y_i\right) \log \left(1-\hat{y}_i\right)\right]$ (15)
where, N is number of data points, $y_i$ is true label, and $\hat{y}_i$ is predicted probability of each sample. When receiving an incorrect prediction, the loss will be greater than when receiving a correct prediction; meanwhile, when predicting with high confidence, an incorrect prediction will also incur a greater penalty than when predicting with low confidence. In this case, the PSO will minimize this specific loss in order to find the optimal parameters for the model.
The purpose of this layer is to take the feature representation learned from the attention mechanism and produce the final prediction output is defined as Eq. (16).
$\hat{y}=\sigma\left(W_z Z+b_z\right)$ (16)
where, $Z$ denotes the context vector produced by the attention mechanism, $W_z$ denotes the weight matrix, and $b_z$ denotes the bias term. The final output, which represents the probability of an individual having a mental illness, has been transformed into a probability value between 0 and 1 using the sigmoid activation function $\sigma(\cdot)$.
SHAP is a tool that is utilized to gain insight into how each feature contributes to the final model prediction is defined as Eq. (10).
$\phi_j=\sum_{P \subseteq F E \backslash\{j\}} \frac{|F E|!(|P|-|F E|-1)!}{|P|!}[f(F E \cup\{j\})-f(P)]$ (17)
where, $\phi_j$ is the amount of feature $j$ contributes according to the input feature set $P, F E$ is the part of the input feature set that does not include feature $j, f(P)$ is the model output with features from subset $S, f(P \cup\{j\})$ is the model output that includes feature $j$. The purpose of is to calculate the average marginal contribution per feature from all possible combinations of input features to create an unbiased attribution to each feature. SHAP provides both global feature importance as well as local feature importance.
To train the Attention-LSTM for each particle, use the respective parameters, evaluate using the loss function defined above, and update the personal $\left(p_i\right)$ and global $(g)$ best solutions based on evaluation. The velocities and positions of each particle are updated for convergence toward the optimal hyperparametric values as determined by the PSO equations. The process will continue until a converging criterion is met (maximum number of iterations or converging threshold have been reached), and then the model will be retrained with the optimal parameter set $\theta^*$, make predictions, and provide SHAP-based explanation for interpretability.
This study developed a new model called PSO-ALSTM, which is created from multiple tools within a Python-based deep learning framework. The experiments are performed on a computer with substantial computational resources, allowing the efficient training of multiple deep NNs. Hyper parameter optimization of the proposed model used a Particle Swarm Optimization-based approach that identified the optimal configuration by minimizing the loss function. Hyper parameters used for optimization included learning rate, number of hidden units, batch size, and dropout rate. Mini-batch learning and Adaptive Optimization algorithm are used to train the proposed model with early stopping used to prevent over fitting.
The search space for the key hyper parameters made up of varying values include: Learning Rate 0.0005–0.01; Number of LSTM Units 32-256; Number of LSTM Layers 1-3; Using each value point within this hyper parameter space will help discover the appropriate dropout rate between 0.1–0.5 and the number of LSTM units in the stack model by testing different batch sizes between 16 and 128. Upon completing the optimization process, the aggregate values resulting from multiple iterations of testing lead to selecting the final developed PSO-ALSTM model setup with the proper learning rate of 0.001, 128 LSTM units, 2LSTM layers stacked, a dropout rate of 0.3, and a batch size of 64. The model will have been trained for no more than 100 epochs using the Adam Optimizer with early stopping to control over fitting and generalization.
For PSO, the particle swarm has 25 particles with a maximum of 40 iterations to reach convergence. The inertia weight is set to 0.7 in order to achieve a balanced approach between exploration and exploitation; the cognitive coefficient and social coefficient are both fixed at 1.5 for an effective balance between learning individually and cooperating as a community of individuals. The empirical values and convergence behavior acquired from preliminary experimentations provide rationale for these parameter settings. In order to provide a fair and unbiased evaluation of the model, we have split the dataset into three subsets: training (70%), validation (15%), and test (15%) partitions using stratified sampling so that the distribution of classes will remain the same in all subsets. We will also use 5-fold cross-validation on the training subset to maximize the robustness of model determinations and minimize the variance of performance estimates. The results will be reported as the average of the performance of all five folds.
