Supervised information gain-based feature selection for multimodal physiological signals in stress prediction

Abstract

Feature selection plays a critical role in identifying the most informative signals across multiple modalities in physiological datasets. While correlation-based approaches have traditionally been used to discover relationships among features, they are often unsupervised and do not incorporate class label information, limiting their discriminative power in supervised classification tasks. In this work, we propose a supervised, Information Gain (IG)-based feature selection framework for multimodal data acquired from wrist-worn sensors, including Blood Volume Pulse (BVP), Acceleration (ACC), Skin Temperature (TEMP), and Electrodermal Activity (EDA).Time–frequency feature representations are extracted from each modality using Short-Time Fourier Transform (STFT) and subsequently ranked according to their correlation with class labels using Information Gain(IG) criterion. The top-ranked features for each modality are then fused into a compact, high-quality representation for classification using Random Forest classifier. Experiments on the WESAD stress recognition dataset demonstrate that our proposed STFT-IG framework along with Random Forest classifier achieves competitive performance under subject-independent evaluation using the Leave-One-Subject-Out (LOSO) protocol. Furthermore, comparative analyses with alternative feature selection methods and feature representations are performed to contextualize the effectiveness of the proposed approach.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-41734-6.

Introduction

In recent years, Internet of Things (IoT) devices have been adopted for research in diverse application domains, particularly in healthcare. Wearable sensors enable continuous monitoring of stress based on physiological signals and behavior. For instance, IoT-enabled healthcare systems that incorporate machine learning, deep learning, and even image-based analysis have demonstrated effectiveness in predicting injury types and identify potential health risks among athletes¹. Continuous monitoring, fewer hospital visits, and prompt intervention are the benefits of integrating IoT and deep learning technologies in medical systems for home environments. These benefits all contribute to better patient care and health outcomes².Further, remote health monitoring can help patients receive personalized feedback where changes in user behavior may signal variations or decline in physical and mental health³⁴. Modern smartphones and wearable devices contain multiple sensors including accelerometers, temperature sensors, heart rate monitors etc. which can record important signals without requiring any modification to surroundings or user behavior. These sensors can function seamlessly in both indoor and outdoor environments and perform real-time monitoring of physiological changes within the body. Physical and mental well-being are both essential components of overall health. According to a survey⁵, many individuals recognize the importance of regular checkups and thorough health examinations, however they avoid getting medical tests due to cost and time requirements. Physiological signals can be used to assess physical health, whereas mental well-being can be inferred through stress and emotional state monitoring. The prevalence of stress is indicated by anxiety and depression, especially for people aged 15 and above⁵. Persistent stress can contribute to various health conditions and negatively impact overall quality of life. Consequently, stress detection using wearable physiological sensors has received increasing research attention⁶. Numerous studies have investigated sensor-based recognition techniques, which include deep learning algorithms and conventional machine learning. Conventional methods rely on handcrafted features in the time, statistical and frequency domains, mapping them to related labels⁷. While handcrafted statistical methods can perform well, they may increase development time, require domain expertise, and depend on user-specific models. In contrast, conventional machine learning models offer reduced computational complexity and faster development, but still rely on careful feature representation and selection. Classifiers like Random Forest(RF), Decision Trees(DT),Support Vector Machines(SVM). Logistic Regression(LR) and Multilayer perceptrons(MLP) have been widely used to analyze vital body signals and predict stress levels. Figure 1 depicts the use of wearable sensors in the collection of important body parameters for analysis of a person’s physical and mental health.

Fig. 1.

Stress monitoring using wearable sensors.

Rationale of the proposed work

The WESAD dataset poses significant challenges for stress recognition due to its multimodal nature, heterogeneous sampling rates and variability among subjects. Existing works on WESAD have used either handcrafted statistical features or deep learning models that achieve high accuracies but fail to generalize when evaluated under subject-independent settings such as LOSO. Moreover, traditional feature representation techniques have limitations in capturing non-stationary changes in physiological signals. Many studies do not address class imbalance in WESAD which directly affects performance and reproducibility. To address these limitations, we propose a generalization-aware methodological framework that integrates three key components: Short-Time Fourier Transform is used to obtain interpretable patterns from physiological signals by providing localized time–frequency representations that preserve transient stress–related changes; Information Gain is used as a feature selection criteria for measuring relationships between class labels and discriminative features; generalization using LOSO for true subject-independent evaluation. The novelty of this work lies in the integration of STFT-based multimodal feature representation with Information Gain–driven feature selection and Random Forest–based classification, evaluated under subject-independent conditions.The combination of these established techniques proves that it is suitable for stress monitoring. Performance is assessed using LOSO validation, and additional experiments on synthetically generated data are conducted in an exploratory manner to examine the applicability and limitations of the proposed approach. This framework offers efficient and interpretable stress recognition pipeline while reducing subject dependence and achieving competitive performance as compared to prior work on WESAD under subject independent evaluation.

