Research on mood monitoring and intervention for anxiety disorder patients based on deep learning wearable devices

Abstract

Background

Anxiety disorders are common mental health issues that have a significant effect on people's quality of life. Conventional techniques for tracking emotional states frequently lack the accuracy and sensitivity needed for successful intervention.

Objectives

This project aims to create a sophisticated monitoring system that uses deep learning methods to evaluate physiological data from wearables, emphasizing heart rate variability (HRV), to forecast patients’ emotional states who suffer from anxiety disorders.

Methods

Wearable equipment monitors physiological characteristics, which we used to obtain patient HRV data. We processed the data using a Bidirectional Long-Short-Term Memory (Bi-LSTM) network to evaluate time-dependent variables and enhance the precision of emotional state predictions. The physiological signals were used to teach the model to recognize different emotional states, such as neutral, happy, and sad.

Results

Outperforming conventional machine learning models, the Bi-LSTM model showed a high accuracy rate of up to 97% in predicting emotional states. The findings suggest that ongoing HRV monitoring can accurately track shifts in emotional states and enable prompt responses.

Conclusion

This work emphasizes the possibility of real-time emotional state monitoring in patients with anxiety disorders with wearable technology and deep learning. The results point to the potential benefits of this strategy for improving emotional regulation and improving anxiety sufferers’ quality of life, opening new avenues for investigation and advancement in the field of mental health therapies.

Keywords

Bi-LSTM wearable devices deep learning people with anxiety disorders mood monitoring and intervention

1 Introduction

Anxiety disorders are a global social and public health problem, and according to the latest statistics, even in 2020, before the epidemic was adjusted, nearly 298 million people were estimated to be affected by anxiety disorders globally, or 6.3% of the global population.¹ This statistic highlights the severity and prevalence of anxiety disorders as a prevalent mental health condition. Traditional methods of assessing anxiety disorders, such as face-to-face counselling through a psychologist or self-report questionnaires, are limited to some extent. It is necessary to consider ethical factors such as patient privacy, data security, informed permission, confidentiality, and avoiding depersonalization while switching from traditional counselling to technology-driven evaluations. Equal access to these technologies is also essential, particularly for more vulnerable communities. These methods rely on patients’ subjective feelings and recollections, may be subject to memory bias, and do not capture patients’ mood swings in real time.² Anxiety disorder diagnostic techniques that are based on subjective self-reported experiences and recollections are typically vulnerable to recall bias and misinterpretations. Additionally, the lack of real-time mood and anxiety level variations in these tests limits their accuracy and depth of knowledge.

With the advancement of technology, wearable devices such as smartwatches and health trackers have been popularized in all aspects of our daily lives, revolutionizing the monitoring of an individual's health.³ This popularity has not only changed the way we grasp health information but has also opened up new opportunities for early identification and intervention of anxiety disorders. Wearable devices can provide uninterrupted monitoring of a user's physiological parameters, which means they can capture subtle changes that indicate an anxiety attack, such as irregular fluctuations in heart rate.⁴ Wearable technology tracks vital signs such as skin conductance and heart rate and reports real-time anomalies that may indicate an anxiety attack. With vibrations, alerts, or visual signals, these devices allow users to take proactive measures, such as practising deep breathing techniques, to reduce anxiety before it gets out of control.

While the proliferation of intelligent wearables offers unprecedented opportunities for monitoring and intervening in mental health conditions, particularly anxiety disorders, it also poses the significant challenge of accurately extracting helpful information from massive amounts of data to support effective interventions.⁵ The underlying algorithms and simplified models currently relied upon show clear limitations in dealing with these data's complexity and high dimensionality, hindering the potential for in-depth analysis and application from a mental health perspective.⁶ In addition, to date, no studies have predicted transient changes in anxiety and avoidance symptoms in populations with clinical levels of stress through the use of these data. This finding opens up the potential for developing timely, individualized adaptive interventions. Therefore, developing new techniques and models that can overcome the limitations of current approaches and effectively utilize the vast amount of data generated by intelligent wearable devices becomes a key issue in addressing mental health monitoring and intervention.⁷ Due to the volume and complexity of data, extracting meaningful insights from wearable devices can be difficult. Other challenges include separating mental health issues from normal variations, ensuring data accuracy, reducing noise, integrating various data types, and protecting user privacy when handling sensitive data. Bi-LSTM models capture relationships between historical and future data, improving data extraction accuracy from wearable devices. This improves real-time monitoring and makes it possible to identify anxiety symptoms and anticipate possible attacks more precisely.