Standard classification metrics, including accuracy, precision, recall, F1 score, and the area under the Receiver Operating Characteristic (ROC) curve (AUC-ROC), will be utilized to evaluate the proposed model, thereby providing a full assessment of the predictive capability of the proposed model. The proposed model will be compared with baseline models (LR, RFs, Support Vector Machine (SVM), and a basic LSTM), in order to provide evidence of the effectiveness of the proposed model. Each experiment will also be repeated several times to provide statistical significance to the reported results, through reporting of mean performance and standard deviation. The determination of whether the improvements of the proposed model are statistically significant will be evaluated using statistical methods such as paired t-tests, when appropriate. An ablation study will also be conducted by removing individual components of the model, including the attention mechanism and the use of PSO optimization, in order to measure the contributions of each component to the overall model performance.
4.1 Dataset description
This research employs a combination of real-time dataset collection and publicly available datasets (NHANES) [28] to ensure that both data types are reliable and specific for predicting mental health among older adults. The NHANES dataset is used as a primary data source. The NHANES dataset provides clinical/physiological variables related to health and includes a number of variables that assess health in relation to body size, clinical laboratory evaluations, and responses to survey-based measures of mental health. Participants in the NHANES study vary in terms of age. Therefore, participants are filtered by age to create a cohort of participants in the study.
The addition of the Community Cognitive Health Dataset [29] provides further focus on the aging population. The Community Cognitive Health Dataset contains information on the functional/cognitive abilities of older adults, as well as information on the presence of depression, lifestyle characteristics, and chronic illness parameters that are particularly relevant to this research question. Together, the two datasets provide a great deal of information about changes in cognition and mental health among older adults. The NPHA dataset [30] is also included in the study to provide information on psychosocial/behavioral factors (e.g., social isolation, emotional well-being) and lifestyle; thus, the information in the NPHA dataset will complement the clinical data collected in the NHANES dataset. The primary dataset includes data collected from 300 individuals in both rural and urban areas to ensure representation of different demographics within the region.
The data are collected through a structured questionnaire and health assessments, capturing the following variables: physical health, sleep, daily activities, social interactions, and self-reported mental health. The inclusion of this dataset adds an additional layer of practical relevance to the study and enables verification of the proposed model in a localized setting. All datasets will be pre-processed before any model training and evaluation by performing steps such as imputing missing values, normalizing variable values, harmonizing dataset features, and so forth to ensure consistency and compatibility with model training and evaluation. By combining multiple datasets within the proposed study, the framework will enable a more complete understanding of the clinical, behavioral, and cognitive aspects of mental health. This will ultimately lead to better predictive performance and generalizability of the results across various populations.
4.2 Model performance on individual datasets
The proposed PSO-ALSTM model is evaluated on the four datasets separately. Performance is consistent and robust for the proposed framework across four data sources with different characteristics (NHANES, Community Cognitive Health Dataset, NPHA, and real-time data).
Table 1 and Figure 2 show the proposed model's performance by dataset. The model achieves the best performance on the Community Cognitive Health Dataset, with an accuracy of 93.2% and an F1 Score of 93.3%. The strong performance is likely due to the domain specificity of this dataset, which includes well-defined feature sets that directly reflect cognitive/mental health issues in older adults, allowing more effective learning of relevant patterns. The model achieves slightly lower performance on the NPHA dataset (accuracy 90.8%), likely due in large part to the preponderance of self-reported and psychosocial variables, which result in greater variation there.
Figure 2. Dataset-wise proposed model performance comparison
Table 1. Performance of the proposed model on individual datasets
|
Dataset |
Accuracy (%) |
Precision (%) |
Recall (%) |
F1-Score (%) |
AUC |
|
NHANES |
91.6 |
90.2 |
92.1 |
91.1 |
0.91 |
|
Community Cognitive Health |
93.2 |
93.5 |
93.1 |
93.3 |
0.93 |
|
NPHA |
90.8 |
89.6 |
91.3 |
90.4 |
0.91 |
|
Real-time Dataset (300 samples) |
92.1 |
91.0 |
93.0 |
92.0 |
0.92 |
The model shows excellent performance (accuracy 91.6%; AUC 0.91) on the NHANES dataset, demonstrating that it will effectively process diverse clinical and physiological data. Additionally, the model produces similar results with the real-time dataset; i.e., accuracy 92.1% and recall 93.0%. This shows that the proposed model is effective in real-life populations of both rural and urban individuals. Therefore, due to the consistently high recall and AUC values across datasets, the reliability and robustness of the proposed model in identifying individuals with mental health concerns are shown to be high.
