Advancing Stress Detection and Health Monitoring with Deep Learning Approaches ^†

Abstract

Numerous studies in the healthcare field conducted in recent years have highlighted the impact of stress on health and its role in the development of several critical illnesses. Stress monitoring using wearable technologies, such as smartwatches and biosensors, has shown promising results in improving the management of this issue. Data from both physical and mental health can be leveraged to enhance medical decision-making, support research on new treatments, and deepen our understanding of complex diseases. However, traditional machine learning (ML) systems often face limitations, particularly in real-time processing and resource optimization, which restrict their application in critical situations. In this article, we present the development of a deep learning-based approach that leverages models such as 1D CNN (Convolutional Neural Networks), LSTM (Long Short-Term Memory), and Time-Series Transformers, alongside classical deep learning techniques. We then highlight the transformative potential of TinyML for real-time, low-power health monitoring, focusing on Heart Rate Variability (HRV) analysis. This approach aims to optimize personalized health interventions and enhance the accuracy of medical monitoring.

1. Introduction

Stress, increasingly prevalent in our daily lives due to our modern lifestyle characterized by a fast-paced rhythm, high professional demands, and constant hyperconnectivity, affects both our mental and physical well-being [1]. Social pressures, multiple responsibilities, and continuous exposure to digital solicitations contribute to rising stress levels. Learning to manage it is crucial for maintaining long-term health [2].

While stress is a natural response to daily challenges, it can have negative effects when excessive or chronic. High stress reduces productivity, affecting concentration, memory, and decision-making [3]. It also impacts mental health, leading to anxiety, sleep disorders, and, in severe cases, depression. Additionally, it has physical health consequences, aggravating diseases such as hypertension and cardiovascular disorders while weakening the immune system. Real-time monitoring of stress is essential to prevent these negative health effects [4].

Modern technologies, such as smart sensors and AI-based solutions, provide non-invasive tools to analyze physiological signals related to stress, enabling proactive management. By using biometric data from the WESAD dataset [5], AI models can analyze heart rate and its variability, directly correlated with stress, helping optimize stress self-regulation, identify stress levels, and generate real-time alerts.

This study aims to automatically detect an individual’s stress state by analyzing physiological data during stressful situations. Using machine learning methods, individuals are classified as stressed or not stressed, and then into different categories (normal, amused, or stressed). We also develop context-aware stress detection approaches using deep learning [6], reducing motion artifacts when assessing stress based on heart rate and electrodermal activity. Finally, we explore the challenges of deploying 1D Convolutional Neural Networks (CNNs) [7] on resource-constrained microcontrollers for real-world stress recognition, while ensuring the privacy of health data [8].

2. Related Studies

A comprehensive study highlights the importance of integrating psychological, physiological, behavioral, and contextual data in stress assessment due to its complex impact on behavior. Datasets like WEMAC [9], used in developing the Bindi system [10], provide essential physiological data for stress recognition. Moreover, LSTM neural networks [11] have been applied to emotion recognition using the DEAP dataset v1.0 [12], achieving 80.64% accuracy by capturing long-term dependencies in brain signals. Additionally, a 1D CNN structure was employed on ECG and GSR data, achieving 75% accuracy in emotion classification using the Amigos dataset [13,14].

The WESAD dataset [5] used in this study tracks emotions by exposing individuals to three emotional states—baseline, stress, and amusement—while collecting corresponding physiological data. Machine learning techniques using this dataset have shown accuracy rates of 80% and 93% for three-category and binary classifications, respectively. However, when using only wrist-based data, accuracy slightly decreases to 75.21% and 87.12%.

Emotion recognition algorithms primarily rely on traditional machine learning (ML) [15] and deep learning (DL) [16]. While ML requires specialized expertise to extract meaningful features from physiological signals, DL techniques have improved emotion recognition by enabling autonomous learning and feature abstraction. Despite these advancements, existing solutions still face challenges such as high energy consumption, lack of scalability, and limited real-time capabilities. Most solutions depend on cloud-based architectures, which introduce issues like delayed data processing, high energy use, and privacy concerns. Additionally, many models are designed for the general population and lack personalized health interventions.

