2.1 Eligibility criteria

In this review, papers that satisfied the following criteria were considered;

1. Papers that examined driver-state detection, recognition or assessment in driving scenarios
2. Papers that used sensing methods like physiological or behavioral assessments, camera-based systems, vehicle dynamic sensors or multi-sensor fusion techniques
3. Papers that used validated behavioral instruments and offered evidence in line with driver-state detection for takeover preparedness
4. Papers that reported empirical behavioral or psychometric outcomes, provided they gave quantitative results or classification measures. Algorithmic frameworks for real-time inference were also considered

Studies were excluded from this review if they were 1) Published before 2021, unless identified as foundational contributions to driver-state monitoring frameworks (two such exceptions, [29,30] were retained: [29] was retained due to its provision of a broad system-level review of driver inattention for intelligent vehicles, classifying inattention and organizing monitoring approaches, supporting the sensing-modality structure used in this review; [30] was retained for providing early empirical evidence of driver-state detection using feature-level fusion and user-specific classification; 2) Not published in English; 3) Studies focused exclusively on fully autonomous (SAE Level 4–5) or fully manual (Level 0–1) driving, with no relevance to conditional automation or takeover transitions.

2.2 Selection process

An initial collection of 459 records was obtained by exporting and consolidating the identified records from each database in EndNote 21. All retrieved records were exported into EndNote 21 for reference management. Duplicate records were identified and removed using EndNote’s automated duplicate detection function, followed by manual verification to ensure that records representing the same study were not retained multiple times. Four duplicates were thus removed by deduplication, leaving 455 unique records. Following duplicate removal, titles and abstracts were screened against predefined eligibility criteria by the review team. Screening decisions were subsequently checked within a team of four reviewers to ensure consistency with the inclusion and exclusion criteria. Records considered potentially relevant were retrieved for full-text assessment. Full-text articles were evaluated based on their relevance to driver-state detection or assessment in conditional automated driving, the sensing or assessment modality used, and the reported behavioral, physiological, psychometric, computational, or takeover-related outcomes. During title and abstract screening, records were excluded when they did not align with the predefined scope of this review. The main reasons for exclusion included studies unrelated to driver-state detection or assessment; studies focusing on general driving behavior without relevance to takeover preparedness; studies investigating fully autonomous (SAE Level 4–5) or fully manual (SAE Level 0–1) driving without transferable relevance to conditional automation. Any uncertainties or disagreements regarding study eligibility were discussed among the reviewers and resolved through consensus. No quantitative measures of inter-rater agreement were used. The selection procedure is summarized in the PRISMA flow diagram (Fig 1). The data items of interest extracted from the papers are also listed in Table 2.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. PRISMA flow diagram for identified studies.

https://doi.org/10.1371/journal.pone.0358700.g001

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. List of variables considered for data extraction.

https://doi.org/10.1371/journal.pone.0358700.t002

2.3 Risk-of-bias assessment

Although driver-state detection is primarily an automotive human-factors research area, many of its assessment approaches rely on physiological and behavioral measurements that share characteristics with biomedical diagnostic evaluations. Signals such as EEG, ECG, EOG, heart-rate variability (HRV) and ocular behavioral measures are used to infer latent human states. QUADAS-2 was thus considered an appropriate methodological foundation because it provides a structured framework for evaluating studies involving an index test, reference standard, and outcome interpretation. However, the original framework was adapted to reflect the specific requirements of driver-state detection research. The domains assessed included driver selection, index test, reference standard and flow and timing, as shown in Table 3. Ratings were categorized as low, high or moderate. QUADAS-2 ratings were summarized using RoBVIS.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Risk of bias domains and their definitions.

https://doi.org/10.1371/journal.pone.0358700.t003

Qualities that characterise a low risk of bias for all domains are also summarised in Table 4.

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Qualities that characterize low-risk of bias.

https://doi.org/10.1371/journal.pone.0358700.t004

2.4 Synthesis method

A quantitative meta-analysis was not conducted due to substantial heterogeneity in study designs, driver-state definitions, experimental contexts. However, findings were synthesized narratively and thematically. Studies were grouped by their experimental contexts and synthesized according to sensing modalities, algorithmic approaches, reported metrics and relevance to takeover preparedness. Table and visual summaries were used to organize the information and present patterns across study characteristics.