4.3 Comparison with baseline models
An assessment of whether or not the suggested method is successful compared against traditional methods, including LR, SVMs, RFs, and standard LSTMs. Table 2 and Figure 3 present the performance of multiple methods using metrics commonly used in the field. LR achieved an accuracy of 81.2% with an AUC of 85%, indicating limited ability to capture complex nonlinear relationships within an individual's health records. While both SVMs and RFs demonstrate modest performance, RFs outperform SVMs, achieving an accuracy of 86.7% and an AUC of 0.89, owing to their ensemble nature.
Table 2. Performance of multiple methods using metrics
|
Model |
Accuracy (%) |
Precision (%) |
Recall (%) |
F1-Score (%) |
AUC |
|
LR |
81.2 |
79.5 |
80.3 |
79.9 |
0.85 |
|
RF |
86.7 |
85.9 |
87.1 |
86.5 |
0.89 |
|
SVM |
84.5 |
83.2 |
85.0 |
84.1 |
0.87 |
|
LSTM |
89.3 |
88.1 |
90.2 |
89.1 |
0.90 |
|
Proposed (PSO-ALSTM) |
92.8 |
91.5 |
93.2 |
92.3 |
0.92 |
Figure 3. Performance of multiple methods using metrics
The LSTM model shows much improved performance, with an accuracy of 89.3% and a recall of 90.2%, demonstrating its ability to capture the temporal dependencies inherent in longitudinal health records. Although it captures temporal dependencies, it does not incorporate mechanisms for weighting or ranking the relative importance of all temporal features, or for optimal parameter tuning.
The proposed model PSO-ALSTM has distinctly better performance in all measures when compared to baseline models: it has achieved an accuracy of 92.8%, a precision of 91.5%, a recall of 93.2%, an F1 score of 92.3%, and an area under the ROC curve of 92%, all of which are the highest of any of the models. The improvements made to this model are primarily attributable to two factors: first, the attention mechanism enhances feature relevance, and second, the PSO method optimizes hyperparameter selection. The increased recall value also reflects this model's capability to correctly classify older adult patients at risk of falling and, therefore, is very useful for supporting decision-making in a health care system by keeping false negatives to a minimum.
4.4 Ablation study
Table 3 shows that the components of the proposed model are evaluated for their contributions to the model's overall accuracy using the NHANES dataset. The baseline LSTM model achieved a total accuracy of 89.3% and an F1-score of 89.1%, confirming its ability to effectively learn temporal dependencies in the health dataset. However, with the inclusion of an attention mechanism, the model’s performance improved to 91.0% accuracy and 90.7% F1-score, confirming that attention enabled the model to focus more effectively on the most relevant time steps and feature patterns associated with the targeted mental health conditions.
Table 3. Ablation study
|
Model Variant |
Accuracy (%) |
F1-Score (%) |
|
LSTM (baseline) |
89.3 |
89.1 |
|
LSTM + Attention |
91.0 |
90.7 |
|
LSTM + PSO |
90.5 |
90.2 |
|
PSO + Attention-LSTM |
92.8 |
92.3 |
Similarly, the integration of PSO with LSTM resulted in an accuracy of 90.5% and an F1-score of 90.2%, further confirming the importance of optimization in selecting optimal hyperparameters and improving model convergence. Of particular note, the highest-performing model (PSO + Attention-LSTM) achieved the highest accuracy (92.8%) and F1-score (92.3%), confirming that the combination of temporal modeling, attention-based feature weighting, and bio-inspired optimization yields the greatest improvement in predictive performance. In summary, all ablation components contributed positively to the model’s predictive capabilities; however, combining all three components yielded the most robust and accurate model.
Table 4 compares the results obtained by the PSO-ALSTM framework proposed in this research with those achieved by other machine learning methods on the NHANES dataset. The AUC-ROC is an evaluation metric, the area under the receiver operating characteristic curve. Lim et al. [31] presented various traditional and ensemble machine learning models and compared them for predicting mental health. The best-performing traditional algorithm is Logistic Regression with an AUC of 0.879; however, it is closely followed by the Support Vector Machine (AUC = 0.876) and Multi-Layer Perceptron (AUC = 0.876) models. Furthermore, both XGBoost (AUC = 0.868) and LightGBM (AUC = 0.858) performed well compared with other algorithms.