3. Materials and Methods

This study focuses on physiological signals, as they are more objective, less influenced by conscious control, and generally require lower energy consumption. Commonly used physiological signals include electrocardiogram (ECG) [17], skin temperature (SKT), blood volume pulse (BVP), and galvanic skin response (GSR), all of which have shown reliability in emotion recognition, particularly in studies involving IoT technologies for real-time emotion analysis and remote monitoring [8,18]. Among these, electrodermal activity (EDA) [19]—measured using wearable devices—has received particular attention for its ability to reflect variations in sympathetic nervous system activity by monitoring sweat gland responses. The integration of EDA sensors into commercial smartwatches has become increasingly common, making this modality practical for real-world applications. Consequently, datasets containing EDA signals from wearable devices are essential. In this context, the WESAD (Wearable Stress and Affect Detection) dataset was selected, as it provides rich, publicly available physiological data collected via wearable sensors for stress and emotion recognition research. Within WESAD, wrist-based signals were prioritized because they offer a good trade-off between information richness and ease of acquisition in realistic settings. This decision also supports the intended deployment of the model on embedded, low-power devices using TensorFlow Lite, since wrist-worn sensors such as smartwatches enable continuous, non-intrusive data collection while meeting the constraints of mobile environments.

3.1. Dataset Processing and Features Analysis

The WESAD dataset is designed for stress and emotion detection using wearable devices. It was collected from 15 participants (12 men and 3 women, aged 24 to 35 years) and contains physiological measurements from two sensors: The Empatica E4 wristband (Empatica Inc., Boston, MA, USA), which integrates an accelerometer, photoplethysmography (PPG), electrodermal activity (EDA), and skin temperature sensors, was used in this study and the RespiBAN chest belt (Biovotion AG, Zurich, Switzerland), which records electrocardiogram (ECG), respiration, temperature, electrodermal activity (EDA), and acceleration data. The data is labeled according to three emotional states: baseline (rest), stress (stressful task), and amusement (watching humorous videos). This dataset, available in .pkl format, can be downloaded under an academic license from platforms like PhysioNet. For our study, we will use only the data from the Empatica E4 wristband. These data are crucial for training and evaluating machine learning models for stress and emotion detection. The following are the main types of signals used in the WESAD dataset, which are summarized in

Table 1. Key characteristics of the WESAD dataset.

Key	Description
Number of participants	15 individuals (12 male, 3 female)
Average age of participants	27.5 years
Total recording duration	36 h of data (all participants combined)
Total number of samples	2.7 million rows (depending on sensors and signals)
Number of raw features	15 raw signals
Classes (labels)	Baseline (rest), Stress, Amusement
Total dataset size	1.2 GB (compressed .pkl format)

:

• EDA (Electrodermal Activity) [20]: Captures skin conductance, related to sweat gland activity.
• PPG (Photoplethysmography) [21]: Measures blood volume changes, indirectly related to heart rate
• Accelerometer [22]: Captures motion data in three axes.
• Skin Temperature [23]: Monitored via a wrist sensor.

3.2. Diagram of the Stress Detection Methodology and Deployment Process

To implement a deep learning algorithm, we cleaned and prepared the data (handling missing values and normalization). We then trained the model on a dedicated dataset so that it could learn the underlying patterns. After that, we evaluated it by making predictions on a test set to measure its performance and its ability to generalize. Once the model was validated, it was optimized for constrained environments using TensorFlow Lite [24]. This included converting the model into a lightweight format using techniques like quantization to reduce its size and improve efficiency on mobile or embedded devices. The optimized model was then integrated into an application or embedded system, enabling fast local execution without relying on a cloud connection, as shown in Figure 1.

3.3. Deep Learning-Based Classification Approaches

In this study, we explored several deep learning architectures for classifying physiological signals from the WESAD dataset. These models, specifically designed for time-series analysis, enable the capture of complex dynamics in physiological signals such as ECG, EDA, and RESP. The algorithms used are as follows:

• 1D CNNs (Convolutional Neural Networks) [25]: The data is prepared by normalizing with StandardScaler and encoding labels using LabelEncoder. The CNN model consists of three Conv1D layers with kernel size 3 and padding = 1 to maintain input dimensions. Each layer is followed by batch normalization and a ReLU activation for better convergence. Dropout regularization of 50% is applied to fully connected layers, with output dimensions of 256, 128, and for the final layer corresponding to the number of classes. Optimization is performed using the Adam optimizer with a learning rate of 0.001. The SparseCategoricalCrossentropy loss function is used for multi-class classification with integer labels. The model’s performance is evaluated using accuracy and F1 score. This approach combines Conv1D layers and advanced optimization techniques, such as Adam and SparseCategoricalCrossentropy, to ensure strong classification results.
• LSTM (Long Short-Term Memory) [26]: We applied advanced regularization and optimization techniques to improve the model’s performance. After normalizing the data using StandardScaler and encoding the labels with LabelEncoder, we used an LSTM model consisting of two LSTM layers, with 128 hidden units and a 50% dropout to prevent overfitting. Batch normalization was applied to enhance stability and convergence. A fully connected layer makes the final prediction, followed by a Softmax activation to obtain probabilities. Optimization was performed using the Adam algorithm (with a learning rate of 0.0005), and the chosen loss function is CrossEntropyLoss, suitable for multi-class classification. Finally, the model’s performance was evaluated by measuring accuracy on the test set.
• Time-Series Transformers [27]: The TimeSeriesTransformer model is defined with several key hyperparameters for time-series classification. The input data is first passed through an embedding layer (nn.Linear), which transforms the features into a higher-dimensional space (d_model = 32). The core of the model is the Transformer encoder, consisting of multiple layers (num_layers = 2) of attention mechanisms, allowing the model to learn dependencies within the time-series data. The attention mechanism uses 8 heads (num_heads = 8), enabling the model to focus on different aspects of the input sequence.

The model uses a dropout rate of 0.1 to prevent overfitting, with an additional 50% dropout applied in the fully connected layers (fc1, fc2, and fc3) to reduce the model’s complexity and improve generalization. The final output layer uses a softmax activation, which is suitable for multi-class classification tasks (output_dim = 2 for binary classification). The loss function used is CrossEntropyLoss, which is well-suited for classification tasks, and optimization is carried out using the Adam optimizer (learning_rate = 0.001), a popular choice in deep learning due to its adaptive learning rate.

The model is trained for 10 epochs (epochs = 10), with the optimizer updating the weights at each step based on the loss calculated from the predictions compared to the true labels. The training process uses mini-batches of size 64 (batch_size = 64), which helps manage memory and computational resources efficiently.

• Autoencoders followed by a Dense Classifier [28]: This approach first compresses signals into a latent space using an autoencoder, followed by classification with a dense neural network.

In this algorithm, we combined two main components: an encoder and a classifier. The encoder consists of two linear layers with a ReLU activation between them, which gradually reduce the dimensionality of the input data to a compressed representation (code_dim = 64). The dense classifier, which follows the encoder, consists of three linear layers with ReLU activations and a 50% dropout rate to reduce overfitting. The final layer of the classifier uses the softmax activation function, which is suitable for multi-class classification tasks, to predict the probabilities of the different classes.

Optimization is performed using the AdamW optimizer, which is a version of the Adam algorithm that includes L2 regularization via the weight_decay = 1 × 10⁻⁵ parameter. This helps prevent overfitting by penalizing overly large weights. The loss function used is CrossEntropyLoss, which is ideal for multi-class classification problems. The model is trained for 30 epochs with a learning rate of 0.001.

4. Results and Discussion

In this study, two classification tasks were performed: a binary classification distinguishing between “stressed” and “non-stressed” states, and a three-class classification between “normal”, “stressed”, and “amused”. The results show that in binary classification, the accuracy reached 92.86%, while in three-class classification, it reached 86.79%.

Table 2. Classification performance for binary and three-class tasks.

Algorithms	Binary (Stressed, Non-Stressed)				Three-Class (Normal, Stressed, or Amused)
	Accuracy	Recall	F1-Score	Precision	Accuracy	Recall	F1-Score	Precision
CNN 1D	92.86	96.23	88.22	81.44	86.79	92.52	81.58	73.63
LSTM	85.18	88.72	81.19	74.84	78.65	82.93	71.12	62.25
Time-Series Transformers	74	70	71	72.03	66	64	61	58.28
Autoencoders	88.29	92.59	85.21	78.94	76.07	77.51	70.95	65.44

below presents the performance of the different algorithms applied.

The results indicate that the CNN 1D classifier achieved the best performance among the tested models, suggesting that this architecture is particularly well-suited for the classification task. In contrast, the Time-Series Transformer classifier showed the weakest performance, which could be due to the increased complexity of this model, making it less efficient in scenarios where temporal data does not require such sophisticated mechanisms. These results suggest that simpler models, like the CNN 1D, can sometimes outperform more complex architectures, especially when the classification task does not require processing long or complex sequences.