4.1 Vision-based sensing: High adoption, variable reliability

Vision-based approaches were the most adopted sensing modality across the reviewed literature (n = 22 studies). Their reported performance however varied across study settings and datasets. Studies using RGB cameras achieved high headline accuracies. A CNN-based YOLOX and RetinaFace system for distraction detection reported object detection accuracy averaging 97.1% [33], a 3D-CNN-LSTM architecture for drowsiness detection achieved 95.3% [35], and an Inception-ResNet-based face recognition classifier reported 98.4–99.2% on standard benchmarks [36]. Attention-based networks such as AG-Net reached 99.70% on the AUC-V1 distracted-driver dataset and 96.65% on AUC-V2 [38]. However, these high figures frequently fail to transfer to naturalistic conditions. Studies using near-infrared cameras demonstrated the sensitivity to environmental factors directly: a user-specific RBF-SVM drowsiness detector operating with NIR video reported no numeric accuracy, citing performance variability across lighting and occlusion conditions [30]. A gaze-zone estimation system using IR cameras reported accuracies ranging from 73.1% to 83.1% depending on time of day and whether the driver wore glasses [32]. Thermal imaging combined with rule-based landmark detection produced no reportable accuracy under real-world conditions [70]. Studies relying on RGB cameras for naturalistic gaze tracking similarly reported non-reportable performance when applied outside controlled settings [71,72]. Face-based emotion recognition using a customized VGG-19 architecture achieved only 66% on the FER2013 and related datasets. This underscores the difficulty of generalizing from controlled image corpora to in-vehicle conditions [28]. The pattern across driver-state detection contexts – distraction [31,38], drowsiness [30,32,35,59,70], emotion [28,36], and gaze [71,73] – is consistent. The average reported accuracies exceeded 90% in several of these studies. Performance however depreciates under changes in illumination, head orientation, and facial occlusion, conditions that are routine in naturalistic driving. This pattern recurs across studies using RGB, near-infrared, and thermal cameras. This suggests that the problem is inherent to the sensing principle rather than to specific hardware choices.

4.2 Physiological sensing: Signal consistency at the cost of practicality

Physiological sensing – primarily EEG, ECG, EOG, and EDA – produced consistent signal outputs across the studies in which it was employed. A stacked autoencoder and attention-LSTM system applied to EEG, facial EMG, pulse rate, respiration, and GSR data achieved 98.63% accuracy in a simulator study [4], and a multi-branch deep CNN fusion of EEG, EDA, and ECG signals achieved 82% under leave-one-subject-out cross-validation [26]. LSTM applied to three-layer EEG signals reached 85.6% in a train-driving context [24]. Ensemble classifiers (hard-voting across Decision Tree, RF, Gradient Boosting, LightGBM, CatBoost, AdaBoost) on ECG-derived features achieved 64–74% in a VR driving simulator [21]. For wrist-worn sensors, a classical ML comparison of KNN, SVM, RF, XGBoost, and LightGBM on PPG and EDA data reached 88.7% binary and 85.6% three-class accuracy [43]. Drowsiness detection from ECG, HR, and HRV using logistic regression on a closed-loop driving track reported AUC of 78–100% in younger drivers and 86–94% in older drivers [59]. Thermal-plus-seat-pressure sensing combined with RBF-SVR for drowsiness prediction achieved 83% in real-highway conditions [46]. Studies on mental workload using physiological signals and classical ML reported metric coverage across accuracy and sensitivity [11,12,65,74]. EEG-based studies however, focused on behavioral responses in emergency driving tasks reported no classification accuracy but provided validated physiological markers through ANOVA and GLMM [22]. Review-level contributions synthesized the physiological sensing evidence base across ECG and HRV for fatigue [25] and the broader landscape of stress, road conditions, and driving environment [75]. A recurring limitation across many physiological studies was equipment intrusiveness: wearable electrode arrays and monitoring devices constrained driver movement and are unlikely to be acceptable in consumer-facing Level 3 systems without significant miniaturization advances [56].

