FDSDS: A Fuzzy-Based Driver Stress Detection System for VANETs Considering Interval Type-2 Fuzzy Logic and Its Performance Evaluation

Abstract

Vehicular Ad Hoc Networks (VANETs) enable Vehicle-to-Vehicle (V2V) and Vehicle-to-Infrastructure (V2I) communications for enhancing road safety. However, reliable driver stress assessment remains challenging due to noisy sensing, inter-driver variability, and context dynamics. This paper proposes a Fuzzy-based Driver Stress Detection System (FDSDS) that employs an Interval Type-2 Fuzzy Logic System (IT2FLS) to model uncertainty. The FDSDS considers four complementary inputs—Heart Rate Variability (HRV), Galvanic Skin Response (GSR), Steering Angle Variation (SAV), and Traffic Density (TD)—to estimate Driver Stress Level (DSL). Extensive simulations (14,641 test points) show monotonic associations between DSL and the inputs, which reveal that physiological indicators dominate average influence (finite-difference sensitivity: GSR 0.357, SAV 0.239, TD 0.239, HRV 0.235). Under severe physiological conditions (HRV = 0.1, GSR = 0.9), the system consistently outputs high stress (mean DSL = 0.813; range 0.622–0.958), while favorable physiological conditions (HRV = 0.9, GSR = 0.1) yield low stress even in challenging traffic (range 0.044–0.512). The IT2FLS uncertainty bands widen for intermediate conditions, aligning with the inherent ambiguity of moderate stress states. These results indicate that combining physiological, behavioral, and environmental factors with IT2FLS yields interpreted, uncertainty-aware stress estimates suitable for real-time VANET applications.

1. Introduction

Vehicular Ad Hoc Networks (VANETs) have become a key component of modern Intelligent Transportation Systems (ITS), enabling Vehicle-to-Vehicle (V2V) and Vehicle-to-Infrastructure (V2I) communications to improve road safety and traffic efficiency [1]. Recent developments in VANET technologies have expanded the possibilities for developing applications dedicated to remote monitoring, traffic management, and safety enhancement through Vehicle-to-Everything (V2X) communications [2]. These advances leverage IEEE 802.11p protocol frequency channels and 5G connectivity to support real-time data exchange critical for ITS applications [3]. However, despite these technological advances, traffic congestion and road accidents continue to become major challenges worldwide. The World Health Organization reports approximately 1.19 million road traffic deaths annually (estimate for 2021) [4].

Driver stress has emerged as a critical factor in transportation safety, with research demonstrating its significant impact on driving performance, reaction time, and decision-making ability. Naturalistic-driving studies report odds ratios of up to approximately ten for certain high-stress states [5], indicating that stressed drivers face substantially elevated crash and near-crash risks. Investigations into multimodal driver stress detection have revealed that stress is among the 10 leading causes of fatal crashes, emphasizing the urgent need for reliable, real-time stress monitoring systems integrated in vehicular networks [6]. Furthermore, driver monitoring systems have gained increasing attention as effective countermeasures for detecting and mitigating hazardous driver states before accidents occur [7].

Various approaches have been proposed to detect driver stress, ranging from physiological monitoring using Heart Rate Variability (HRV) and Galvanic Skin Response (GSR) [8,9] to behavioral analysis through steering wheel patterns [10] and vehicle context dynamics [11]. Fuzzy Logic (FL)-based systems have demonstrated promising results due to their ability to handle imprecise data, with some Type-1 Fuzzy Logic Systems (T1FLS) achieving up to 99% accuracy in controlled settings [12]. Machine Learning (ML) and Deep Learning (DL) approaches have also shown strong performance, with multimodal convolutional neural networks achieving accuracies exceeding 95% by combining short-term physiological signals [13]. However, most existing systems have limitations in handling the uncertainty and variation inherent in real-world driving environments. Traditional binary classification approaches and T1FLS often fail to adequately model the complex relationships between multiple stress indicators and their varying degrees of uncertainty for different drivers and driving conditions. The gap between laboratory accuracy (often exceeding 90%) and real-world performance (typically 70–85%) highlights the need for more robust uncertainty modeling approaches [6].

This paper introduces a Fuzzy-based Driver Stress Detection System (FDSDS) using an Interval Type-2 Fuzzy Logic System (IT2FLS) for stress evaluation in VANET environments. The proposed system considers four parameters: HRV, GSR, Steering Angle Variation (SAV), and Traffic Density (TD). By using IT2FLS, our system addresses the limitations of existing approaches by handling uncertainty and providing more reliable stress detection under varying driving conditions. To the best of our knowledge, there is not any other approach that combines these three categories of characteristics: physiological, behavioral, and environmental.

The contributions of this paper are as follows.

• We propose an IT2FLS-based framework that integrates physiological, behavioral, and environmental parameters for driver stress detection.
• We show the effectiveness of the proposed IT2FLS architecture in capturing the synergistic effects of combining HRV and GSR as physiological indicators with SAV and TD data available through VANET communications.
• We provide detailed simulation results showing the system’s performance under different driving scenarios and driver stress level.

The remainder of this paper is organized as follows. Section 2 reviews related work in driver stress detection and FL applications. Section 3 provides an overview of FL principles, particularly IT2FLS. Section 4 details the proposed FDSDS and implementation. Section 5 presents simulation results and performance evaluation. Finally, Section 6 concludes the paper and discusses future research directions.

2. Related Work

Physiological approaches for driver stress detection have received considerable attention due to their direct connection with Autonomic Nervous System (ANS) responses. Rastgoo et al. [6] conducted a comprehensive critical review of techniques for multimodal driver stress detection. They show that stress is among 10 leading causes of fatal crashes. Healey and Picard [8] used multiple physiological signals, achieving 97% accuracy by combining Electrocardiography (ECG), Electromyography (EMG), GSR, and respiration data. However, their approach requires extensive sensor arrays that may be impractical for everyday driving scenarios. More recently, Liu et al. [9] and Castaldo et al. [14] investigated ultra-short HRV analysis (30–180 s) for stress detection. They showed that a good accuracy is achievable with ≤3 min windows, which improves real-time applicability.