Table 4. Performance comparison with the state-of-the-art methods
|
Reference |
Dataset Used |
Frameworks Used |
AUC-ROC |
|
Lim et al. [31] |
NHANES |
Logistic Regression |
0.879 |
|
NHANES |
Random Forest |
0.833 |
|
|
NHANES |
Support Vector Machine |
0.876 |
|
|
NHANES |
LightGBM |
0.858 |
|
|
NHANES |
Multi-Layer Perceptron |
0.876 |
|
|
NHANES |
XGBoost |
0.868 |
|
|
NHANES |
K-Nearest Neighbour |
0.829 |
|
|
Koelstra et al. [32] |
NHANES |
Recursive Feature Elimination (RFE) based Logistic Regression |
0.83 |
|
NHANES |
RFE-based Random Forest |
0.834 |
|
|
NHANES |
RFE-based Support Vector Machine |
0.82 |
|
|
NHANES |
RFE-based XGBoost |
0.839 |
|
|
NHANES |
RFE-based Stacking Logistic Regression |
0.848 |
|
|
NHANES |
RFE-based Stacking Support Vector Machine |
0.822 |
|
|
Zheng and Lu [33] |
NHANES |
Logistic Regression |
0.66 |
|
NHANES |
Random Forest |
0.65 |
|
|
NHANES |
Naïve Bayes |
0.68 |
|
|
NHANES |
Support Vector Machine |
0.68 |
|
|
NHANES |
LightGBM |
0.62 |
|
|
Proposed |
NHANES |
PSO-ALSTM |
0.92 |
RF and K-Nearest Neighbor performed significantly worse than their counterparts because they do not handle complex nonlinear relationships and heterogeneous healthcare data well. Koelstra et al. [32] described how Recursive Feature Elimination (RFE) was introduced to improve mental health predictive performance through an RFE-based approach. Since RFE combined with XGBoost and stacking produced slight improvements, with the highest AUC from the RFE-based stacking LR model (AUC = 0.848), it is clear that both feature optimization and the use of ensemble learning are critical to achieving more accurate predictions of mental health. However, neither approach has provided sufficient success due to their lack of temporal learning mechanisms or deep contextual extraction capabilities.
Zheng and Lu [33] showed that traditional machine-learning algorithms (including LR, RFs, Naive Bayes, SVM, and light gradient boosted models) achieved an average AUC between 0.62 and 0.68. This low performance is due to the inherent limitations of traditional machine learning models for predicting mental health disorders without advanced optimization techniques or sequential learning.
Compared with traditional machine learning methods, the proposed Particle Swarm Optimization and Attention-LSTM (PSO-ALSTM) model achieves the highest average AUC of 0.92. This result significantly exceeds all previous results obtained for the same purpose and is achieved by applying an attention-LSTM to capture temporal constraints and important behavioral patterns, in conjunction with particle swarm optimization for hyperparameter tuning and improved convergence. Overall, the results suggest that integrating temporal deep learning with bio-inspired optimization techniques will yield greater success than traditional machine learning or ensemble modeling methods for predicting individuals' mental health.
The proposed approach has demonstrated consistent performance on a real-time dataset, confirming that the framework is practically applicable to populations in both urban and rural settings. Overall, our analysis shows that there is a strong advantage to combining temporal deep learning and bio-inspired optimization techniques for accurately and reliably predicting the mental health of senior citizens.
In this study, a new hybrid deep learning framework (NTSM-Net) comprised of neuro-inspired sparse encoding, Bayesian temporal reasoning, micro-event analysis, and individual personality modulation for accurate detection of anxiety and depression was introduced. Through extensive experimentation, it is determined that the NTSM-Net achieved an accuracy of 96.8% and an F1-score of 96.5%, representing a performance improvement of 4.7%–9.4% over state-of-the-art models, including SAAM, T-GAT, Transformer-Encoder, and hybrid CNN-GRU networks. The model also demonstrated better generalization, greater robustness to noisy signals, and enhanced ability to detect weak behavioral cues across all evaluated datasets.
An ablation study of NTSM-Net's individual components demonstrated that the Bayesian Temporal Transformer and Personality Modulation layer substantially enhanced NTSM-Net's overall performance, indicating that uncertainty-aware temporal modeling techniques and individualized behavior adaptations are effective tools for maximizing performance in this application area. In addition to classification, the NTSM-Net system generates interpretable emotion trajectory maps and provides probabilistic severity estimates for clinically meaningful mental health assessment. Future work will focus on real-time deployment of the framework using lightweight edge devices, examining the impact of cross-cultural behavioral differences, and investigating continual learning strategies for long-term mental health monitoring.