Therefore, the deep learning method based on bi-directional long and short-term memory network (Bi-LSTM) proposed in this study further refined the analysis of time-dependent features in the data of physiological parameter of heart rate variability (HRV) with its ability to process the time-series data sensitively, which can more accurately monitor and predict the changes of stress and anxiety symptoms. The method improves data extraction accuracy from wearable devices and products and provides real-time monitoring and intervention. Smart wearables must include feedback mechanisms and real-time data processing to anticipate and control anxiety symptoms. These systems keep an eye on physiological cues, spotting anxiety symptoms and giving prompt feedback, which encourages early intervention and a more adaptable approach to mental health care. Through in-depth analysis of the physiological markers of psychological stress, this study opens a new path for mental health monitoring and intervention, making personalized and timely interventions possible.

The contribution points of this paper are as follows:

1.
By comparing existing anxiety monitoring methods, the Bi-LSTM-based model proposed in this paper significantly improves the accuracy of monitoring anxiety and stress states, achieving a prediction accuracy of up to 83%.
2.
with the wearable technology and Bi-LSTM deep learning model, this study achieves real-time monitoring and personalized intervention of the emotional state of patients with anxiety disorders, which is difficult to achieve in traditional mental health assessment and intervention methods.
3.
The innovative use of the method proposed in this paper provides a compelling new way to more accurately identify and predict an individual's psychological state, especially when dealing with highly dynamic and complex time-series physiological data.

2 Related work

2.1 Detection of mood by wearable devices

Wearable devices have become more accurate and easier to use, making them a powerful support for continuous monitoring of an individual's emotional state in daily life. This development not only helps to recognize mood changes in individuals promptly, which can lead to earlier interventions to alleviate anxiety symptoms, but also provides a new way to study the physiological basis of subtle mood changes. Healthcare practitioners incorporate wearable device data into patient care plans using real-time insights for customized actions. Data actionability and its ability to support clinical decision-making are among the challenges, along with privacy, security, interoperability, and interpretation. For usage to be practical, clinicians must be trained.

Hickey et al.⁸ attempted to monitor an individual's stress level using a wearable device to capture electrical skin activity (EDA) data. EDA, the conductance response of the skin, determines physiological activation levels, which are usually strongly correlated with emotional reactions and stress states. Choi et al.⁹ optimized a personal model specifically for individual data to improve prediction accuracy, and among machine-learning multilayer perceptron models, the generic model showed a high recall of 80%. In comparison, the individually adapted model had an average recall of 82.7%. Shu et al.¹⁰ introduced an innovative strategy to analyze heart rate data collected from intelligent wearable bracelets to identify people's emotional states, called “Neutral + Target,” and explore people's responses to three distinct emotional stimuli: neutral, happy, and sad. CAN et al.¹¹ proposed an advanced intelligent emotion perception method to detect emotional states in daily life effectively. Continuous monitoring data may predict emotional states by examining physiological signs such as skin conductance and heart rate. Correlations between these indicators and emotional states are found using machine learning techniques. Reliability, however, depends on individual variances, the prediction models’ resilience, and the data quality. Therefore, continuous improvement is required. The emotion recognition algorithm can filter out the most critical features from the collected physiological data and utilize the output of the automatic emotion detection system as a new input dimension to enhance the overall performance of the system. Wang et al.¹² emotion recognition technology, which is based on the identification of the current activity scene in which the user is located and through the analysis of a variety of perceptual data containing blood volume pulse, galvanic skin response and skin temperature of these physiological indicators to adapt to the emotional state of different people, the accuracy of its emotion detection is as high as 74.3%.

However, despite the rich physiological data that existing wearable devices can collect, it remains a challenge to accurately and efficiently identify and predict emotional states from this data.

2.2 Deep learning for data analytics

Deep learning techniques show great potential for processing and analyzing large amounts of complex data, opening up new possibilities for emotion monitoring and analysis. Compared with traditional data analysis methods, deep learning can mine deep features and complex data patterns to more accurately identify emotional states.