The main contributions of the work are as follows:

• The generalization-aware stress recognition framework combines supervised STFT-based feature representation with Information Gain–driven feature selection and evaluation of performance under subject-independent LOSO validation, thereby avoiding bias caused by subject overlap. The approach directly computes Information Gain between features and class labels, making it label-aware and enabling the selection of the most discriminative subsets per modality (BVP, ACC, TEMP, and EDA).

• The use of Short-Time Fourier Transform (STFT) captures both spectral information and temporal dynamics, demonstrating improved efficiency compared to alternative techniques such as Hilbert transform and Quadratic kernel.

• The approach uses Information Gain to select features that show the strongest non-linear relationship with psychological stress labels. The modality-wise ranking of features using IG enhances interpretability, stability, and classification performance across BVP, ACC, TEMP, and EDA modalities.

• The proposed framework explicitly addresses a key limitation in prior WESAD studies i.e subject leakage because of non-independent train-test splitting, by performing subject-independent evaluation. The performance of Random Forest classifier is systematically compared with other classifiers to provide contextual insights.

• An exploratory evaluation using synthetically generated data is conducted to examine the applicability and limitations of the approach.

The remaining paper is structured as follows: Sect. 2 explores the related literature in the domain of stress monitoring. Section 3 details the proposed methodology whereas Sect. 4 describes the experiments and results. Section 5 concludes the paper.

Background and related work

Recent research in stress detection has increasingly shifted toward multimodal physiological sensing and supervised learning frameworks. The most recent studies carried out in the year 2024 and 2025 emphasize deep multimodal architectures. For instance, Xiang et al.⁸ propose an approach which integrates EDA, ST, and HR signals using a CNN-based time–frequency fusion architecture and report  91% accuracy, however the limitations of the same are high model complexity and limited cross-subject generalization. Transformer-based and CNN-Transformer hybrid models have also been introduced for wearable biosignal stress prediction⁹ improving feature representations but limitations include requirement of a large annotated dataset and challenges in application to real-world scenarios. Other approaches implemented focus on multimodal sensor fusion using lightweight classical ML. For instance, Shedage et al.¹⁰ use BVP, HR and EDA from WESAD with decision trees, RF and LR to achieve around 99%, though their model depends heavily on simple hand-crafted features which fail to generalize. Radhika et al.¹¹ propose an autoencoder-based joint modality representation combining ECG, UECG and EDA, but the framework lacks interpretability and is computationally complex. Rashid et al.¹² propose context-aware fusion but require additional contextual sensors which make it unsuitable for real-time wrist-worn systems.

Between 2023 and 2021, several studies explored wearable-only machine learning models. Federated learning–based EDA recognition¹³, CNN-based stress detection with wrist-worn signals¹⁴, and PRV/HRV-based models¹⁵ have demonstrated strong performance, with accuracies upto 98%. Systematic reviews¹⁶ highlight that classical models such as LR, SVM, DT, RF, Naive Bayes, and Gradient Boosting give superior performances especially when using features from EDA, BVP and temperature. However, many works rely predominantly on summary statistics such as mean, min, max, and STD, while features such as skewness and kurtosis show limited utility for stress classification¹⁷. These findings motivate the need for richer time–frequency feature representations.

Earlier multimodal studies (2020–2019) involved heavy laboratory setups incorporating EEG, EMG, EOG, ECG, respiration, GSR and BVP signals¹⁸, or stress datasets collected with wearable sensors and¹⁹. Reported accuracies range from 70 to 95%, but these systems suffer from high sensitivity to changes in artifacts and require complex sensor configurations. Classic deep learning studies–including CNN architectures for multichannel time series²⁰, RNN-based feature extraction²¹, and LSTM models²² offer powerful temporal modeling yet suffer from overfitting, and lacking interpretability.

Deep multimodal architectures such as CNN-based fusion models⁸ and Transformer-based networks⁹ offer strong accuracy but suffer from high computational cost, large training-data requirements, poor interpretability, and limited cross-subject generalization. These models often require very large datasets and struggle with real-world noise and subject variability. Autoencoder-based representation learning¹¹ and context-aware fusion frameworks¹² further increase model complexity and require additional contextual sensors.

On the other hand, several classical ML works on WESAD and related datasets including decision trees, random forests, and logistic regression¹⁶¹⁰ rely on simple summary statistics (mean, min, max, standard deviation). These features are fast to compute but may not capture spectral variations in physiological signals. Prior studies have already shown that methods such as skewness and kurtosis contribute little to improving stress prediction¹⁷, proving the need to design richer signal representations.