Vehicle-based measurements offer non-intrusive alternatives for stress detection. Balters et al. [10] proposed a novel steering-wheel-based approach using signals from an unmodified steering wheel to detect driver stress. While achieving approximately 77% accuracy with minimal sensor requirements, this method alone cannot capture the full complexity of stress responses. Warnecke et al. [15] advanced this approach by attaching printed, flexible polyurethane ECG electrodes to the steering wheel, enabling ECG-based Heart Rate (HR) monitoring for approximately 45.62% of driving time. Other studies utilizing acceleration patterns, brake pressure, and lane-change trajectory have shown promising results. However, they depend heavily on driving conditions and individual driving styles [11].

Multimodal sensor fusion approaches have demonstrated superior performance compared to single-modality methods. Rigas et al. [16] detected driver stress events in real time by fusing 10 s physiological features with driving-event cues in Bayesian networks. They improved the accuracy from 82% (physiology only) to 96% with event information and online adaptation. Memar and Mokaribolhassan [17], by using a single foot GSR amplitude feature with an Analysis of Variance (ANOVA)-based statistical classifier, achieved 95.83% accuracy for three-level stress classification. Iqbal et al. [18] presented a wrist-worn Photoplethysmogram (PPG)-based pilot study and the Stress-Predict dataset collected under laboratory protocols, highlighting the need and challenges for wearable continuous stress monitoring. Can et al. [19] conducted real-life, continuous stress detection with wrist-worn sensors and underscored the difficulty of maintaining accuracy in unconstrained settings.

ML/DL approaches can be used for stress classification. Rastgoo et al. [20] developed a Convolutional Neural Network (CNN)–Long Short-Term Memory (LSTM)-based multimodal fusion model that fuses ECG, vehicle data, and contextual information. They achieved 92.8% average accuracy and outperformed traditional ML methods with handcrafted features. Lee et al. [13] proposed a multimodal CNN that uses continuous recurrence-plot representations of short-term (10–30 s) foot and hand GSR and HR signals, achieving 95.67% accuracy for two-class stress classification. Amin et al. [21] combined CNN and hybrid CNN–LSTM fusion models with a fuzzy-based approach for evaluation, demonstrating automatic feature extraction capabilities of DL. These approaches, while achieving high accuracy in controlled environments, often require substantial computational resources and large training datasets, which may limit their applicability in resource-constrained VANET environments.

FL systems have been applied for stress detection due to their ability to handle imprecise and uncertain data while maintaining interpretability. De Santos Sierra et al. [12] achieved very high accuracy (99.5%) using T1FLS with GSR and HR data. However, T1FLS is limited in its ability to model uncertainty in Membership Functions (MFs). Recent advances in Type-2 FL, particularly IT2FLS, offer improved capabilities for handling higher-order uncertainties [22]. Li et al. [23] introduced a Type-2 fuzzy LSTM that improves long-term traffic prediction accuracy and interpretability under uncertainty. Unlike DL approaches that require extensive training data and lack interpretability, FL systems can incorporate expert knowledge directly through rule bases and provide transparent decision-making processes.

The integration of stress detection systems with VANETs presents unique opportunities and challenges. While VANET infrastructure can provide real-time TD information and enable V2V stress alerts, existing studies have not fully used this potential. Most current approaches focus on either physiological monitoring or vehicle dynamics alone. Thus, they miss the synergistic benefits of utilizing multiple data sources within the VANET framework.

Despite these advances, the following limitations remain in current approaches.

• Lack of comprehensive integration between physiological and environmental factors.
• Limited handling of uncertainty in multi-parameter fusion.
• Limited use of VANET capabilities for stress detection.
• Absence of standardized evaluation frameworks for real-world deployment.
• Trade-off between accuracy and computational complexity for real-time applications.

This paper addresses these gaps by proposing an IT2FLS-based system that integrates multiple stress indicators while using VANET infrastructure to improve detection accuracy and practical applicability.

Table 1. Summary of related works in driver stress detection.

Ref.	Method/Algorithm	Advantages	Limitations
[8]	ECG, EMG, GSR, respiration (multi-physiological)	High accuracy (97%) in real driving; foundational dataset	Intrusive sensors; higher computational complexity
[9]	Ultra-short-term HRV analysis	≤3 min windows with high accuracy; real-time capability	Requires consistent signal quality; sensitive to artifacts
[10]	Steering-wheel signal from unmodified wheel	Non-intrusive; ∼77% accuracy without extra hardware	Limited to steering-based cues; subject-specific tuning
[12]	T1FLS with GSR & HR	Accuracy up to 99.5% in controlled settings; interpretable rules	Primarily binary stress classification; limited real-world validation
[13]	Multimodal CNN with GSR & HR	High accuracy (95.67%); short-term signals; automatic feature extraction	Requires multiple sensors; computationally intensive; limited datasets
[14]	Ultra-short-term HRV analysis	94% accuracy (3 min); 88% accuracy (1 min); validated surrogate features	Limited to mental stress detection; requires quality ECG signal
[15]	ECG electrodes in steering wheel	Non-intrusive; 45.62% driving time coverage	Limited signal quality; contact-dependent; requires validation
[16]	Multimodal fusion in Bayesian networks	Real-time stress-event detection; accuracy improved from physiology-only to with events	Complex sensor/event setup; practicality concerns
[17]	ANOVA classifier with single GSR	95.83% accuracy; reduced computational burden; single sensor	Limited to GSR amplitude feature; requires careful sensor placement
[18]	Wearable PPG, ECG, EDA (Electrodermal Activity) sensors	Pilot study and Stress-Predict dataset	Optimal measurement approach unclear; challenges in real-world deployment
[19]	Smart wrist-worn wearables	Real-life continuous monitoring; artifact removal; discriminated stress levels	Limited to wrist-worn devices; subject to motion artifacts
[20]	CNN–LSTM fusion with ECG, vehicle data, context	92.8% accuracy; outperforms traditional ML; automatic features	Requires large training data; computationally expensive; black-box nature
[21]	CNN and CNN–LSTM with fuzzy EDAS evaluation	Automatic feature extraction; multiple evaluation metrics	Complex model; requires extensive training; interpretability challenges
[23]	Type-2 fuzzy LSTM for traffic forecasting	Handles time-series uncertainty; improved robustness	Not designed for stress detection; indirect relevance
Proposed FDSDS	IT2FLS with HRV, GSR, SAV, TD	Uncertainty-aware; non-intrusive; VANET-integrated; interpretable	Simulation-based evaluation (requires field validation)

provides a summary of existing research in driver stress detection, highlighting the methodologies, advantages, and limitations of various approaches.