Tizzano et al.¹³ proposed to mine human emotional states from publicly available datasets by using Long Short-Term Memory Networks (LSTMs) as the core of a deep learning framework, which improves the overall accuracy of emotion recognition using highly detailed features extracted by the LSTMs and combining these features with Support Vector Machines (SVM) classifiers. Bi-LSTM encounters difficulties in recognizing emotions, such as overfitting, computing capacity, interpretability of the model, and data quality. Overfitting may result from insufficient data, inadequate computational resources, and low-quality data, which may impede its implementation on devices with limited resources. Bobade et al.¹⁴ adopted a deep learning approach to process and analyze multimodal physiological datasets from wearable devices to improve accuracy in recognizing human stress conditions. They ultimately achieved an accuracy of 84.32% in a triple classification task and 95.21% in a more straightforward binary classification problem. Khan et al.¹⁵ innovative the Deep Neural Network (DNN) structure for analyzing unprocessed RF data combined with processed RF signals to accurately classify and visualize human emotional states. With this approach, a high classification accuracy of 71.67% on an independent test population is now achieved, along with 71%, 72% and 71% in terms of accuracy, recall and F1 score. Nakisa et al.¹⁶ designed a deep learning framework fusing time-series multimodal data, which efficiently captures the complex nonlinear relationships between electroencephalogram (EEG) and blood volume pulse (BVP) signals, as well as within each of them, to optimize the recognition and classification of emotional states. To further enhance the efficacy of multimodal learning in complex sentiment analysis tasks, Bhatti et al.¹⁷ creatively introduced a novel strategy to improve cross-modal information flow and mutual understanding by establishing attention mechanism-oriented connectivity between separate Convolutional Neural Networks (ConvNet) to efficiently realize the relationship between the electrical skin activity and the sharing of intermediate representations between ECG signals. Naresh Kumar Reddy Panga¹⁸ investigates how fraud detection in healthcare might be enhanced by applying deep learning and machine learning approaches. It demonstrates notable advancements in detecting fraudulent activities using CNNs, RNNs, logistic regression, decision trees, and support vector machines. 99.9% of the time, the Decision Tree Classifier is accurate. The protection of the brain, heart, and lungs and the preservation of movement depend heavily on bones. They may break and result in immediate neurovascular problems. A deep learning system is required to identify and cure bone fractures immediately. This study suggests using a Systematized Attention Gate UNet (SAG-UNet) to diagnose bone fractures accurately. Using improved attention gates and unsupervised learning, the model tests bone break pictures with 89% accuracy.¹⁹

Although deep learning techniques have significant potential for processing and analyzing complex datasets in emotion monitoring, their ability to learn the dynamic changes in emotional states from continuous streams of physiological parameters is limited. The inability to understand the complex dynamic patterns of emotional changes in continuous physiological data leads to a lack of more accurate monitoring and prediction of emotional states. Therefore, this paper proposes to utilize Bi-LSTM to continuously detect the emotions of patients with anxiety disorders and make timely interventions.

3 Materials and methods

3.1 Overall construction

In this paper, the heart rate variability (HRV) data of patients with anxiety disorders is first collected using a wearable device. Heart Rate Variability (HRV) data, which is gathered from wearable technology, aids in understanding anxiety disorders by leading actions to enhance physiological functioning and monitoring the balance of the autonomic nervous system, particularly the sympathetic and parasympathetic responses. Then, this data is fed into a deep-learning network for processing. During this process, the deep learning network analyzes this data. It identifies the emotional states that the patient may be in, including but not limited to sad, regular, and happy, as shown in Figure 1. This monitoring and prediction of emotional states enables continuous monitoring of the emotions of patients with anxiety disorders and timely interventions. These interventions can be personalized suggestions for specific emotional states, such as emotion regulation techniques, psychological counselling, or relaxation training, thus helping patients better manage their emotional states and improve their quality of life.

Figure 1.

Overall architecture diagram.