[1] Mansoor, M.A., Ansari, K.H. (2024). Early detection of mental health crises through artificial-intelligence-powered social media analysis: A prospective observational study. Journal of Personalized Medicine, 14(9): 958. https://doi.org/10.3390/jpm14090958
[2] Wei, Y.T., Zhang, Y.Z., Zhang, S.Y., Zhang, H. (2023). CANAMRF: An attention-based model for multimodal depression detection. In PRICAI 2023: Trends in Artificial Intelligence, pp. 111-116. https://doi.org/10.1007/978-981-99-7022-3_10
[3] Ruslim, M.A., Spencer, M.J., Hogendoorn, H., Meffin, H., Lian, Y., Burkitt, A.N. (2024). Emergence of sparse coding, balance and decorrelation from a biologically-grounded spiking neural network model of learning in the primary visual cortex. bioRxiv. https://doi.org/10.1101/2024.12.05.627100
[4] Müller, S., Hollmann, N., Arango, S.P., Grabocka, J., Hutter, F. (2021). Transformers can do Bayesian inference. arXiv preprint arXiv:2112.10510. https://doi.org/10.48550/arXiv.2112.10510
[5] Xiao, Y.M., Shao, H.D., Wang, J., Yan, S., Liu, B. (2024). Bayesian variational transformer: A generalizable model for rotating machinery fault diagnosis. Mechanical Systems and Signal Processing, 207: 110936. https://doi.org/10.1016/j.ymssp.2023.110936
[6] Dwivedi, R., Kumar, D. (2022). Challenges of facial micro-expression detection and recognition: A survey. In International Conference on Neural Information Processing, pp. 483-492. https://doi.org/10.1007/978-981-99-1648-1_40
[7] Wang, Z.F., Zhang, K.H., Luo, W.H., Sankaranarayana, R. (2024). HTNet for micro-expression recognition. Neurocomputing, 602: 128196. https://doi.org/10.1016/j.neucom.2024.128196
[8] Ringwald, W.R., Wright, A.G. (2021). The affiliative role of empathy in everyday interpersonal interactions. European Journal of Personality, 35(2): 197-211. https://doi.org/10.1002/per.2286
[9] Liu, Z.Y., Yin, S.X., Lin, G.Y., Chen, N.F. (2024). Personality-aware student simulation for conversational intelligent tutoring systems. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, Miami, Florida, USA, pp. 626-642. https://doi.org/10.18653/v1/2024.emnlp-main.37
[10] Song, M.S., Yang, Z.J., Triantafyllopoulos, A., et al. (2024). Empowering mental health monitoring using a macro-micro personalization framework for multimodal-multitask learning: Descriptive study. JMIR Mental Health, 11: e59512. https://doi.org/10.2196/59512
[11] Castillo-Bolado, D., Davidson, J., Gray, F., Rosa, M. (2024). Beyond prompts: Dynamic conversational benchmarking of large language models. Advances in Neural Information Processing Systems, 37: 42528-42565. https://doi.org/10.52202/079017-1347
[12] Jia, X.W., Chen, J.X., Liu, K.X., Wang, Q., He, J.L. (2025). Multimodal depression detection based on an attention graph convolution and transformer. Mathematical Biosciences and Engineering, 22(3): 652-676. https://doi.org/10.3934/mbe.2025024
[13] Ahmad, A., Li, Z., Iqbal, S., et al. (2024). A comprehensive bibliometric survey of micro-expression recognition system based on deep learning. Heliyon, 10(5): e27392. https://doi.org/10.1016/j.heliyon.2024.e27392
[14] Jain, A., Bhakta, D. (2024). Micro-expressions: A survey. Multimedia Tools and Applications, 83: 53165-53200. https://doi.org/10.1007/s11042-023-17313-6
[15] Lukac, M. (2024). Speech-based personality prediction using deep learning with acoustic and linguistic embeddings. Scientific Reports, 14: 30149. https://doi.org/10.1038/s41598-024-81047-0
[16] Han, X., Chen, F.Y., Ban, J.R. (2024). FMFN: A fuzzy multimodal fusion network for emotion recognition in ensemble conducting. IEEE Transactions on Fuzzy Systems, 33(1): 168-179. https://doi.org/10.1109/TFUZZ.2024.3373125
[17] Xu, Z.R., Gao, Y., Wang, F., et al. (2025). Depression detection methods based on multimodal fusion of voice and text. Scientific Reports, 15: 21907. https://doi.org/10.1038/s41598-025-03524-4