3. FL Overview

FL provides a mathematical framework for addressing uncertainty and imprecision, making it well-suited for modeling complex human-centric systems such as driver stress detection. Unlike classical binary logic that limits variables to true or false values, FL enables variables to take any value within the continuous range of [0, 1] [24]. This characteristic is essential for driver stress assessment, where stress levels show gradual transitions rather than changes between discrete states.

In the context of driver stress detection, traditional binary approaches fail to capture the uncertainty nature of stress indicators. For instance, HRV value cannot be simply classified as “stressed” or “not stressed” without losing important information about intermediate stress states. FL naturally handles these intermediate conditions through degrees of membership, allowing representations such as “moderately stressed” or “slightly relaxed”.

IT2FLS extends traditional T1FLS by introducing an additional dimension of uncertainty in the MFs themselves. This improvement is particularly valuable for driver stress detection, where the relationship between physiological signals and stress levels varies across individuals and driving conditions. Figure 1 illustrates the structure of IT2FLS, which consists of five main components.

• Fuzzifier: It transforms crisp sensor readings (heart rate, skin conductance, steering angle) into interval type-2 fuzzy sets using upper and lower MFs. This captures both the measurement value and its associated uncertainty.
• Rule Base: It contains the collection of IF–THEN rules encoding expert knowledge about driver stress patterns. It includes relationships between multiple physiological and behavioral indicators to determine stress levels.
• Inference Engine: It processes the fuzzified inputs through the rule base using interval type-2 fuzzy operators. This generates type-2 fuzzy outputs that preserve uncertainty information throughout the inference process.
• Type-Reducer: It converts the type-2 fuzzy output sets into type-1 fuzzy sets using algorithms such as the Karnik–Mendel (KM) iterative procedure [22]. This manages computational complexity while preserving essential uncertainty information.
• Defuzzifier: It transforms the reduced type-1 fuzzy sets into crisp stress-level values suitable for real-time decision making in VANET applications. This typically uses centroid or weighted-average methods.

The IT2FLS framework offers significant advantages for driver stress detection in VANET environments. The additional degree of freedom in modeling uncertainty enables the system to handle variations in sensor accuracy, individual physiological differences, and environmental noise more effectively than traditional approaches. This robustness is important for practical deployment where perfect sensor data and uniform driver responses cannot be guaranteed.

4. Proposed Fuzzy-Based System

Figure 2 illustrates the overall driver stress detection framework for VANETs. The framework comprises three modules: the Data Collection Module, the Driver Stress Evaluation Module, and the Driver Stress Detection Module. The Data Collection Module collects data using different kinds of sensors and cameras. The proposed Fuzzy-based Driver Stress Detection System (FDSDS) is implemented in the Driver Stress Evaluation Module. Based on the evaluation results, the Driver Stress Detection Module determines the driver stress level. The structure of the proposed FDSDS is shown in Figure 3.

4.1. Selection and Definition of Parameters for FDSDS

The validity and robustness of a driver stress detection model depend on the choice of input parameters. In this study, input parameters are selected under the operational constraints of everyday driving and VANET traffic situation as follows.

• Non-invasiveness and practical deployability.
• Real-time measurability.
• VANET data acquisition.
• Complementarity inputs, where each input provides distinct information about driver stress.

Table 2. Comparison of candidate input parameters based on the selection criteria.

Parameters	Non-Invasiveness	Real-Time	VANET Data Acquisition	Complementary	Notes
HRV	✓	Δ	✓	✓	Wearable-friendly; stress/workload effects are well supported, but stable estimation may require a longer window than purely reactive signals.
GSR	✓	✓	✓	✓	Fast sympathetic response; robust wearable acquisition via skin-contact electrodes.
SAV	✓	✓	✓	✓	Directly available from CAN (Controller Area Network) bus; reflects driver control variability under distraction/stress.
TD	✓	✓	✓	✓	Accessible via V2V/V2I; represents external driving demand linked to stress and workload indices.
EEG	✕	Δ	✕	✓	Intrusive setup and motion artifacts make it impractical for everyday driving despite rich cognitive-state information.
Respiration	Δ	Δ	Δ	Δ	Measurement is feasible with dedicated sensors but is easily confounded by speech and ambient conditions; deployability varies by setup.
Camera-based (pupil/facial)	Δ	✓	Δ	Δ	Real-time is feasible, but reliability can be affected by illumination/occlusion; deployment may raise privacy and mounting constraints.
Brake/acceleration patterns	✓	✓	✓	Δ	CAN-accessible, but strongly confounded by traffic flow and vehicle-specific calibration (e.g., brake sensitivity).
Roadway conditions	✓	Δ	Δ	Δ	Relevant to stress/workload, but real-time availability and standardization depend on infrastructure, maps, and sensing fidelity.

✓: suitable, Δ: conditionally suitable, ✕: unsuitable.

summarizes a structured comparison among candidate parameters considered in this work, including physiological (HRV, GSR, EEG (Electroencephalography), respiration), behavioral (SAV, brake/acceleration patterns), and environment-related indicators (traffic density and roadway conditions).

Based on this comparison, we selected four input parameters (HRV, GSR, SAV, and TD) to cover (1) physiological stress responses with complementary autonomic dynamics (HRV, GSR) [25,26,27,28,29], (2) behavioral stress in vehicle control obtainable without additional sensors (SAV) [30,31,32], and (3) environmental stress caused by external driving demand accessible through VANET communications (TD) [33,34]. Other parameters such as EEG [35] and respiration [36,37] were excluded mainly due to deployability and artifact sensitivity. Camera-based approaches [35,38] were treated as conditionally suitable due to illumination/occlusion and practical constraints. Vehicle-based indicators such as brake/acceleration patterns can correlate with stress [11] but are more sensitive to traffic context and calibration differences than steering-based variability under typical driving conditions [32]. Roadway geometry and surface conditions (e.g., mixed roads and rough pavements) have also been reported to induce distinct levels of driving stress as reflected in HRV responses [39].

Adding more input parameters in general increases the number of rules that must be designed, validated, and maintained, which can increase the computational cost and reduce interpretability in practical deployment. The selected four inputs therefore provide a minimal but complementary set that supports real-time VANET deployment while preserving multi-modal robustness.

The input and output parameters of FDSDS are explained in the following.