A further limitation across many existing studies is the lack of a supervised, discriminative feature-selection mechanism. Most approaches employ either unsupervised correlation-based filtering or model-specific feature-ranking, which do not explicitly exploit class label information and are sensitive to class imbalance and subject variability.

Therefore, the motivation for our work arises from the need to combine:

• Richer time–frequency representations which can interpret physiological variations.

• Supervised feature selection to rank features using class-label information.

• Multimodal fusion across BVP, ACC, EDA, and TEMP signals.

• Subject-independent evaluation to ensure generalizability across individuals

• Lower computational overhead compared to deep learning frameworks.

To address these gaps, we propose a Short-Time Fourier Transform (STFT) based multimodal feature representation pipeline, followed by Information Gain (IG) driven supervised feature selection. This enables the model to retain only the most discriminative features from each physiological modality. The resulting compact representation is then used with Random Forest, which gives competitive performance with significantly reduced computational cost. Additionally, we compare the IG approach against several traditional feature-selection methods to demonstrate its effectiveness. The proposed framework is compared with other classifiers and cross-validation is done using LOSO technique.

Feature selection techniques

This section explores the most popular feature selection techniques which can be used to rank important features before classification using ML models.

Filter methods

Filter methods evaluate each feature independently with respect to the target variable. The features with high correlation are selected as they lead to improved prediction performance. Due to their low computational cost, these methods are widely used for high-dimensional physiological datasets. Information Gain measures the reduction in entropy achieved by knowing the value of a feature. For a feature f and a class label , the IG is defined as:

where denotes Shannon entropy. A higher IG value indicates that the feature carries stronger discriminative information about the class labels.

The selected feature subset for modality i is:

where is a predefined threshold. Alternatively, the top-K features may be chosen by ranking features in decreasing IG order.

Other techniques include Chi-Square test which measures the dependency between a feature and the target variable. The Chi-Square statistic is defined as:

where and represent observed and expected frequencies for the i-th category. Larger values indicate more informative features. Fisher Score evaluates the discriminatory ability of each feature independently. Larger score indicates better class separation. Pearson correlation coefficient measures the degree of linear association between two variables X and Y:

where is covariance, and , are standard deviations. Other methods such as ANOVA F-test evaluates whether a numerical feature exhibits significant variance between classes compared to within classes. Variance-based techniques (Variance Threshold, Mean Absolute Deviation, Dispersion Ratio) rely on dispersion statistics and do not utilize class labels.

Wrapper methods

Wrapper methods evaluate subsets of features by repeatedly training a predictive model and selecting the combination that yield the best performance. Common wrapper techniques include forward selection which starts with empty set of features and iteratively adds the most informative feature at each step, and backward selection which begins with all features and iteratively removes the least informative feature. On the other hand, Recursive Feature Elimination (RFE) removes the least important feature and repeats until the desired subset is obtained.

Embedded methods

Embedded methods integrate feature selection into the model training process. A representative example is L1-regularized regression (LASSO), which introduces an L1 penalty that drives less informative feature coefficients toward zero, resulting in a sparse feature subset. Other techniques include Gradient Boosting, Tree-based modeling etc.

Canonical correlation analysis based supervised feature selection

To exploit inter-modal relationships among physiological signals, Canonical Correlation Analysis (CCA) is explored as a correlation-based dimensionality reduction technique A two-stage hybrid technique is followed in which CCA is first applied to generate correlated multimodal representations, followed by supervised feature ranking using Information Gain.The goal is to exploit cross-modal relationships while still selecting the most discriminative features for stress classification.

Time-frequency feature extraction using STFT

Physiological signals such as BVP, ACC, ST, EDA are non-stationary signals. In order to capture their temporal and spectral characteristics, we employ the Short-Time Fourier Transform (STFT) technique described as follows:

1

where is the signal from modality i, is a window function, and is the resulting time-frequency representation. Using these representations, the statistical descriptors such as mean, variance, entropy, peak frequency etc are calculated to obtain feature sets for each modality. From this representation, we compute statistical parameters which contribute significantly to each modality. Table 1 compares several feature selction techniques in terms of their strengths and limitations.

Table 1.

General Comparison of Feature Selection Techniques.