3.2 Data pre-processing

In dealing with non-steady-state heart rate variability (HRV) signals, the critical task is to effectively capture the constant changes exhibited by the signal and the diversity of features it is rich in. Assuming the signal is stationary, the STFT partitions it into brief periods, making it an effective tool for analyzing non-steady-state HRV data. The HRV signals’ dynamics may be captured as they change over time by a time-varying spectrum analysis made possible by this. This makes it a helpful tool for tracking momentary variations in HRV data already-state signals; HRV exhibits fluctuations in frequency throughout the monitoring cycle, which requires researchers to employ analytical methods capable of revealing such dynamic changes. Monitoring HRV frequency fluctuations poses several obstacles, including non-stationarity of the signal, susceptibility to noise, and temporal precision. It can be challenging to select the proper analysis windows or methodologies for these problems, which require adaptive approaches like time-frequency analysis, pre-processing solid techniques, and balancing temporal resolution with frequency resolution. In this paper, in order to pre-process the HRV data, the method of Short-Time Fourier Transform (STFT) is introduced, which copes with the characteristics of this type of non-stationary signal by segmenting the data. To gain knowledge of the physiological foundations of anxiety disorders, extracting frequency components from human relative voluntary variation (HRV) data using the Short-Time Fourier Transform (STFT) approach is necessary. This technique offers insights into the dynamic nature of autonomic control. In the implementation of the STFT, the data is divided into multiple shorter time windows, and the signal within each window is then determined to be approximately stable. Subsequently, a spectral analysis is performed within each window to compute the frequency domain characteristics for that time range. Physiological signals, such as heart rate variability, may be identified using frequency domain analysis, which reveals patterns and rhythms associated with altered mental or physical health. Assisting in separating and examining various frequency bands, it also sheds light on the functioning and reaction of the autonomic nervous system. Short, quasi-stationary periods may be analyzed by segmenting HRV data into various time frames for dynamic change identification. This aids in the identification of sporadic occurrences or variations linked to emotional states or autonomic nervous system activity. Identifying specific patterns related to emotional states or stress reactions is enhanced by breaking down the HRV signal into smaller pieces for analysis. By moving the time windows along the signal axis and repeating the above process, a complete time-frequency image can be constructed detailing the time-frequency characteristics of the entire HRV signal. The Hamming window function decreases spectral leakage at the analysis window edges, improving time-frequency analysis for HRV data. As a result, edge effects are lessened and frequency representation is improved, improving the accuracy of HRV analysis and enabling the trustworthy identification and tracking of physiological changes linked to worry. The duration of the study's Hamming window was used to strike a compromise between time and frequency precision, enabling lower frequency components to be included while yet permitting rapid changes in HRV analysis. By striking a balance, it is possible to do efficient time-frequency analysis without unduly smoothing the data, which can mask significant physiological differences in anxiety disorders. In time-frequency analysis, the Hamming Window is a method used to reduce spectral leakage in non-stationary signals such as HRV. Although it lacks dynamic temporal adaptability, it smoothes edges. To handle complicated temporal patterns and capture long-term interdependence, LSTM networks employ temporal gating methods to regulate input flow. Using the Hamming window approach, one may better understand the physiological mechanisms behind anxiety disorders by reducing spectral leakage, improving frequency resolution, and enhancing signal representation in time-frequency analysis of HRV data. Unlike wavelet transform and empirical mode decomposition, STFT is a time-frequency analysis technique with a set window size, making it more straightforward to use and understand. It works well with both rapid and slow components of HRV signals; however, it can be noisy and computationally demanding.

In the practice of time-frequency analysis, the use of the Hamming window function helps to mitigate the spectral leakage that can be caused by window truncation. This window function implements the effect of smoothing the edges of the signal by applying a weighting process to both ends of the signal. Specifically for the study in this paper, a Hamming window with a duration of one second was selected and configured with a strategy that causes the window to slide once every second. Such a parameter setting is designed to achieve an optimal balance between time resolution and frequency resolution, allowing the ability to observe instantaneous changes in the signal while maintaining a frequency resolution accuracy that meets analytical needs.

3.3 Research on mood monitoring and intervention based on BILSTM

The Long Short-Term Memory Network (LSTM) introduces a complex gated mechanism and memory cell, effectively solving the gradient explosion and gradient vanishing problems prevalent in traditional Recurrent Neural Networks (RNNs). With these innovative structures, the LSTM can retain critical temporal information while discarding those that are not important for prediction, thus significantly enhancing the model's ability to deal with long-term dependency problems, as shown in Figure 2. x_t is the input at time t, h_t is the hidden layer state at time t, C_t is the internal state of the cell at time t, σ and tanh are the activation functions, W denotes weight matrix, b represents the bias, and [,. ] denotes the concatenation of two vectors.

Figure 2.

Structure of the LSTM cell.

In the LSTM architecture, the forgetting gate f_t determines which information needs to be maintained from the previous state of the cell and which should be discarded. In the LSTM architecture, the forgetting gate plays a critical role in supporting and deleting information from past cell states, limiting the buildup of obsolete or unnecessary data, and guaranteeing that only relevant data impacts future predictions—that is, noise or redundant data is avoided. This decision-making process involves the integration of the input vector x_t from the current time step t with the hidden state h_t−1 from the previous time step t-1, which is subsequently accomplished through the manipulation of linear transformations and nonlinear activation functions. The activation value ft of the forgetting gate can be calculated using Equation (1).

f_{t} = σ (W_{f} [h_{t - 1}, x_{t}] + b_{f})

(1)

Similar to forgetting gates, the operation of input gates can be computed using a specific formula. This computational process is critical in evaluating and determining which new input information is important enough at each time step to be added to the unit's state. The activation value of the input gate is shown in Equations. (2)-(3):

\begin{aligned} i_{t} & = σ (W_{i} [h_{t - 1}, x_{t}] + b_{i}) \end{aligned}

(2)

\begin{aligned} g_{t} & = \tanh (W_{g} [h_{t - 1}, x_{t}] + b_{g}) \end{aligned}

(3)

where g_t is the internal state at the time of candidacy.