• Heart Rate Variability (HRV): HRV measures the variation in time intervals between consecutive heartbeats. In this work, HRV is represented by time-domain features such as Root Mean Square of Successive Differences (RMSSD) and proportion of NN50 count (pNN50), normalized to a [0, 1] scale [40].
• Galvanic Skin Response (GSR): GSR measures variations in skin conductance. This parameter is represented using both tonic (baseline) and phasic (response) components and normalized to a [0, 1] scale before fuzzification [41,42].
• Steering Angle Variation (SAV): SAV is computed from steering wheel angle signals obtained from the vehicle CAN bus. In this work, SAV characterizes the dispersion and temporal variations of steering movements using time- and frequency-domain descriptors, normalized to a [0, 1] scale [10,32].
• Traffic Density (TD): TD is quantified as the vehicle count per unit road length acquired through V2V/V2I communications, and normalized to a [0, 1] scale [33].
• Driver Stress Level (DSL): DSL ranges from DSL1 (minimal stress) to DSL9 (maximum stress), enabling detailed stress monitoring for safety interventions.

The integration of these four parameters through sensor fusion techniques enhances detection reliability compared to unimodal approaches [43].

4.2. MFs and Rule Base for FDSDS

Table 3. Linguistic variables and term sets for FDSDS parameters.

Parameter	Linguistic Term Sets
Heart Rate Variability (HRV)	Low → Lo, Medium → Me, High → Hi
Galvanic Skin Response (GSR)	Low → Lw, Medium → Md, High → Hg
Steering Angle Variation (SAV)	Small → S, Medium → M, Large → L
Traffic Density (TD)	Light → Li, Moderate → Mo, Heavy → He
Driver Stress Level (DSL)	DSL1, DSL2, DSL3, DSL4, DSL5, DSL6, DSL7, DSL8, DSL9

defines the linguistic term sets for each parameter. The IT2FLS implementation uses two-sided Gaussian MFs with uncertain spreads for the physiological inputs (HRV and GSR) and triangular/trapezoidal functions for the behavioral and environmental inputs (SAV and TD), as illustrated in Figure 4. For each term, an interval type-2 Footprint Of Uncertainty (FOU) is formed by Lower MF (LMF) and Upper MF (UMF). The FOU is visualized as the shaded region between the upper and lower MFs, as illustrated in Figure 5. The output stress levels (DSL) are modeled with triangular MFs. The type reduction uses the Enhanced KM (EKM) algorithm, balancing computational efficiency and accuracy. The defuzzification process applies the centroid method to generate crisp stress values suitable for real-time VANET applications.

We define in following interval type-2 triangular (f), trapezoidal (g), and two-sided Gaussian (h) MFs. The triangular MF $f(x;x_0,{\overline{w}}_0,{\underline{w}}_0,{\overline{w}}_1,{\underline{w}}_1)$ has the UMF support $[x_0-x_{{\overline{w}}_0},x_0+x_{{\overline{w}}_1}]$ and the LMF support $[x_0-x_{{\underline{w}}_0},x_0+x_{{\underline{w}}_1}]$. The trapezoidal function $g(x;x_0,x_1,{\overline{w}}_0,{\underline{w}}_0,{\overline{w}}_1,{\underline{w}}_1)$ contains plateau $[x_0,x_1]$, the UMF support $[x_0-x_{{\overline{w}}_0},x_1+x_{{\overline{w}}_1}]$, and the LMF support $[x_0-x_{{\underline{w}}_0},x_1+x_{{\underline{w}}_1}]$. For the two-sided Gaussian function $h(x;{\underline{\sigma }}_0,{\underline{c}}_0,{\underline{\sigma }}_1,{\underline{c}}_1,{\overline{\sigma }}_0,{\overline{c}}_0,{\overline{\sigma }}_1,{\overline{c}}_1)$, the UMF has a unit plateau $[{\overline{c}}_0,{\overline{c}}_1]$ with left and right half-Gaussian tails of standard deviations ${\overline{\sigma }}_0$ and ${\overline{\sigma }}_1$, and the LMF has a unit plateau $[{\underline{c}}_0,{\underline{c}}_1]$ with tails of standard deviations ${\underline{\sigma }}_0$ and ${\underline{\sigma }}_1$.

For FDSDS, the input parameters are defined as follows.

$$\begin{array}{ccc}\hfill T\left(\mathrm{HRV}\right)& =& \{\mathrm{Low}\left(\mathrm{Lo}\right),\mathrm{Medium}\left(\mathrm{Me}\right),\mathrm{High}\left(\mathrm{Hi}\right)\}\hfill \\ \hfill T\left(\mathrm{GSR}\right)& =& \{\mathrm{Low}\left(\mathrm{Lw}\right),\mathrm{Medium}\left(\mathrm{Md}\right),\mathrm{High}\left(\mathrm{Hg}\right)\}\hfill \\ \hfill T\left(\mathrm{SAV}\right)& =& \{\mathrm{Small}\left(\mathrm{S}\right),\mathrm{Medium}\left(\mathrm{M}\right),\mathrm{Large}\left(\mathrm{L}\right)\}\hfill \\ \hfill T\left(\mathrm{TD}\right)& =& \{\mathrm{Light}\left(\mathrm{Li}\right),\mathrm{Moderate}\left(\mathrm{Mo}\right),\mathrm{Heavy}\left(\mathrm{He}\right)\}\hfill \end{array}$$

The MFs for input parameters are defined using interval type-2 representations.