Method	What it Measures	Strengths	Limitations
Information Gain	Reduction in entropy and non-linear dependence between feature and class	Fast and works well for multimodal physiological signals	Slight bias toward continuous features; depends on discretization scheme
Chi-Square Test	Statistical dependency between categorical variables	Simple, fast, interpretable	Less suitable for continuous sensor data
ANOVA F-test	Ratio of between-class to within-class variance	Efficient for continuous features	Assumes linear separability and ignores feature interactions
Pearson Correlation	Linear correlation between feature and label	Extremely fast	Misses complex physiological patterns
Variance Threshold	Removes low-variance features	Fastest method	Ignores class labels
LASSO (L1 Regularization)	Embedded selection	Produces sparse, interpretable models	Removes correlated features
RFE (Recursive Feature Elimination)	Iteratively removes least-important features using a classifier	Captures interactions	Computationally expensive
Canonical Correlation Analysis (CCA)	Extracts latent variables that maximize correlation	Suitable for multimodal fusion	Not a supervised feature selector; can be combined to produce ranked features

Proposed methodology

To evaluate the performance of the proposed STFT–IG framework, a series of experiments are conducted on the WESAD dataset, which contains multimodal physiological signals acquired from wrist-worn sensors, including BVP, ACC, TEMP, and EDA. The results are compared against widely used feature-selection techniques such as the ANOVA F-test, Chi-square test, Pearson correlation, and supervised Canonical Correlation Analysis (CCA) etc. The time-frequency features are derived from each modality using the Short-Time Fourier Transform (STFT) with a fixed window size of 30 seconds. The selection of 30-second sliding windows for feature representation is a choice motivated by prior physiological stress-recognition literature, where stress-related autonomic responses shift typically require 20–60 seconds. A 30-second window provides a balance between physiological stability ensuring that short fluctuations do not affect the features. In the study conducted by the authors in²³, the effectiveness of different window sizes for stress detection using the WESAD dataset was systematically examined. It was observed that shorter window sizes upto 30 seconds were more sensitive to transient stress responses and consistently outperformed longer windows. The results showed that Decision Tree model achieved better accuracy with 10–30 second windows than with windows exceeding 90 seconds. The study also highlighted that 30 and 60-second windows are most commonly used in the literature based on the assumption that physiological stress responses unfold within these durations. Segmentation of sensor signals using a sliding window of 30 seconds is also based on the studies of^24–26 and²⁷. While longer window durations allow the model to account for trends in physiological signals that may contain relevant information about the subject’s emotional state, shorter window durations on the other hand may result in the loss of important temporal relationships and patterns. While Longer durations may increase the computational load and memory requirements, and processing time of the model. This is especially critical for real-time applications, which require efficient processing. Consequently, a 30-second window offers an effective trade-off between temporal resolution and computational efficiency. Additionally, selecting the top 50 features helps to preserve variability caused due to all four modalities (BVP, ACC, EDA, TEMP). It thus maintains a balance between preserving discriminative information and avoiding overfitting. In order to find out the best value of k among multiple feature subset sizes (10, 20, and 50), experiments were conducted on WESAD dataset without subject-independent evaluation, as shown in Appendix A.1. It was observed that the 50-feature configuration consistently produced better results across subjects. Therefore, the top-50 IG-based features represent the most stable and discriminative combination for stress recognition. To increase sample density while maintaining temporal coherence, we apply a 50% overlap, which is a consistent practice in time–frequency signal processing. Overlapping windows reduce the risk of missing short events while preserving independence between samples.For subject-level distribution, each participant contributes a proportional number of windows based on their recorded duration. In LOSO evaluation, the samples of the test subject are completely isolated, guaranteeing independence between training and testing data. From these representations, statistical descriptors such as mean, variance, entropy, and dominant frequency are computed to generate compact feature vectors. Subsequently, the Information Gain algorithm is applied to each feature to measure its discriminative power with respect to the class labels. In the classification stage, we employ the Random Forest classifier as it is robust and has minimal requirements for hyperparameter optimization. The results achieved in terms of standard performance metrics such as Precision, Recall, F1 score and accuracy evaluate the stress prediction performance and also evaluate the discriminative ability of the proposed feature selection framework. For completeness, comparative results are also reported for Decision Trees (DT), Support Vector Machines (SVM), and Multilayer Perceptrons (MLP). The overall methodology is illustrated in Fig. 2.

Fig. 2.

Proposed Methodology.

Dataset description

The dataset used for stress monitoring and prediction is the WESAD (Wearable Stress and Affect Detection) dataset. This multimodal dataset contains physiological signals recorded using wrist-worn and chest-worn wearables. Data was collected from 15 subjects during a controlled laboratory study. The following sensor modalities are included: blood volume pulse (BVP), electrodermal activity (EDA), body temperature (TEMP), and three-axis acceleration (ACC). The dataset bridges the gap between previous lab studies on stress and emotion, by providing recordings of three distinct affective states: neutral (baseline), stress, and amusement. In addition, self-reports of the subjects were obtained using several established questionnaires. Raw sensor data in WESAD is recorded with two devices: a chest-worn device (RespiBAN) and a wrist-worn device (Empatica E4). The RespiBAN device provides the following signals: electrocardiogram (ECG), electromyogram (EMG), respiration, body temperature, and three-axis acceleration, all sampled at 700 Hz. The Empatica E4 device provides the following signals: blood volume pulse (BVP, 64 Hz), electrodermal activity (EDA, 4 Hz), body temperature (4 Hz), and three-axis acceleration (32 Hz). For experimentation in this work, the data collected using Empatica E4 was pre-processed and employed.