4.3 Vehicle-performance sensing: Structured data, limited representation

The use of steering entropy, SDLP, and pedal pressure as proxies for driver attentiveness was underrepresented in the reviewed literature despite offering a non-intrusive, equipment-free alternative to physiological and vision-based approaches. A generalized Pareto distribution applied to simulator-derived takeover-time data for takeover-threshold detection achieved a classification accuracy of 90.38% for older drivers [2], while an LightGBM classifier on pre-lane-change driving features reached 92.3% [49]. In contrast, the majority of vehicle-performance studies did not report classification metrics. Repeated-measures ANOVA on steering angle and lateral position data yielded validated behavioral indicators of distraction without a classifier output [3]. Simulator-plus-eye-tracker studies of lane-changing behavior, visual workload and situational awareness in freeway exit areas similarly reported no accuracy values [57,61,62,76]. Studies on tunnel driving behavior used CAN-bus, Mobileye, camera, and GPS combined with K-means clustering and ANOVA to identify lane-deviation patterns. These however didn’t produce a classifier accuracy figure [51]. Additional vehicle-performance contributions examined takeover response timing and driver workload in conditionally automated scenarios [42,77,78], drowsiness indicators through naturalistic driving data [48,54], and intersection risk warning performance [79]. The structured, numerical nature of vehicular data proved well-suited to classical ML classifiers. The primary limitation identified across these studies was ecological validity: the behavioral cues produced are highly dependent on the specific simulator design and takeover event configuration [63]. This limits the generalizability of findings across experimental settings.

4.4 Multimodal fusion: Performance benefits and implementation trade-offs

Multimodal fusion approaches – combining two or more of the above sensing modalities – consistently produced balanced performance profiles across all reported metrics. Compared with single-modality methods, several fusion studies reported a broader range of performance metrics. This gave a more balanced evaluation profile. A multi-sensor fusion system combining galvanic skin response, driving simulator outputs, and webcam data for emotion classification achieved 92.0–92.9% accuracy [65]. A real-world expressway study combining ECG, EDA, video (front and rear), and RF/SVM/NN classification with CNN-based vehicle detection reached 97.8% accuracy [12]. A deep CNN fusion of EEG, EDA, and ECG in a driving simulator context achieved 82% under cross-subject validation [26]. Studies combining physiological signals with simulator-based behavioral measures of takeover performance found that emotional instability – assessed via joint physiological and vehicle-performance indicators predicted takeover quality [53]. Train-driving attention monitoring using multi-biosignal fusion with LSTM achieved 85.6% [24]. Driver mental engagement studies applying deep CNN fusion to EEG, EMG, and EDA in simulator conditions demonstrated the advantage of combining physiological modalities over single-signal approaches [26,65]. Other fusion configurations incorporating CAN-bus, camera, GPS, Mobileye, and inertial guidance system data identified driving pattern deviations in urban tunnels [51] and characterized drowsiness progression under real-highway thermal comfort conditions [46]. Conceptual frameworks specifically addressing multimodal cooperation between driver and automation in Level 3 systems provided methodological guidance for fusion architecture design. Utilizing multimodal fusion techniques covers up the bottlenecks of individual modalities. This yields a more robust representation of driver state [64]. Despite the ability to compensate for the limitations of individual sensing modalities, higher performances recorded were achieved under controlled experimental conditions. These depend on accurate sensor synchronization, stable signal quality, reliable missing-data handling and increased computational resources. Therefore, multimodal fusion should not just be judged by classification accuracy or metric coverage, but also by its latency, implementation complexity, robustness to real-world driving conditions and feasibility for continuous in-vehicle monitoring.

4.5 Data-flow architecture and real-time constraints in driver state detection

The studies reviewed provide extensive coverage of sensing modalities and how they are classified. However, less attention is given to data-flow architecture that govern the conversion of sensor inputs into driver-state outputs. Driver-state detection systems are not defined just by their sensing modality or classification algorithm; but by the sequence of processing stages through which data is transformed. These stages are signal acquisition, pre-processing, feature extraction, model inference and decision output. Across the reviewed literature, the degree of processing required between input signals and classification output were studied. From that, some dominant data flow architecture were identified.

4.6 Inconsistent data and reporting methods

The reviewed studies reveal that there are a number of problems in the reporting and interpretation of the driver-state detection and assessment in conditional automated driving. The inconsistencies in sensor results, inconsistencies in data collection circumstances, and inconsistencies in report methods prevent the comparability of the results across the literature. Several review and conceptual contributions mapped these challenges at the field level. A systematic review of physiological fatigue and drowsiness monitoring [25], a broad review of driver stress across traffic and environmental conditions [75], a cognitive-load and task-switching analysis for semi-autonomous vehicles [10], a review of existing technologies for identifying abnormal driving [80], a naturalistic driving study of awareness and speed behavior [81], and a qualitative study of autopilot as a default cognitive mode [82] collectively document the fragmentation of the evidence base. Non-sensing and psychometric studies contributed evidence on emotional regulation and road aggression [7,69], personality traits and driving distraction [67], mentalization and road aggression [60], fatigue and persistent performance impairment in new drivers [6], driver uncertainty in lane-change decisions [61], situation awareness assessment by experts [83], and automated vehicle acceptance [68]. These were investigated without algorithmic sensing, highlighting that driver state has dimensions that camera and biosignal systems alone do not capture. Ambient conditions such as the change of the illumination and the change of the head orientation affected vision-based sensors. Such factors caused the instability in the gathered facial or postural cues, which led to drastic changes in sensitivity and F1-scores. Vision-based systems, even though widely used in the studies under analysis, did not allow generalizing their performance to various tasks, devices, or scenarios on the road, due to the variety of operational environments.