${\tilde{C}}_{t}$ Is used to describe the input state at the moment t. Since the hidden layer output h_t−₁ at the moment t-1 and the input x_t at the moment, t undergoes a linear change after information integration to form a new state quantity as shown in Equation (4)

{\tilde{C}}_{t} = \tanh (W_{c} * [h_{^{t - 1}}, x_{t}] + B_{c})

(4)

The unitary state C_t maintains and conveys important information across time steps. Maintaining long-term dependencies, preventing vanishing gradients, and managing the update process via input, forget, and output gates are ways that the unitary state (Ct) in LSTM networks efficiently retains and conveys important information. This guarantees a steady memory state by enabling the network to transfer fresh information and remove unnecessary data selectively. The update process involves updating the internal state C_t−₁ of the previous moment to the internal state C_t of the current moment, as shown in Equation (5).

C_{t} = f_{t} * C_{t - 1} + i_{t} * {\tilde{C}}_{t}

(5)

The role of the output gate is to control the relationship between the output O_t and the cell state Ct at any time step t to ensure that the output reflects the currently crucial internal information, as shown in Equation (6).

O_{t} = σ (W_{o} * [h_{^{t - 1}}, x_{t}] + B_{o})

(6)

The hidden state output at each time step is determined by a combination of two key factors: the internal state at the current moment and the activation of the output gate, as shown in Equation (7). This strategy allows the LSTM to make informed decisions to keep or discard information at each point in the sequence to effectively manage the internal state's storage space.

h_{t} = O_{t} * \tanh (C_{t})

(7)

When processing time-series information such as heart rate variability (HRV) data, conventional LSTM captures the features of the data in a unidirectional manner only in chronological order, which may not be able to depict the complex patterns in the signal fully. In time-series data, Bidirectional LSTM (Bi-LSTM) performs better than Unidirectional LSTM because it can detect intricate patterns in both forward and backward directions. Because it takes into account both the past and the future environment, this enables improved prediction in tasks like sentiment analysis, voice recognition, and HRV data analysis. Bi-LSTM is a reliable and effective technique that enhances pattern identification, captures temporal connections, and helps categorize emotional states. However, it is more prone to overfitting, especially in small or imbalanced datasets, due to extra parameters, longer training durations, and higher computing resource requirements. The features of time-series data, such as ECG signals, are not only related to the state of the current moment but are also strongly influenced by the states of previous and subsequent moments.

Applying the Bi-LSTM network to emotion recognition, especially in detecting the emotions of patients with anxiety disorders, its ability to capture such bidirectional information is essential, as shown in Figure 3. Modelling complicated nonlinear connections, improving classification performance, and translating Bi-LSTM features to higher-dimensional spaces are all made possible by the Support Vector Machines (SVM) kernel function. The kernel selection directly influences the model's capacity to categorize high-dimensional input correctly. By comprehensively analyzing the time-series information in HRV data, Bi-LSTM can help the model identify and understand subtle but decisive patterns in their impact. Bi-LSTM increases classification accuracy by efficiently capturing essential elements in HRV data. These include temporal dependences, frequency-domain elements like LF and HF power, nonlinear dynamics, and time-domain features like SDNN and RMSSD. These aspects aid in the identification of intricate interactions and abnormalities in HRV, which may indicate physiological or emotional states. These patterns may be slight differences between normal and abnormal heart rhythms or changes in HRV characteristics caused by emotional changes. Thus, Bi-LSTM improves the accuracy of arrhythmia classification and also enhances the accuracy of monitoring changes in emotional states such as anxiety disorders.

Figure 3.

Network structure of Bi-LSTM.

In addition, this paper combined the feature extraction capability of Bi-LSTM with the classification capability of SVM. Upon completion of Bi-LSTM, deep features captured by the model rather than the direct output were extracted and transformed into feature vectors for downstream SVM processing. The combination of Bi-LSTM and SVM performs better in classification tasks than other models because of their ability to extract features, the power of SVM in high-dimensional spaces, and their kernel trick. SVM performs well in high-dimensional spaces and manages non-linearly separable data by employing various kernel functions, whereas Bi-LSTM can extract intricate, high-dimensional features. Incorporating the features into SVM for classification enables the definition of an optimized decision boundary in the multidimensional feature space, accurately classifying the emotional state of patients with anxiety disorders. SVM using linear kernels provides several benefits for emotion classification in high-dimensional feature fields. They are resilient because they generalize effectively on unknown data, lowering the chance of overfitting. They are computationally less demanding, quicker to train, more straightforward to understand, and successful in high-dimensional domains. Using the linear kernel of SVM can find hyperplanes in the feature space that can segment different emotional states at maximum intervals, which improves the generalization ability of classification, as shown in Equation (8):