The dataset consists of the following experimental conditions:

• Baseline 20-minute period of standing/sitting and reading magazines.

• Amusement Subjects watched a set of eleven funny video clips.

• Stress Trier Social Stress Test (TSST), consisting of public speaking and mental arithmetic.

Table 2 depicts the features in the WESAD dataset.

Table 2.

Processed features after data collection.

Feature	Description
Label	Stress level class assigned to sample
Subject ID	Identifier of individual subject
Timestamp	Time of recording the activity
Motion (ACC)	Three-axis acceleration of subject
Blood Volume Pulse (BVP)	Used to estimate heart rate and Inter-Beat Interval (IBI)
Body Temperature (TEMP)	Current body temperature of subject
Electrodermal Activity (EDA)	Electrical conductance of the skin

Experiments and results

All continuous features in the WESAD dataset were standardized before model training using z-score normalization independently for every feature dimension as per the formula:

where and denote the mean and standard deviation computed over the training set. This ensures that all modalities are on a comparable scale and prevents classifiers from being biased towards certain physiological signals.

Furthermore, the class imbalance in WESAD was handled using balanced class weights. For each classifier, the weight assigned to class c was computed as:

where N is the total number of samples, C is the number of classes, and is the number of samples belonging to class c. These weights were passed to all supervised models during training to reduce bias toward majority classes and ensure reliable performance.

Table 3 depicts the important parameters for time-frequency feature representation using the proposed STFT-IG approach.

Table 3.

STFT Parameter Configuration.

Parameter	Value
Window Type	Hamming Window
Window Length	30 seconds
Hop Size (Overlap)	50%
FFT Size	256 (ACC, ST); 512 (BVP, EDA)
Sampling Rate (ACC)	32 Hz
Sampling Rate (BVP)	64 Hz
Sampling Rate (EDA)	4 Hz
Sampling Rate (TEMP)	4 Hz
Frequency Resolution	Determined by FFT size and sampling rate
STFT Features per Modality	Magnitude spectrogram statistics (mean, variance,
	Spectral entropy, band-power features)
Final Selected Features	Reduced to 50 using Information Gain

The application of STFT to each window helps to obtain a spectrogram. From this spectogram, statistical descriptors are obtained. The distributions of STFT features across the three class labels (Baseline, Stress, and Amusement) are depicted using kernel density estimation (KDE) in Fig. 3. For each modality, the first three retrieved features are plotted. These STFT-based descriptors include statistical measures such as mean, variance, entropy, and frequency-domain characteristics as obtained from the spectrograms. The resulting plots provide insight into how the feature values are distributed across the different class labels. It is observed that certain features, particularly from EDA and ST, exhibit highly skewed distributions, whereas BVP and ACC features demonstrate broader variability but substantial overlap between classes. This indicates that all features are not discriminative on their own, thus use of supervised feature selection methods such as Information Gain (IG) or the F-test is important to identify the most relevant features for stress prediction.

Fig. 3.

Kernel density estimates (KDE) plots for feature ranking of STFT-extracted features across three class labels.

This is followed by feature selection using Information Gain measure which identifies the top 50 features from the STFT derived features. Other statistical and correlation tests have also been used for the purpose of feature selection. The next step involves training of ML classifiers such as RF, SVM, DT and MLP for stress level prediction.

Leave one subject out cross validation and cross generalization

The WESAD dataset is explicitly designed for subject-independent stress detection, meaning that a trained model should generalize to new, unseen individuals rather than only to the subjects included in training. Physiological responses are inherently subject-specific, and random train–test splits often lead to data leakage, as samples from the same subject share baseline patterns and sensor characteristics. This can result in inflated accuracy and misleading conclusions.

For these reasons, LOSO cross-validation is followed for WESAD in the proposed work. Under LOSO, the model is trained on all subjects but one subject and tested on the held-out subject. This ensures true subject-independent evaluation and removes overlapping of subject-specific patterns between training and testing data. Although LOSO yields lower performance than random splits, but it provides a more accurate assessment of how the system will behave on new users, which is critical for wearable stress recognition applications. Random Forest is the best-performing model under LOSO for the WESAD dataset, offering the strongest subject-independent generalization and the most stable performance across participants.

Table 4 depicts the performance of proposed framework with cross validation using LOSO framework.

Table 4.

LOSO Evaluation Results for Random Forest Classifier on WESAD Dataset.