$$\begin{array}{ccc}\hfill {\mu }_{\mathrm{Lo}}\left(\mathrm{HRV}\right)& =& h(\mathrm{HRV};{\mathrm{Lo}}_{{\underline{\sigma }}_0},{\mathrm{Lo}}_{{\underline{c}}_0},{\mathrm{Lo}}_{{\underline{\sigma }}_1},{\mathrm{Lo}}_{{\underline{c}}_1},{\mathrm{Lo}}_{{\overline{\sigma }}_0},{\mathrm{Lo}}_{{\overline{c}}_0},{\mathrm{Lo}}_{{\overline{\sigma }}_1},{\mathrm{Lo}}_{{\overline{c}}_1})\hfill \\ \hfill {\mu }_{\mathrm{Me}}\left(\mathrm{HRV}\right)& =& h(\mathrm{HRV};{\mathrm{Me}}_{{\underline{\sigma }}_0},{\mathrm{Me}}_{{\underline{c}}_0},{\mathrm{Me}}_{{\underline{\sigma }}_1},{\mathrm{Me}}_{{\underline{c}}_1},{\mathrm{Me}}_{{\overline{\sigma }}_0},{\mathrm{Me}}_{{\overline{c}}_0},{\mathrm{Me}}_{{\overline{\sigma }}_1},{\mathrm{Me}}_{{\overline{c}}_1})\hfill \\ \hfill {\mu }_{\mathrm{Hi}}\left(\mathrm{HRV}\right)& =& h(\mathrm{HRV};{\mathrm{Hi}}_{{\underline{\sigma }}_0},{\mathrm{Hi}}_{{\underline{c}}_0},{\mathrm{Hi}}_{{\underline{\sigma }}_1},{\mathrm{Hi}}_{{\underline{c}}_1},{\mathrm{Hi}}_{{\overline{\sigma }}_0},{\mathrm{Hi}}_{{\overline{c}}_0},{\mathrm{Hi}}_{{\overline{\sigma }}_1},{\mathrm{Hi}}_{{\overline{c}}_1})\hfill \\ \hfill {\mu }_{\mathrm{Lw}}\left(\mathrm{GSR}\right)& =& h(\mathrm{GSR};{\mathrm{Lw}}_{{\underline{\sigma }}_0},{\mathrm{Lw}}_{{\underline{c}}_0},{\mathrm{Lw}}_{{\underline{\sigma }}_1},{\mathrm{Lw}}_{{\underline{c}}_1},{\mathrm{Lw}}_{{\overline{\sigma }}_0},{\mathrm{Lw}}_{{\overline{c}}_0},{\mathrm{Lw}}_{{\overline{\sigma }}_1},{\mathrm{Lw}}_{{\overline{c}}_1})\hfill \\ \hfill {\mu }_{\mathrm{Md}}\left(\mathrm{GSR}\right)& =& h(\mathrm{GSR};{\mathrm{Md}}_{{\underline{\sigma }}_0},{\mathrm{Md}}_{{\underline{c}}_0},{\mathrm{Md}}_{{\underline{\sigma }}_1},{\mathrm{Md}}_{{\underline{c}}_1},{\mathrm{Md}}_{{\overline{\sigma }}_0},{\mathrm{Md}}_{{\overline{c}}_0},{\mathrm{Md}}_{{\overline{\sigma }}_1},{\mathrm{Md}}_{{\overline{c}}_1})\hfill \\ \hfill {\mu }_{\mathrm{Hg}}\left(\mathrm{GSR}\right)& =& h(\mathrm{GSR};{\mathrm{Hg}}_{{\underline{\sigma }}_0},{\mathrm{Hg}}_{{\underline{c}}_0},{\mathrm{Hg}}_{{\underline{\sigma }}_1},{\mathrm{Hg}}_{{\underline{c}}_1},{\mathrm{Hg}}_{{\overline{\sigma }}_0},{\mathrm{Hg}}_{{\overline{c}}_0},{\mathrm{Hg}}_{{\overline{\sigma }}_1},{\mathrm{Hg}}_{{\overline{c}}_1})\hfill \\ \hfill {\mu }_{\mathrm{S}}\left(\mathrm{SAV}\right)& =& g(\mathrm{SAV};{\mathrm{S}}_0,{\mathrm{S}}_1,{\mathrm{S}}_{{\overline{w}}_0},{\mathrm{S}}_{{\underline{w}}_0},{\mathrm{S}}_{{\overline{w}}_1},{\mathrm{S}}_{{\underline{w}}_1})\hfill \\ \hfill {\mu }_{\mathrm{M}}\left(\mathrm{SAV}\right)& =& f(\mathrm{SAV};{\mathrm{M}}_0,{\mathrm{M}}_{{\overline{w}}_0},{\mathrm{M}}_{{\underline{w}}_0},{\mathrm{M}}_{{\overline{w}}_1},{\mathrm{M}}_{{\underline{w}}_1})\hfill \\ \hfill {\mu }_{\mathrm{L}}\left(\mathrm{SAV}\right)& =& g(\mathrm{SAV};{\mathrm{L}}_0,{\mathrm{L}}_1,{\mathrm{L}}_{{\overline{w}}_0},{\mathrm{L}}_{{\underline{w}}_0},{\mathrm{L}}_{{\overline{w}}_1},{\mathrm{L}}_{{\underline{w}}_1})\hfill \\ \hfill {\mu }_{\mathrm{Li}}\left(\mathrm{TD}\right)& =& g(\mathrm{TD};{\mathrm{Li}}_0,{\mathrm{Li}}_1,{\mathrm{Li}}_{{\overline{w}}_0},{\mathrm{Li}}_{{\underline{w}}_0},{\mathrm{Li}}_{{\overline{w}}_1},{\mathrm{Li}}_{{\underline{w}}_1})\hfill \\ \hfill {\mu }_{\mathrm{Mo}}\left(\mathrm{TD}\right)& =& f(\mathrm{TD};{\mathrm{Mo}}_0,{\mathrm{Mo}}_{{\overline{w}}_0},{\mathrm{Mo}}_{{\underline{w}}_0},{\mathrm{Mo}}_{{\overline{w}}_1},{\mathrm{Mo}}_{{\underline{w}}_1})\hfill \\ \hfill {\mu }_{\mathrm{He}}\left(\mathrm{TD}\right)& =& g(\mathrm{TD};{\mathrm{He}}_0,{\mathrm{He}}_1,{\mathrm{He}}_{{\overline{w}}_0},{\mathrm{He}}_{{\underline{w}}_0},{\mathrm{He}}_{{\overline{w}}_1},{\mathrm{He}}_{{\underline{w}}_1})\hfill \end{array}$$

where overlined parameters ($\overline{w}$) specify the UMF boundaries and underlined parameters ($\underline{w}$) determine the LMF boundaries, collectively establishing the FOU.