K (x_{i}, x_{j}) = x_{i}^{T} x_{j}

(8)

where K represents the kernel function and (x_i) and (x_j) are the sample vectors in the feature space. The core idea of the SVM algorithm is to find an optimal hyperplane in the high-dimensional or primitive feature space to maximize the separation of different data classes. The sample points near the decision boundary, i.e., the support vectors, play a decisive role in determining the hyperplane. By applying SVM in the feature space extracted via Bi-LSTM and using a linear kernel to construct the decision boundary, a final model of anxiety disorder patients’ emotions capable of effective classification based on high-dimensional features is formed. With the help of extracted features, the SVM reduces misclassification and improves generalization on unknown data by optimizing decision boundaries in 42-dimensional feature space by dividing data points, maximizing the margin between classes, and using kernel functions to handle complex feature relationships.

4 Experimental results

4.1 Environment configuration

In this experiment, an Intel Core i7 CPU and NVIDIA GTX 2080 GPU are equipped to cope with deep learning model training and inference demands. The NVIDIA GTX 2080 GPU is a well-liked option for deep learning because of its affordability, performance, and accessibility. Large datasets benefit significantly from its high CUDA and Tensor Core counts, which allow for effective parallel processing, quick data transmission, and interoperability with deep learning frameworks like TensorFlow and PyTorch. The computer has a 1TB hard disk capacity and 64GB of RAM, providing sufficient computing and storage resources. Python was adopted as the primary development language for the development environment, and PyCharm was chosen as the development tool. PyTorch was selected as a deep learning framework. PyTorch makes it easy to build complex neural network structures and utilize GPU acceleration for model training and inference, as shown in Table 1.

Table 1.
Experimental environment configuration.

Software Configuration parameters Hardware Configuration parameters

CPU Intel Core i7 Operating systems Windows7 64 bit

GPU NVIDIA GTX 2080Ti Development languages python

Hard Disk 1TB Development tools Pycharm

RAM 64GB Development framework Pytorch

Software	Configuration parameters	Hardware	Configuration parameters
CPU	Intel Core i7	Operating systems	Windows7 64 bit
GPU	NVIDIA GTX 2080Ti	Development languages	python
Hard Disk	1TB	Development tools	Pycharm
RAM	64GB	Development framework	Pytorch

4.2 Model parameter settings

It is crucial to ensure the experiment's reproducibility when performing experimental simulations. For validity, scientific integrity, and future research, reproducibility in scientific studies is essential. It allows for future development and validates that the findings are consistent and dependable under comparable circumstances. Because of this crucial premise, the results are solid, trustworthy, and beneficial to the scientific community. According to the setting requirements of the experimental parameters, the fundamental parameters were configured, as shown in Table 2. Among them, batch size refers to the amount of data processed by the model at one time, which directly affects the model's learning efficiency and memory usage; embedding size determines the size of the dimension of the vector space that the model converts into when processing the input data, which affects the model's ability to capture features. Optimizer is an algorithm that adjusts the model parameters to reduce the error. An optimizer such as Adam or SGD uses gradient descent, individual learning rate adaptation, and minimal convergence to minimize error in model parameter adjustments. This lessens prediction error and minimizes error over time as the model converges to optimal settings. The Adam optimizer was chosen because its adaptive learning rate feature helps accelerate the convergence of the model and improves stability. The adaptive learning rate feature of the Adam optimizer enhances stability, especially in early training stages, by modifying learning rates according to gradient moments. This ensures that more significant gradients receive minor updates and smaller gradients receive larger ones.

Table 2.
Model parameter setting.

Name Value Name Value

Batch size 128 Embedding 64

Epoch 64 dropout 0.2

BiLSTM layers 8 Learning rate 8e-7

Name	Value	Name	Value
Batch size	128	Embedding	64
Epoch	64	dropout	0.2
BiLSTM layers	8	Learning rate	8e-7

4.3 Data sets

Heart Rate Variability (HRV) dataset and electrocardiogram (ECG) data obtained from the Amigos dataset provide a rich data source for HRV analysis. It offers the opportunity to extract a range of HRV measures, including Time-domain features: metrics such as mean RR interval, SDNN, RMSSD, and pNN50 in the dataset can provide insights into intrinsic heartbeat interval variability and overall cardiac health in different emotional and social settings. Frequency-domain features: Spectral components such as LF and HF and their ratios (LF/HF) found or calculated in the dataset reveal autonomic nervous system dynamics in response to emotional stimuli. Nonlinear features: Given the complexity of physiological data, nonlinear metrics such as sample entropy or fractal dimension can be derived to understand the predictability of heart rate time series under different emotional states and social interactions.