Subject	Accuracy	Macro-F1	Precision	Recall
1	0.717857	0.515860	0.552670	0.534979
2	0.646853	0.626195	0.641159	0.663842
3	0.825784	0.618919	0.586207	0.662698
4	0.750853	0.551830	0.537754	0.583500
5	0.589041	0.443728	0.441012	0.489631
6	0.752577	0.543964	0.495358	0.616034
7	0.763699	0.568376	0.571429	0.590909
8	0.893103	0.855911	0.870730	0.882602
9	0.820598	0.606526	0.573323	0.649881
10	0.750000	0.603252	0.625539	0.605064
11	0.765306	0.565754	0.564897	0.593870
12	0.581633	0.335234	0.519713	0.389513
13	0.901695	0.866694	0.879994	0.894160
14	0.9000000	0.900000	0.900000	0.900000
15	0.714286	0.499420	0.458725	0.549313

The performance under the Leave-One-Subject-Out (LOSO) scheme decreases and ranges between 75% and 90%. Prior literature on WESAD shows an accuracy between 60 to 80% with LOSO evaluation using machine learning and deep learning models²⁸. In summary, the non-LOSO evaluations reflect within-subject performance, whereas LOSO reflects true cross-subject generalization to unseen individuals. Therefore, the highest LOSO accuracy achieved is around 90.01% obtained in this study is both consistent with existing research and represents real-world deployment constraints. To compare results on WESAD dataset under 80:20 train-test split,additional experiments are conducted and results are summarized in Supplementary Information (Table S1 and Table S2).

Exploratory analysis on synthetically generated data

Further, exploratory analysis was conducted using the proposed model on a synthetically generated dataset to assess its generalization capability. A controlled synthetic dataset that mimics the structure of WESAD was generated with total of 15 synthetic subjects , each containing 1,500 samples (500 samples per class for baseline, stress and amusement), yielding 22,500 rows overall. For each subject, synthetic physiological signals (ACC, BVP, EDA and TEMP) were produced using Gaussian distributions with class-specific mean and variance shifts. Subject-level variability was introduced by applying random biases and scaling factors. Feature representations were then derived using STFT-based spectral features. The results indicate a substantial drop in performance, with an overall accuracy of 49.67%. This degradation is expected because generated data typically differs from WESAD in several critical aspects, including signal dynamics,temporal structure, and stress related distribution. Since the model was trained exclusively on WESAD data, its learned boundaries do not fully transfer to synthetically generated input distributions, leading to mismatches in feature statistics and weaker discriminative power. The low accuracy is attributed to machine learning models learning dataset-specific features and not generic stress features.The synthetically generated stress responses may be different. All these factors could influence the low generalization capabilities of the proposed approach on synthetic dataset. Table 5 depicts the performance metrics achieved by trained model which is tested on synthetically generated data as a part of the exploratory analysis.

Table 5.

Comparison of Model Performance on WESAD and Generated Data.

Dataset	Accuracy	Macro F1
WESAD (STFT + IG + RF)	0.759	0.59
Generated Data ((STFT + IG + RF)	0.497	0.466

Although the model’s performance on the generated dataset is noticeably lower than on WESAD, the results are still meaningful for exploratory purposes. The model was trained only on WESAD, not on the new domain. WESAD is a clean, controlled laboratory dataset collected under standardized stress protocols. In contrast, the generated dataset resembles unseen, noisy, and real-world physiological patterns. Generated data lacks exact temporal, spectral, or sensor characteristics of WESAD. Even though STFT + IG features were extracted, the underlying physiological dynamics differ in many aspects. Even with domain shift, some high-entropy features still separate stress vs non-stress across subjects.

Performance comparison using feature selection methods

Table 6 reports the classification performance using Random Forest classifier for different feature selection methods applied on the STFT features extracted using WESAD dataset under LOSO protocol. It is observed that Information Gain(IG) achieved the highest performance, reaching an accuracy of 75.9% with Top-50 features (as depicted in Table 5). Pearson correlation also demonstrated strong discriminative power, yielding 75.4% accuracy at Top-50 features. In contrast, the F-test and Chi-square achieved an accuracy of 60.1% and 59.5% respectively. The Canonical Correlation Analysis technique combined with IG gives an accuracy of 70% for Top-50 features. For all CCA blocks, the number of canonical components was fixed at five. The pairwise projected components (from ACC–BVP, BVP–TEMP, TEMP–EDA, and EDA–ACC) were concatenated to form a unified fused feature matrix. The solver used for SVD-based and no penalty was applied. Before applying CCA, each modality was standardized using z-score normalization to ensure equal scaling in the canonical space. Overall, IG emerged as the most effective feature selection strategies for stress prediction using STFT-based physiological features. Table 7 summarizes the mean performance obtained using the different feature selection techniques.