[18] Huang, X.S., Wang, F., Gao, Y., et al. (2024). Depression recognition using voice-based pre-training model. Scientific Reports, 14: 12734. https://doi.org/10.1038/s41598-024-63556-0
[19] Nykoniuk, M., Basystiuk, O., Shakhovska, N., Melnykova, N. (2025). Multimodal data fusion for depression detection approach. Computation, 13(1): 9. https://doi.org/10.3390/computation13010009
[20] Tan, Z.X., Xiong, R.C., Wan, Y.L., et al. (2025). Detecting emotional dynamic trajectories: An evaluation framework for emotional support in language models. arXiv. https://doi.org/10.48550/arXiv.2511.09003
[21] Zhang, Y., Wang, Z.Y., Ding, Z.Y., et al. (2025). Employing machine learning and deep learning models for mental illness detection. Computation, 13(8): 186. https://doi.org/10.3390/computation13080186
[22] Zhou, S., Gao, C., Delbruck, T., Verhelst, M., Liu, S.C. (2025). Exploiting neuro-inspired dynamic sparsity for energy-efficient intelligent perception. Nature Communications, 16: 9928. https://doi.org/10.1038/s41467-025-65387-7
[23] Bahadi, S., Plourde, E., Rouat, J. (2024). Efficient sparse coding with the adaptive locally competitive algorithm for speech classification. arXiv preprint arXiv:2409.08188. https://doi.org/10.48550/arXiv.2409.08188
[24] Mamidisetti, S., Reddy, A.M. (2023). A stacking-based ensemble framework for automatic depression detection using audio signals. International Journal of Advanced Computer Science and Applications, 14(7): 603-612. http://doi.org/10.14569/IJACSA.2023.0140767
[25] Tveter, M., Tveitstøl, T., Hatlestad-Hall, C., et al. (2024). Advancing EEG prediction with deep learning and uncertainty estimation. Brain Informatics, 11: 27. https://doi.org/10.1186/s40708-024-00239-6
[26] Xie, S.S., Cheng, W., Nie, Z.L., et al. (2025). Bayesian cooperative probabilistic Transformer for remaining useful life prediction with uncertainty estimation in industrial equipment. Advanced Engineering Informatics, 67: 103515. https://doi.org/10.1016/j.aei.2025.103515
[27] Haraldsen, I.H., Hatlestad-Hall, C., Marra, C., et al. (2024). Intelligent digital tools for screening of brain connectivity and dementia risk estimation in people affected by mild cognitive impairment: The AI-Mind clinical study protocol. Frontiers in Neurorobotics, 17: 1289406. https://doi.org/10.3389/fnbot.2023.1289406
[28] Kopalidis, T., Solachidis, V., Vretos, N., Daras, P. (2024). Advances in facial expression recognition: A survey of methods, benchmarks, models, and datasets. Information, 15(3): 135. https://doi.org/10.3390/info15030135
[29] Malik, P., Singh, J., Ali, F., Sehra, S.S., Kwak, D. (2025). Action unit based micro-expression recognition framework for driver emotional state detection. Scientific Reports, 15: 27824. https://doi.org/10.1038/s41598-025-12245-7
[30] Hosseini, M., Sohrab, F., Gottumukkala, R., et al. (2025). A multimodal stress detection dataset with facial expressions and physiological signals. Scientific Data, 12: 1844. https://doi.org/10.1038/s41597-025-05812-0
[31] Lim, E., Jhon, M., Kim, J.W., Kim, S.H., Kim, S., Yang, H.J. (2025). A lightweight approach based on cross-modality for depression detection. Computers in Biology and Medicine, 186: 109618. https://doi.org/10.1016/j.compbiomed.2024.109618
[32] Koelstra, S., Muhl, C., Soleymani, M., Lee, J.S., Yazdani, A., Ebrahimi, T. (2012). DEAP: A database for emotion analysis using physiological signals. IEEE Transactions on Affective Computing, 3(1): 18-31. https://doi.org/10.1109/T-AFFC.2011.15
[33] Zheng, W.L., Lu, B.L. (2015). Investigating critical frequency bands and channels for EEG-based emotion recognition with deep neural networks. IEEE Transactions on Autonomous Mental Development, 7(3): 162-175. https://doi.org/10.1109/TAMD.2015.2431497