The output linguistic parameter is DSL. The term set for DSL is defined as follows.

$$T\left(\mathrm{DSL}\right)=\left[\begin{array}{c}\mathrm{Driver}\mathrm{Stress}\mathrm{Level}1\\ \mathrm{Driver}\mathrm{Stress}\mathrm{Level}2\\ ⋮\\ \mathrm{Driver}\mathrm{Stress}\mathrm{Level}9\end{array}\right]=\left[\begin{array}{c}\mathrm{DSL}1\\ \mathrm{DSL}2\\ ⋮\\ \mathrm{DSL}9\end{array}\right]$$

(1)

The MFs of DSL for FDSDS are defined as follows.

$$\begin{array}{ccc}\hfill {\mu }_{\mathrm{DSL}1}\left(\mathrm{DSL}\right)& =& f(\mathrm{DSL};\mathrm{DSL}{1}_0,\mathrm{DSL}{1}_{{\overline{w}}_0},\mathrm{DSL}{1}_{{\underline{w}}_0},\mathrm{DSL}{1}_{{\overline{w}}_1},\mathrm{DSL}{1}_{{\underline{w}}_1})\hfill \\ \hfill {\mu }_{\mathrm{DSL}2}\left(\mathrm{DSL}\right)& =& f(\mathrm{DSL};\mathrm{DSL}{2}_0,\mathrm{DSL}{2}_{{\overline{w}}_0},\mathrm{DSL}{2}_{{\underline{w}}_0},\mathrm{DSL}{2}_{{\overline{w}}_1},\mathrm{DSL}{2}_{{\underline{w}}_1})\hfill \\ \hfill {\mu }_{\mathrm{DSL}3}\left(\mathrm{DSL}\right)& =& f(\mathrm{DSL};\mathrm{DSL}{3}_0,\mathrm{DSL}{3}_{{\overline{w}}_0},\mathrm{DSL}{3}_{{\underline{w}}_0},\mathrm{DSL}{3}_{{\overline{w}}_1},\mathrm{DSL}{3}_{{\underline{w}}_1})\hfill \\ \hfill {\mu }_{\mathrm{DSL}4}\left(\mathrm{DSL}\right)& =& f(\mathrm{DSL};\mathrm{DSL}{4}_0,\mathrm{DSL}{4}_{{\overline{w}}_0},\mathrm{DSL}{4}_{{\underline{w}}_0},\mathrm{DSL}{4}_{{\overline{w}}_1},\mathrm{DSL}{4}_{{\underline{w}}_1})\hfill \\ \hfill {\mu }_{\mathrm{DSL}5}\left(\mathrm{DSL}\right)& =& f(\mathrm{DSL};\mathrm{DSL}{5}_0,\mathrm{DSL}{5}_{{\overline{w}}_0},\mathrm{DSL}{5}_{{\underline{w}}_0},\mathrm{DSL}{5}_{{\overline{w}}_1},\mathrm{DSL}{5}_{{\underline{w}}_1})\hfill \\ \hfill {\mu }_{\mathrm{DSL}6}\left(\mathrm{DSL}\right)& =& f(\mathrm{DSL};\mathrm{DSL}{6}_0,\mathrm{DSL}{6}_{{\overline{w}}_0},\mathrm{DSL}{6}_{{\underline{w}}_0},\mathrm{DSL}{6}_{{\overline{w}}_1},\mathrm{DSL}{6}_{{\underline{w}}_1})\hfill \\ \hfill {\mu }_{\mathrm{DSL}7}\left(\mathrm{DSL}\right)& =& f(\mathrm{DSL};\mathrm{DSL}{7}_0,\mathrm{DSL}{7}_{{\overline{w}}_0},\mathrm{DSL}{7}_{{\underline{w}}_0},\mathrm{DSL}{7}_{{\overline{w}}_1},\mathrm{DSL}{7}_{{\underline{w}}_1})\hfill \\ \hfill {\mu }_{\mathrm{DSL}8}\left(\mathrm{DSL}\right)& =& f(\mathrm{DSL};\mathrm{DSL}{8}_0,\mathrm{DSL}{8}_{{\overline{w}}_0},\mathrm{DSL}{8}_{{\underline{w}}_0},\mathrm{DSL}{8}_{{\overline{w}}_1},\mathrm{DSL}{8}_{{\underline{w}}_1})\hfill \\ \hfill {\mu }_{\mathrm{DSL}9}\left(\mathrm{DSL}\right)& =& f(\mathrm{DSL};\mathrm{DSL}{9}_0,\mathrm{DSL}{9}_{{\overline{w}}_0},\mathrm{DSL}{9}_{{\underline{w}}_0},\mathrm{DSL}{9}_{{\overline{w}}_1},\mathrm{DSL}{9}_{{\underline{w}}_1})\hfill \end{array}$$

The Fuzzy Rule Base (FRB) consists of 81 rules covering all combinations of input parameter values. These rules encode expert knowledge about stress patterns considering multiple indicators.

Table 4. FRB for FDSDS.

Rules 1–27						Rules 28–54						Rules 55–81
No.	HRV	GSR	SAV	TD	DSL	No.	HRV	GSR	SAV	TD	DSL	No.	HRV	GSR	SAV	TD	DSL
1	Lo	Lw	S	Li	DSL3	28	Me	Lw	S	Li	DSL2	55	Hi	Lw	S	Li	DSL1
2	Lo	Lw	S	Mo	DSL4	29	Me	Lw	S	Mo	DSL3	56	Hi	Lw	S	Mo	DSL2
3	Lo	Lw	S	He	DSL5	30	Me	Lw	S	He	DSL4	57	Hi	Lw	S	He	DSL3
4	Lo	Lw	M	Li	DSL4	31	Me	Lw	M	Li	DSL3	58	Hi	Lw	M	Li	DSL2
5	Lo	Lw	M	Mo	DSL5	32	Me	Lw	M	Mo	DSL4	59	Hi	Lw	M	Mo	DSL3
6	Lo	Lw	M	He	DSL6	33	Me	Lw	M	He	DSL5	60	Hi	Lw	M	He	DSL4
7	Lo	Lw	L	Li	DSL5	34	Me	Lw	L	Li	DSL4	61	Hi	Lw	L	Li	DSL3
8	Lo	Lw	L	Mo	DSL6	35	Me	Lw	L	Mo	DSL5	62	Hi	Lw	L	Mo	DSL4
9	Lo	Lw	L	He	DSL7	36	Me	Lw	L	He	DSL6	63	Hi	Lw	L	He	DSL5
10	Lo	Md	S	Li	DSL4	37	Me	Md	S	Li	DSL3	64	Hi	Md	S	Li	DSL2
11	Lo	Md	S	Mo	DSL5	38	Me	Md	S	Mo	DSL4	65	Hi	Md	S	Mo	DSL3
12	Lo	Md	S	He	DSL6	39	Me	Md	S	He	DSL5	66	Hi	Md	S	He	DSL4
13	Lo	Md	M	Li	DSL5	40	Me	Md	M	Li	DSL4	67	Hi	Md	M	Li	DSL3
14	Lo	Md	M	Mo	DSL6	41	Me	Md	M	Mo	DSL5	68	Hi	Md	M	Mo	DSL4
15	Lo	Md	M	He	DSL7	42	Me	Md	M	He	DSL6	69	Hi	Md	M	He	DSL5
16	Lo	Md	L	Li	DSL6	43	Me	Md	L	Li	DSL5	70	Hi	Md	L	Li	DSL4
17	Lo	Md	L	Mo	DSL7	44	Me	Md	L	Mo	DSL6	71	Hi	Md	L	Mo	DSL5
18	Lo	Md	L	He	DSL8	45	Me	Md	L	He	DSL7	72	Hi	Md	L	He	DSL6
19	Lo	Hg	S	Li	DSL6	46	Me	Hg	S	Li	DSL5	73	Hi	Hg	S	Li	DSL4
20	Lo	Hg	S	Mo	DSL7	47	Me	Hg	S	Mo	DSL6	74	Hi	Hg	S	Mo	DSL5
21	Lo	Hg	S	He	DSL8	48	Me	Hg	S	He	DSL7	75	Hi	Hg	S	He	DSL6
22	Lo	Hg	M	Li	DSL7	49	Me	Hg	M	Li	DSL6	76	Hi	Hg	M	Li	DSL5
23	Lo	Hg	M	Mo	DSL8	50	Me	Hg	M	Mo	DSL7	77	Hi	Hg	M	Mo	DSL6
24	Lo	Hg	M	He	DSL9	51	Me	Hg	M	He	DSL8	78	Hi	Hg	M	He	DSL7
25	Lo	Hg	L	Li	DSL8	52	Me	Hg	L	Li	DSL7	79	Hi	Hg	L	Li	DSL6
26	Lo	Hg	L	Mo	DSL9	53	Me	Hg	L	Mo	DSL8	80	Hi	Hg	L	Mo	DSL7
27	Lo	Hg	L	He	DSL9	54	Me	Hg	L	He	DSL9	81	Hi	Hg	L	He	DSL8