4.4 Evaluation indicators

In this paper, three main standard metrics are used to evaluate the model's performance: precision (Pre) and recall (Rec), F1. AUC-ROC, F1-Score, confusion matrix, accuracy, precision, and recall are standard measures used to assess the efficacy of a model. Recall gauges the model's capacity to recognize pertinent occurrences, precision gauges optimistic predictions, and accuracy gauges how well the model classifies data points. As shown in Equation (9), Pre refers to the proportion of correctly predicted positive classes to the proportion of all predicted positive classes. As shown in Equation (10), Rec refers to the proportion of the positive samples predicted as favourable to the positive samples. As shown in Equation (11), F1 combines Pre and Re, and the higher the F1 score, the more reliable the proposed algorithm is.

\begin{aligned} P r e c i s i o n (P r e) & = \frac{T P}{T P + F P} \end{aligned}

(9)

\begin{aligned} R e c a l l (R e c) & = \frac{T P}{T P + F N} \end{aligned}

(10)

\begin{aligned} F 1 & = \frac{2 \times P r e \times R e c}{P r e + R e c} \end{aligned}

(11)

4.5 Model analysis and comparison

As shown in Figure 4, using the Bi-LSTM approach with parameters Batch size = 128 and Epoch = 64, the precision of mood monitoring for anxiety patients will be 97%.

Figure 4.

The precision of the model.

In the experiments, a standard model evaluation method, k-fold cross-validation, was used, where the value of k was taken as 10. In each division, the ratio of the training data to the test data was set at 70% to 30%. Such a setting means that most of the data was used for training the model, while about one-third of the data was reserved for testing the model's effectiveness, as shown in Figure 5.

Figure 5.

Precision rate of the model in the test set and training set.

In addition, in this research to accurately monitor the emotional state of patients with anxiety disorders, a combination of Bi-LSTM and branched SVM was used. SVM performs well in high-dimensional spaces to categorize emotional monitoring data, although it may have trouble with complicated sequences and temporal correlations. Although it takes more training data and increases computing complexity, integrated Bi-LSTM and SVM may capture complicated patterns and temporal correlations. Thereby, the emotional state of anxiety patients was accurately determined, which assisted doctors in developing a more precise treatment plan. In addition, to focus on the model's ability to recognize anxiety in the dichotomous mode (i.e., anxious vs. non-anxious) and its ability to distinguish between different emotions in the more complex trichotomous modes (“happy”, “sad”, and “regular”) in determining the accuracy of various emotional states. The performance of SVM,⁵ LSTM,²⁰ LSTM + SVM,²¹ and Bi-LSTM²² models are compared, as shown in Tables 3 and 4.

Table 3.

Anxiety detection accuracy in a dichotomous classification model.

	Anxiety	Non-anxiety
SVM	67%	63%
LSTM	73%	71%
LSTM + SVM	86%	84%
Bi-LSTM	93%	90%
Bi-LSTM + SVM	97%	96%

Table 4.

Anxiety detection accuracy in three classification models.

	Happy	Regular	Sad
SVM	62%	58%	70%
LSTM	77%	75%	78%
LSTM + SVM	85%	80%	86%
Bi-LSTM	93%	94%	92%
Bi-LSTM + SVM	99%	97%	95%

The data in Tables 3 and 4 show that in emotion monitoring for patients with anxiety disorders, the use of a combination of Bi-LSTM and SVM significantly improves the accuracy of emotion recognition in both different classification modes compared to the use of SVM, LSTM, or a combination of LSTM and SVM individually.

The following trend for the dichotomous classification model is that the SVM model has a lower recognition accuracy in anxious and non-anxious states, 67% and 63%, respectively. Class overlap, unbalanced data, and feature homogeneity might cause the SVM model's reduced recognition performance in anxious and non-anxious states. The model's ability to correctly identify various states may be hampered by overlapped physiological data, under-representation of one class, and a lack of discriminative characteristics. The LSTM model's performance improves, with recognition accuracies increasing to 73% and 71%. When LSTM was combined with SVM, the recognition accuracies increased significantly to 86% and 84%. Bi-LSTM achieved the highest recognition accuracies when combined with SVM, 97% and 96%, respectively, and the model achieved the highest accuracies in the triple categorization task, 99%, 97%, and 95%, which is almost close to perfect analysis.