The existence of single and multi-metric studies (Table 11) led to the absence of balance in the interpretation of performance, which made it difficult to determine uniform results. These discrepancies could be seen through the dispersion of measures on the 85 trials analyzed. This inconsistency also varied across sensing modalities, with substantial differences in the reporting frequencies of accuracy, sensitivity, F1-score, AUC, and reaction time (Fig 4).

Download:

• PNG
  larger image
• TIFF
  original image

Table 11. Summary of evaluation metrics reported across the included studies.

https://doi.org/10.1371/journal.pone.0358700.t011

The inconsistent reporting of evaluation metrics limits cross-study interpretation. This inconsistency matters because driver-state detection is often concerned with identifying rare but safety-critical states rather than simply maximizing overall correct classification. Cross-study numerical comparisons should therefore be interpreted cautiously, since reported values of accuracy, F1-score, AUC, or sensitivity are influenced by differences in driver-state labels, class distributions, sampling strategies, sensing modalities, experimental tasks, simulator fidelity, environmental conditions, and validation procedures. A high accuracy value obtained from a controlled, balanced dataset may not indicate the same practical reliability as a lower value obtained under more naturalistic or operationally complex conditions.

4.7 Inconsistent definition of driver-state constructs

A further challenge was inconsistency in both the definition and operationalization of driver-state constructs. Constructs such as distraction were investigated using different operational measures, including behavioral indicators, physiological markers, secondary-task performance, and simulator-based task manipulations. These approaches are not necessarily contradictory, as they may provide complementary measures of the same underlying construct; however, differences in thresholds, temporal windows, task designs, and ground-truth criteria can lead to substantial variation in labelling and evaluation across studies.

Studies that used 2x2 classification tables proved to be very sensitive to variations in the true positives, false positives, false negatives, and true negatives. Any changes in the balance of these classes resulted in significant changes in sensitivity, specificity and F1-scores, although the underlying performance of the classifier was the same.

4.8 Heterogeneity in methodology used

Much of the challenge regarding methodological consistency was also revealed in RoBVIS evaluation (Table 12 & Fig 6). In seventy percent of the studies, low-risk was established in the index-test domain but high-risk assessments were located mostly in the areas of participant-selection and flow-and-timing. The seventy high-risk instances between the two domains showed that many of the studies used non-representative samples, lapsed time between sensor data and that they failed to synchronise behavioral, visual, and physiological sensor data. The domain of reference standard had twenty-two high-risk ratings which can be explained by the reliance on subjective ground-truth definitions as opposed to objective physical standards. Such issues reduced the confidence in reported performance claims. They also revealed major methodological gaps across many studies.

Download:

• PNG
  larger image
• TIFF
  original image

Table 12. Summary of RoBVIS ratings across the included studies.

https://doi.org/10.1371/journal.pone.0358700.t012

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. Risk-of-bias assessment across four RoBVIS domains.

https://doi.org/10.1371/journal.pone.0358700.g006

4.9 Limitations of the evidence and review process

Many of the reviewed studies used simplified environment or constant vigilance procedures, which do not sufficiently assess the complexity of real-time changes in Level 3 driving. Though these environments helped in the precise measurement of physiological or behavioral responses, the controlled environment also restricted the use of the results to real-life scenarios. Such experimental conditions often did not capture environmental unpredictability [84,85], multitasking demands, effects of motion and the real-time system requirements characteristic of real-world Level 3 driving.

Furthermore, the recognition of driver-state was based on the other factors such as the cognitive load, the visual load and the emotional state. These determinants required measurement to be consistent across research, but they were defined differently. There was a varied influence of emotional factors on driving behavior in various research environments, such as anger, annoyance, and stress. The variability in these determinants led to the dissimilarity in the operationalisation of driver state.

Other problems arose in the time aspects of the acquisition performance. Reaction-time effects were measured with different methodologies; some of them used gaze shifts, some used data of pedal-pressure or button clicks. The differences made the comparison hard and complicated the meaning of a driver who is distracted, fatigued or emotionally disturbed to react to a takeover request.