presents the mapping between input conditions and stress levels.

The rule design follows physiological and behavioral principles as follows.

• Low HRV combined with high GSR indicates higher stress regardless of other factors.
• Large SAV increases stress assessment when combined with physiological indicators.
• Heavy TD serves as a stress multiplier, increasing the overall stress level.
• The absence of stress indicators (high HRV, low GSR, small SAV, light TD) results in minimal stress classification.

The FRB enables the FDSDS to provide detailed stress assessments that align with physiological stress responses while maintaining robustness against sensor uncertainties and individual variations. The integration with VANET infrastructure facilitates dynamic parameter updates and collective stress monitoring.

5. Simulation Results

This section presents simulation results demonstrating the performance and behavior of the proposed FDSDS under different driving conditions. For simulations, we implemented in Rust Type-1 fuzzy inference pipeline of FuzzyC [44]. We have utilized FuzzyC for many applications such as policing mechanism, call admission control, handover decision, sensor networks, sensor–actor networks, and peer-to-peer systems. The inference engine of FuzzyC is implemented as domain-independent fuzzy functions that are extensible for different applications. We also extended the fuzzy inference pipeline with IT2FLS-specific processing, including UMF/LMF representation of MFs and EKM-based type-reduction while preserving the computational flow of the original T1FLS implementation. We evaluate the system through systematic parameter variation and analyze both quantitative and qualitative aspects of driver stress detection across different scenarios.

5.1. Performance Under Various Stress Conditions

Figure 6 illustrates the FDSDS output across three typical physiological states, demonstrating the system’s response to varying SAV and TD conditions. Each subfloat displays the crisp DSL output (solid lines) along with the IT2FLS uncertainty bounds (shaded regions), with separate curves for three SAV levels (0.1, 0.5, 0.9).

Figure 6a illustrates the high-stress physiological state (HRV = 0.1, GSR = 0.9). The system generates high stress assessments across all SAV and TD combinations, with DSL values ranging from 0.62 to 0.96. This demonstrates that the system reliably detects severe stress conditions regardless of driving behavior.

Figure 6b illustrates the moderate physiological state (HRV = 0.5, GSR = 0.5). The system exhibits intermediate stress levels with greater sensitivity to behavioral and environmental factors, yielding DSL values from 0.27 to 0.76. In this intermediate state, both SAV and TD have a stronger influence on the stress assessment.

Figure 6c illustrates the low-stress physiological state (HRV = 0.9, GSR = 0.1). The system generates minimal stress assessments with DSL values between 0.04 and 0.51, even under challenging driving conditions.

5.2. Quantitative Analysis

Table 5. Quantitative analysis of DSL under typical conditions.

Condition	Min DSL	Max DSL	Mean DSL	Std Dev	Mean FOU
Case 1 (HRV = 0.1, GSR = 0.9)	0.622	0.958	0.813	0.101	0.068
Case 2 (HRV = 0.5, GSR = 0.5)	0.265	0.756	0.519	0.132	0.139
Case 3 (HRV = 0.9, GSR = 0.1)	0.044	0.512	0.276	0.123	0.054

presents quantitative analysis of DSL distributions across the three typical conditions. The results demonstrate the system’s ability to distinguish stress levels with high granularity while maintaining consistent uncertainty bounds.

The analysis reveals several key findings. First, case 1 (HRV = 0.1, GSR = 0.9) maintains high DSL values with low variability (std dev = 0.101), indicating reliable detection of severe stress states. Then, case 2 (HRV = 0.5, GSR = 0.5) exhibits the highest sensitivity, as evidenced by the largest standard deviation (0.132) and FOU width (0.139). This increased uncertainty reflects the inherent ambiguity in intermediate stress states. Finally, case 3 (HRV = 0.9, GSR = 0.1) exhibits constrained DSL ranges, thereby avoiding false alarms during normal driving.

Statistical analysis across all 14,641 test cases reveals a monotonic association (Spearman’s $\rho$) between DSL and each input; decreasing HRV and increasing GSR, SAV, and TD are associated with higher DSL.

To quantify individual parameter effects, we performed a finite-difference sensitivity analysis on the evaluation grid. The mean absolute sensitivities were: GSR 0.357, TD 0.239, SAV 0.239 and HRV 0.235. The ordering (GSR > TD ≈ SAV > HRV) indicates that physiological indicators have the strongest influence on average. However, these sensitivities vary across the parameter space. In low-stress regions, GSR sensitivity dominates, whereas in high-stress regions, all parameters contribute more evenly to the final assessment.