Table 6.

Subject-wise LOSO accuracy comparison of different feature selection strategies combined with Random Forest (RF) classifier on the WESAD dataset.

Subject	ANOVA	Chi-square	Pearson	CCA + IG
1	0.682	0.700	0.829	0.721
2	0.503	0.531	0.731	0.693
3	0.760	0.725	0.829	0.758
4	0.590	0.638	0.803	0.702
5	0.695	0.555	0.781	0.675
6	0.557	0.485	0.739	0.648
7	0.692	0.575	0.798	0.738
8	0.752	0.710	0.772	0.701
9	0.472	0.625	0.721	0.662
10	0.608	0.696	0.838	0.745
11	0.571	0.568	0.759	0.713
12	0.531	0.534	0.503	0.59
13	0.610	0.566	0.749	0.728
14	0.840	0.823	0.769	0.781
15	0.150	0.193	0.691	0.602
Mean Accuracy	0.601	0.595	0.754	0.70

Table 7.

Mean LOSO accuracy comparison of different STFT-based feature selection strategies combined with Random Forest classifier.

Method	Mean Accuracy(%)
STFT + ANOVA + RF	60.09
STFT + Chi2 + RF	59.49
STFT + Pearson + RF	75.4
STFT + IG + RF	75.9
STFT + CCA + IG + RF	70.8
STFT + LASSO(L1 Regularization) + RF	71.2
STFT + Variance Threshold + RF	71.6
STFT + Recursive Feature Elimination + RF	73.5

Figure 4 depicts the most discriminative features which are correlated with the class labels using the STFT-IG framework.

Fig. 4.

Top-50 discrimminative features extracted using STFT-IG framework.

The analysis of top-50 Short-Time Fourier Transform (STFT) features, selected using Information Gain and evaluated with a Random Forest classifier, highlights the dominance of blood volume pulse (BVP) descriptors. Features related to BVP emerge as the most discriminative, showing the strong contribution of heart rate variability in stress prediction. Temperature-related features also rank highly, confirming the role of thermoregulatory changes under stress conditions. EDA features appear consistently across the middle ranks, indicating their complementary role in differentiating stress states. Accelerometer-based descriptors provide less dominant contributions, primarily reflecting motion differences during amusement versus stress conditions. Overall, the distribution of feature importance is skewed towards BVP and ST modalities, establishing them as the most reliable indicators for stress detection within the WESAD dataset, while EDA and motion features act as supportive modalities.

Performance comparison with other classifiers

The performance of the proposed STFT+IG framework was further evaluated by training three different classifiers, SVM,DT and MLP using the Top-50 features selected through Information Gain. The results indicate that the Information Gain-based feature selection combined with Random Forest classifier provides a highly discriminative representation of the STFT-extracted features from the WESAD dataset. Each classifier was trained with tuned hyperparameters that balance performance and computational efficiency. The Support Vector Machine (SVM) employed the radial basis function (RBF) kernel with . The Decision Tree (DT) classifier was trained with a maximum depth of 10 to avoid overfitting while capturing non-linear patterns in the feature space. The Multi-Layer Perceptron (MLP) was configured with two hidden layers of sizes (100, 50), ReLU activation, and the Adam optimizer, trained for a maximum of 500 iterations with a dropout rate of 0.5 to improve generalization. Table 8 shows the hyperparameters of all the ML classifiers used in the framework after extraction of time-frequency features using STFT.

Table 8.

Hyperparameter settings for classifiers.

Classifier	Hyperparameters
SVM	Kernel = RBF, ,
Logistic Regression	max_iter = 1000
MLP (Neural Net)	Hidden layers = (100, 50), Activation = ReLU, Optimizer = Adam,
	Max iterations = 500, Dropout = 0.5, random_state = 42

Table 9 depicts the performance of all classifiers under LOSO cross validation, with reported mean accuracies, precision, recall and F1 score.

Table 9.

Comparison of Classifier Performance under LOSO Evaluation on WESAD Dataset.

Classifier	Accuracy (Mean)	Accuracy (Std)	F1 (Mean)	F1 (Std)
Random Forest	0.759	0.1014	0.5992	0.1543
SVM (RBF Kernel)	0.6388	0.1757	0.5417	0.1475
MLP	0.6561	0.1826	0.5282	0.1538
Logistic Regression	0.6709	0.1395	0.6130	0.1384

Performance comparison of STFT with Hilbert Transform and Quadratic Kernel Techniques

The Hilbert Transform is widely used to derive signals which provide instantaneous amplitude, phase, and frequency information. These components are particularly relevant for physiological stress detection, for capturing changes in EDA, BVP, and other signals.Hilbert features capture fine-grained oscillations and small fluctuations in physiological signals that time-domain summary statistics often miss.These features perform well when signals vary quickly, unlike windowed Fourier methods which rely on fixed time-frequency trade-offs. In contrast, quadratic kernel mapping projects the original feature space into a second-order polynomial space, enabling classifiers such as Random Forest to model nonlinear relationships between features. By explicitly capturing pairwise feature interactions, quadratic mappings allow the classifier to exploit second-order dependencies that may be relevant for stress discrimination. Table 10 depicts the difference in accuracies achieved using Hilbert transform, Quadratic kernel and STFT techniques using LOSO evaluation protocol.