Problems connected to sensor stability, changes in data sets, differences in metric reporting and bias patterns, divergent definitions, and environmental limitations make finding reliability and generalizability to real-world problems in driver-state detection methods challenging.

Methodological limitations exist for the review process utilised: First, the search was limited to just 4 databases. Although these databases cover a substantial proportion of engineering, transportation, computer science and human factors literature, relevant studies indexed outside these sources may have been missed. Also, because the included studies varied across several domains, a quantitative meta-analysis was not conducted. The findings should therefore be interpreted as a narrative and thematic synthesis.

5.1 Gap 1: Absence of a Level 3-specific benchmark dataset

No publicly available, standardised dataset exists that combines synchronized multimodal data collected during realistic SAE Level 3 takeover events. The reviewed studies used a wide range of proprietary, simulator-based, and repurposed datasets (e.g., FER2013, LISA V2) that were not designed for this context. This gap makes cross-study comparison unreliable and prevents the development of generalizable models. Future research should prioritise the creation and open publication of a Level 3 driver-state benchmark dataset, collected from demographically stratified participants in both simulator and instrumented real-vehicle environments, with standardised TOR event designs and synchronized multi-sensor logging.

5.2 Gap 2: Inconsistent definition and operationalization of driver-state constructs

The constructs of ‘distraction,’ ‘fatigue,’ ‘drowsiness,’ and ‘cognitive load’ were operationalised inconsistently across the reviewed literature, making nominal comparisons between studies substantively misleading. No community-agreed operational definitions currently exist for these constructs in the Level 3 takeover context. Future work should establish a taxonomy of driver-state constructs with agreed measurement operationalisations, ideally anchored to objective physiological or behavioral thresholds rather than subjective ratings.

5.3 Gap 3: Underrepresentation of diverse driver populations

The RoBVIS assessment identified convenience sampling as the dominant source of bias in Participant Selection (38.8% high-risk). The majority of reviewed studies recruited young, university-educated participants, systematically excluding older drivers, drivers with disabilities, and commercial drivers. Given that Level 3 vehicles will be operated across the full age spectrum and in diverse driving environments, future studies must recruit demographically representative samples and report participant characteristics (age, gender, driving experience, health status) in sufficient detail to enable subgroup analyses.

5.4 Gap 4: Insufficient real-world and longitudinal validation

Only 12 out of the 85 reviewed studies were conducted in direct Level 3 automation settings; the remainder used simulators or proxy tasks. Simulator validation is a necessary but insufficient step for systems intended for real-world deployment. Future research must include on-road validation studies with instrumented Level 3 vehicles, conducted under naturalistic conditions including night driving, adverse weather, and high-traffic environments. Longitudinal studies assessing driver-state detection system performance over extended periods (weeks to months) are also absent from the literature and are needed to evaluate reliability under long-term use conditions.

5.5 Gap 5: Absence of minimum reporting standards for performance evaluation

The finding that 44.7% of studies reported a numeric accuracy value, while only 10.6% reported AUC (Fig 5), reflects the absence of community-agreed reporting standards. This prevents safety validation, since accuracy-only reporting cannot detect the class-imbalance failure modes most relevant to real-world driver monitoring. The field needs a minimum reporting checklist for driver-state detection studies. At minimum, this checklist should require: classification accuracy, F1-score, sensitivity, AUC or specificity, sample size and class balance, and dataset origin (public or proprietary). A proposed checklist is provided in the supplementary material.

5.6 Policy and practice implications

UN Regulation No. 157 [86] requires vehicles equipped with Automated Lane Keeping System to incorporate a driver-availability recognition system capable of detecting whether the driver is present and available to take over the driving task. The regulation also calls for real-world assessment by engineers to supplement the manufacturer’s technical documentation. The European Union, Regulation 2019/2144 [87] further mandates driver drowsiness and attention warning systems: this emphasizes the regulatory importance of reliable driver-state monitoring. Reliable regulatory deployment depends on clearly defined driver states, validation under real-world conditions, robust sensing under degraded or variable conditions and evidence that driver-state monitoring can operate accurately within real-time transition constraints. This need for clearer and more harmonised requirements is consistent with broader fatigue-management evidence showing that inconsistent enforcement and regulatory interpretation can weaken safety practice in transport settings [88]. However, these legal instruments focus primarily on functional requirements, validation procedures, and privacy-related constraints, rather than prescribing a standardized academic reporting set based on metrics such as sensitivity, F1-score, or AUC.