5.3. Qualitative Analysis

The simulation results show several important qualitative behaviors of the FDSDS. The IT2FLS uncertainty bands (shaded regions in Figure 6) show dynamic widths that adapt to input conditions. In case 2, uncertainty increases in moderate physiological states where stress classification is naturally more unclear, while becoming narrower in case 1 and case 3, where stress levels are clear. This dynamic adaptation allows the system to maintain robust decision-making even in the presence of sensor noise and inter-driver variability.

The system shows suitable nonlinear response characteristics. In case 3 (Figure 6c), even large SAV and heavy TD produce only moderate stress assessments, showing the protective effect of healthy physiological indicators. In contrast, in case 1 (Figure 6a), the system keeps high DSL even with small SAV and light TD, focusing on physiological distress signals.

The TD shows a multiplying effect on stress assessment. Across all conditions, DSL increases more quickly with TD when SAV is large, showing synergistic effects between environmental stressors and behavioral indicators. This interaction is especially clear in case 2 (Figure 6b), where TD transitions from 0.5 to 0.7 produce high DSL increase when combined with large SAV.

The SAV parameter shows threshold-type behavior in its influence on DSL. Small SAV values (0.1) produce flat DSL curves across varying TD, while large SAV values (0.9) create steeper gradients. This indicates that steering irregularities increase the impact of environmental stressors. This behavior agrees with psychological research showing that motor control difficulties worsen stress responses to external demands.

5.4. Computational Performance Evaluation

The proposed system can be implemented as a software module deployed on vehicle gateway devices, In-Vehicle Infotainment (IVI) systems, or smartphones interfacing with the vehicle’s CAN bus and V2X communication systems. Physiological inputs (HRV, GSR) can be obtained from wearable devices via Bluetooth, while behavioral (SAV) and environmental (TD) inputs can be obtained from the CAN bus and V2I communications, respectively.

Table 6. Theoretical computational complexity of FDSDS.

Phase	Complexity
Fuzzification	$O(n\times m)$
Rule Firing	$O(R\times n)$
Type Reduction (EKM)	$O(R\times k)$
Defuzzification	$O\left(1\right)$

$n=4$ inputs, $m=3$ terms/input, $R=81$ rules, k = Number of EKM iterations.

summarizes the theoretical computational complexity of the FDSDS per inference cycle, based on the standard IT2FLS inference process [45,46].

To evaluate real-time feasibility, performance measurements were conducted in a resource-constrained virtual machine environment (1 vCPU, 1 GB RAM, Ubuntu Server 22.04). This configuration represents a conservative lower bound of computational resources available in modern automotive application processors used in IVI systems and vehicle gateways, which typically feature multi-core processors with 512 MB to 2 GB of RAM. The experimental results over 14,641 inference cycles are presented in

Table 7. Performance measurements of FDSDS.

Metric	Value
Total Inference Time	5446 ms
Average Time per Inference	$371.97$$\mathsf{\mu }$s
Fuzzification	$0.22$$\mathsf{\mu }$s (0.1%)
Rule Firing	$13.75$$\mathsf{\mu }$s (3.7%)
Type Reduction (EKM) *	$334.31$$\mathsf{\mu }$s (89.9%)
Average EKM Iterations	$1.29$

* Includes defuzzification.

.

The results indicate that a single inference cycle completes in approximately $372\mathsf{\mu }$s, which is well within update intervals on the order of seconds (e.g., 1–5 s) reported in prior driver stress monitoring architectures [47,48]. The EKM algorithm converges rapidly with an average of $1.29$ iterations, consistent with convergence properties [49]. Note that the above results quantify only the on-device computation latency of the IT2FLS inference engine. End-to-End (E2E) latency in VANET deployments additionally includes sensing/preprocessing (e.g., HRV/GSR extraction), Bluetooth/CAN access, and V2X/V2I communication delays, which are platform- and network-dependent.

These findings align with previous work demonstrating the feasibility of interval type-2 fuzzy controllers on low-cost microcontrollers [50], indicating that FDSDS is suitable for practical deployment in automotive computing environments.

6. Conclusions and Future Work

This paper presented FDSDS, which uses IT2FLS for driver stress evaluation in VANET environments. The proposed system combines four parameters (HRV, GSR, SAV, and TD) to provide reliable and detailed DSL assessment suitable for real-time safety applications. Through detailed simulation analysis, we draw the following conclusions.

• Physiological parameters have the strongest influence on stress assessment, with mean absolute sensitivities of 0.368 (GSR). Under severe physiological distress (HRV = 0.1, GSR = 0.9), the system maintains high DSL (mean 0.813) across all driving conditions.
• Behavioral and environmental parameters (SAV and TD) increase stress in moderate physiological states. In case 2 (HRV = 0.5, GSR = 0.5) with heavy traffic (TD = 1.0), increasing SAV from 0.1 to 0.9 increases DSL from 0.512 to 0.756 (+47.6%).
• The system generates minimal stress under favorable conditions as in case 3 (HRV = 0.9, GSR = 0.1, SAV = 0.1, TD = 0.0, DSL = 0.044) while approaching maximum stress under worst-case conditions as in case 1 (HRV = 0.1, GSR = 0.9, SAV = 0.9, TD = 1.0, DSL = 0.958).

Despite these results, there are some limitations. The current evaluation relies only on simulation and has not yet been validated with real-world driver stress measurements. Additionally, the approach requires reliable sensors and VANET connectivity, which may not always be available in practice. While the computational cost of type reduction is manageable in our setup, further optimization may be needed for resource-constrained platforms.

In future work, we will address these limitations and extend this study as follows.

• Real-world Validation: We will validate FDSDS in real driving scenarios by collecting synchronized physiological signals and driving in VANET context, and by defining stress labels using standardized self-reports from NASA Task Load Index and event-level annotations.
• Degraded-input Robustness: We will improve robustness for sensor noise and V2X requirement (delay, packet loss) by introducing confidence-aware inference so that uncertainty bounds remain meaningful under degraded inputs.
• Online Personalization: We will develop methods for adapting FOUs and preserving interpretability constraints (e.g., monotonicity and rule-base transparency), enabling driver-specific calibration without retraining.
• Fair Benchmarking: We will conduct comparative evaluation against representative ML/DL baselines (e.g., CNN–LSTM fusion models [20]) under identical input features and evaluation protocols, reporting not only classification accuracy but also calibration and uncertainty quality metrics.