Table 10.

Comparison of Feature Representation Techniques Under LOSO Settings.

Method	With LOSO
Hilbert + IG + RF	68.76%
Quadratic Kernel + IG + RF	60.52%
STFT Features + RF	75.9%

Short-Time Fourier Transform (STFT) consistently outperforms both Hilbert Transform and quadratic-kernel–expanded features because it captures stable band-level features, suppresses subject specific noise an captures global frequency domain patterns that generalize well across subjects. In contrast, Hilbert Transform features are highly sensitive to individual physiological differences, sensor placement, and noise. On the other hand, quadratic kernel method improves non-linear separability but it fails to compensate for lack of explicit time–frequency structure in the input features. Overall, STFT provides a more robust representation of multimodal physiological dynamics, offering superior cross-subject generalization and emphasizing on frequency-domain stability for stress recognition.

Performance comparison with existing studies using LOSO evaluation

Early frequency-spectrum–based approaches²⁹ rely on handcrafted spectral features extracted from physiological signals. While these methods demonstrate promising subject-dependent accuracies exceeding 90%, their LOSO performance typically drops to the 65–75% range, indicating limited cross-subject robustness.The work done in³⁰ revealed that the simpler, non-temporal Logistic Regression baseline model achieved a final aggregated accuracy of 74.67% on WESAD using LOSO, outperforming the more complex LSTM model, which achieved an aggregated accuracy of 73.33%.

In comparison, the proposed framework achieves a mean LOSO accuracy of up to 75.9% on multimodal sensor data, demonstrating competitive performance compared to prior work while maintaining computational efficiency. The improvement can be attributed to the use of supervised feature selection via Information Gain, which helps retain discriminative, subject-invariant time–frequency descriptors and reduces degradation for subjects with high variability. Table 11 provides a comparative summary of LOSO results from related studies.

Table 11.

Comparison with Previous Subject-Independent (LOSO-Based) Stress Recognition Studies.

Study	Modalities	Method	LOSO Accuracy (%)
31	Wrist modalities	Adaboost
29	Wrist modalities	Adaboost	67
32	Wrist modalities	Convolutional Autoencoder + CNN
30	Wrist modalities	Lightweight ML models	74.67
Proposed Method	Wrist modalities	STFT+ IG+ RF	75.9

Conclusion

In this paper, we propose a supervised multimodal feature- selection framework that integrates Short-Time Fourier Transform (STFT)–based time–frequency representations with Information Gain (IG)–based feature ranking. By incorporating label-aware feature evaluation, our approach identifies the most discriminative descriptors from wrist-worn physiological signals obtained from WESAD dataset such as BVP, EDA, ACC, and TEMP. The selected features are then fused into a compact representation and classified using several machine learning models. Experimental results under the Leave-One-Subject-Out (LOSO) evaluation protocol show that the proposed STFT–IG framework, combined with a Random Forest classifier, demonstrates competitive performance compared to prior work under subject-independent evaluation. The analysis further compares different feature representation and feature selection strategies, highlighting the effectiveness of the proposed approach.

Overall, the findings indicate that the proposed approach proves generalizable for stress recognition across individuals. Future work will explore nonlinear correlation measures and deep representation learning to better capture complex multimodal dependencies.

Author contributions

S.A.C has worked on the manuscript and implemented the entire methodology

Funding

This work has been done as a part of SEED grant funding given by Dr BR Ambedkar National Institute of Technology, Jalandhar.

Data availability

The WESAD dataset used in this study is publicly available from the UCI Machine Learning Repository: https://archive.ics.uci.edu/dataset/465/wesad+wearable+stress+and+affect+detection To ensure reproducibility of the proposed methodology, the WESAD data used, STFT feature extraction script and the dataset comprising derived features, other signal processing modules (Hilbert transform), feature selection techniques, and LOSO evaluation code developed in this study are publicly available on Zenodo at: https://doi.org/10.5281/zenodo.18608743

Code availability

All code and scripts are publicly available in a DOI repository on Zenodo: https://doi.org/10.5281/zenodo.18608743

Declarations

Competing of Interests

There is no conflict of interest.