5. Deployment of the 1D CNN Model on TensorFlow Lite

Before deploying the 1D CNN model on TensorFlow Lite, several optimizations were applied to improve its performance and efficiency. First, MaxPooling1D layers were integrated after each convolution to reduce feature dimensions while preserving essential information, thereby accelerating training and mitigating overfitting. Additionally, a fourth convolutional layer was added with Conv1d(in_channels = 128, out_channels = 256, kernel_size = 3, stride = 1, padding = 1), enhancing the model’s ability to extract complex patterns. Regarding the activation function, LeakyReLU(0.1) replaced ReLU, helping to alleviate the issue of vanishing gradients. To stabilize training, the model’s output now uses LogSoftmax(dim=1) instead of Softmax, preventing numerical instabilities related to the CrossEntropyLoss() function. In terms of optimization, we adopted the AdamW algorithm with a learning rate of 0.001 and a weight decay of 1 × 10⁻⁴, improving regularization. A ReduceLROnPlateau scheduler was also introduced to dynamically adjust the learning rate based on loss reduction (factor = 0.5, patience = 5). To optimize training, we replaced single-batch data processing with mini-batch training using DataLoader(batch_size = 32, shuffle = True), reducing memory usage and improving convergence. Finally, INT8 quantization was applied during the conversion to TensorFlow Lite with converter.optimizations = [tf.lite.Optimize.DEFAULT], reducing the model size and speeding up execution on embedded devices.

Despite these optimizations, a slight performance decrease was observed, with an F1 score and accuracy of 83% for the binary classification task (distinguishing between stress and non-stress) and 74% for the three-class classification (differentiating between normal, stressed, and amused states). However, these results remain acceptable for prediction, ensuring a good balance between accuracy and efficiency, particularly in a deployment context on mobile or embedded platforms.

To deploy the 1D CNN model on TensorFlow Lite, we follow a multi-step process to ensure compatibility with mobile and embedded devices. First, the PyTorch model is converted into an ONNX (Open Neural Network Exchange) intermediate format using the torch.onnx.export() function, facilitating its import into TensorFlow. Then, the onnx-tf library is used to transform the ONNX model into a TensorFlow-compatible format via prepare(onnx_model).export_graph(). Once this conversion is complete, TensorFlow Lite Converter is used to transform the TensorFlow model into a TFLite model, applying INT8 quantization with converter.optimizations = [tf.lite.Optimize.DEFAULT], which reduces the model size and improves execution speed on resource-constrained devices. After conversion, the model is loaded and executed on a mobile device using tflite.Interpreter(), where input and output tensors are retrieved via get_input_details() and get_output_details(). Finally, interpreter.invoke() is used to perform real-time predictions from new input data. This approach enables efficient deployment of the model on smartphones, IoT devices, and embedded systems, ensuring optimal performance while maintaining minimal resource consumption.

6. Conclusions and Future Work

The research work we conducted primarily aimed at understanding the structure of the WESAD dataset. Crucial steps of data cleaning, normalization, and scaling were performed to ensure the quality, consistency, and relevance of the data used, which helped optimize the performance of the classification models. Among the models tested, the 1D CNN achieved the best performance, highlighting its effectiveness for the classification task, particularly in the context of stress detection from physiological signals. This architecture appears to be particularly well-suited to this type of data, thanks to its ability to extract complex features from time series while remaining computationally efficient, making it ideal for real-time applications.

Moreover, the 1D CNN model proved to be the most suitable for deployment on lightweight platforms such as TensorFlow Lite and microcontrollers. Its low energy consumption and computational efficiency allow for optimized performance while minimizing the energy footprint, a crucial aspect for wearable devices in real-world deployment. These results pave the way for practical applications on wearable devices capable of continuously monitoring stress while ensuring prolonged battery life.

For future work, we plan to expand our analysis by applying hybrid models that combine different deep learning architectures in order to leverage synergies between approaches. Indeed, hybrid models, such as the combination of CNN and LSTM, could enable better capture of temporal dependencies and complex patterns in physiological data. We also aim to explore other datasets with varying characteristics to test the robustness of the models in different contexts. These potential improvements could lead to a significant increase in classification performance. Furthermore, we will deploy other algorithms on TensorFlow Lite and perform energy classification for each algorithm to compare their efficiency in terms of energy consumption and performance on resource-constrained platforms. Thus, future work will focus on improving the accuracy, generalization, and efficiency of the models while exploring the best strategies for optimal deployment on low-resource platforms.