5 Conclusion and perspective

In this research, an approach based on combining a Bi-LSTM network and Support Vector Machine SVM was adopted through in-depth analysis and processing of HRV data to achieve high-accuracy monitoring and timely intervention on the emotional state of patients with anxiety disorders. The method improves the accuracy of data extraction from wearable devices and dramatically enhances the ability of real-time monitoring and personalized intervention of mental health conditions. By comparing with existing techniques, this study demonstrates the superiority of using Bi-LSTM in conjunction with SVM in anxiety monitoring, especially regarding accuracy and utility when dealing with highly dynamic and complex time-series physiological data.

Future research could explore more physiological data, such as electrical skin activity (EDA) and blood volume pulse (BVP), to further improve the accuracy and comprehensiveness of emotional state monitoring. In addition, the interpretability of deep learning models is another important research direction, and developing interpretable AI models can help us better understand the basis of the model's judgment.

Footnotes

Author contributions

Xiao Gu is responsible for designing the framework, analyzing the performance, validating the results, and writing the article. Xuedan Hu is responsible for collecting the information required for the framework, provision of software, critical review, and administering the process.

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data availability

No datasets were generated or analyzed during the current study.

References

Jacobson

Lekkas

Huang

, et al. Deep learning paired with wearable passive sensing data predicts deterioration in anxiety disorder symptoms across 17–18 years. J Affect Disord 2021; 282: 104–111.

Abd-Alrazaq

AlSaad

Aziz

, et al. Wearable artificial intelligence for anxiety and depression: scoping review. J Med Internet Res 2023; 25: e42672.

Masino

Forsyth

Nuske

, et al. M-Health and autism: recognizing stress and anxiety with machine learning and wearables data. In: 2019 IEEE 32nd international symposium on computer-based medical systems (CBMS). IEEE, 2019, pp.714–719.

Tsai

Chen

Liu

, et al. Panic attack prediction using wearable devices and machine learning: development and cohort study. JMIR Med Inform 2022; 10: e33063.

Nath

Thapliyal

. Machine learning-based anxiety detection in older adults using wristband sensors and context feature. SN Comput Sci 2021; 2: 359.

Moshe

Terhorst

Opoku Asare

, et al. Predicting symptoms of depression and anxiety using smartphone and wearable data. Front Psychiatry 2021; 12: 625247.

Jin

Gao

Yang

, et al. Attention-block deep learning-based features fusion in wearable social sensors for mental well-being evaluations. Ieee Access 2020; 8: 89258–89268.

Hickey

Chalmers

Newton

, et al. Smart devices and wearable technologies to detect and monitor mental health conditions and stress: a systematic review. Sensors 2021; 21: 3461.

Choi

Lee

Kim

, et al. Depressed mood prediction of elderly people with a wearable band. Sensors 2022; 22: 4174.

10.

Shu

Chen

, et al. Wearable emotion recognition using heart rate data from a smart bracelet. Sensors 2020; 20: 718.

11.

Can

Ersoy

. Smart affect monitoring with wearables in the wild: an unobtrusive mood-aware emotion recognition system. IEEE Trans Affect Comput 2022; 14: 2851–2863.

12.

Wang

Zhao

, et al. Emotionsense: an adaptive emotion recognition system based on wearable smart devices. ACM Trans Comput Healthc 2020; 1: 1–7.

13.

Tizzano

Spezialetti

Rossi

A deep learning approach for mood recognition from wearable data. In: 2020 IEEE international symposium on medical measurements and applications (MeMeA). IEEE, 2020, pp.1–5.

14.

Bobade

Vani

. Stress detection with machine learning and deep learning using multimodal physiological data. In: 2020 second International conference on inventive research in computing applications (ICIRCA). IEEE, 2020, pp.51–57.

15.

Khan

Ihalage

, et al. Deep learning framework for subject-independent emotion detection using wireless signals. Plos One 2021; 16: e0242946.

16.

Nakisa

Rastgoo

Rakotonirainy

, et al. Automatic emotion recognition using temporal multimodal deep learning. IEEE Access 2020; 8: 225463–74.

17.

Bhatti

Behinaein

Rodenburg

, et al. Attentive cross-modal connections for deep multimodal wearable-based emotion recognition. In: 2021, The 9th international conference on effective computing and intelligent interaction workshops and demos (ACIIW). IEEE, 2021, pp.01–05.

18.

Panga

NKR

. Financial fraud detection in healthcare using machine learning and deep learning techniques. Secur Commun Netw 2021; 2021: 9293877.

19.

Devi

Aruna

Almufti

, et al. Bone feature quantization and systematized attention gate UNet-based deep learning framework for bone fracture classification. Intell Data Anal (Preprint) 2024: 1–29.

20.