Bayesian network approach to building an affective module for a driver behavioural model

Carmen Armero; Dorota M{\l}ynarczyk; Francisco Palmi-Perales; Gabriel Calvo; Ursula Martinez-Iranzo; Virgilio G\'omez-Rubio

REVIEW 4 major objections 4 minor 13 references

Bayesian network approach to building an affective module for a driver behavioural model

T0 review · 4 major / 4 minor · reviewed 2026-08-09 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read This paper claims that a Bayesian network fitted to 1,892 driving observations estimates mental load probability 0.94 and active fatigue probability 0.78 when mean heart rate exceeds 100 beats per minute and respiration exceeds 20 breaths…

desk verdict A clean but unvalidated BN fit to driver physiology; the repeated-measures structure is ignored, so the headline probabilities are point estimates without cluster-robust support. read the letter →

arxiv 2502.03254 v1 pith:TX36UJLM submitted 2025-02-05 stat.AP

Dorota M{l}ynarczyk , Gabriel Calvo , Francisco Palmi-Perales , Carmen Armero , Virgilio Gómez-Rubio , Ursula Martinez-Iranzo This is my paper · ORCID

classification stat.AP

keywords Bayesiannetworksdriverbehaviourmentalloadactivefatigueheartratevariabilityrespirationaffectivemoduleprobabilisticinference

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper builds a probabilistic affective module for a driver behavioural model by learning a Bayesian network from physiological and questionnaire data collected on 56 drivers. The network links heart-rate variability indices, mean heart rate, and respiration rate to two binary mental states, mental load and active fatigue, and the fitted structure places the mental states as direct influences on heart-function variables. The headline result is \(P(\mathrm{ML}=1\mid \mathrm{MeanHR}>100,\mathrm{RespRate}>20)=0.94\) and \(P(\mathrm{AF}=1\mid\text{same conditions})=0.78\). If the fitted probabilities are reliable, a vehicle could estimate a driver's mental state from live vital-sign measurements, a step toward fatigue alerts and safer autonomous-vehicle decisions.

What carries the argument

The central object is a Bayesian network, a probabilistic model represented by a directed acyclic graph in which nodes are variables and directed edges encode conditional dependencies, so the joint distribution factorizes as \(f(y)=\prod_v f(y_v\mid y_{\mathrm{pa}(Y_v)})\). The machinery does the work of breaking a seven-variable joint distribution into local conditional models: structure is selected with the Bayesian Information Criterion, parameters are estimated by maximum likelihood, and the fitted factors are recombined to answer conditional-probability queries about the mental states.

What would settle it

Train the network on any 55 of the 56 participants and use it to predict the held-out driver's observations that have mean heart rate above 100 and respiration above 20; if the predicted probability of mental load or active fatigue differs noticeably from the observed rate among held-out drivers, or if adding a per-driver random effect changes the selected graph, the model's real-world probability estimates are called into question.

Watch

Extended reading notes

Core claim

The central discovery is a directed acyclic graph in which mental load (ML) and active fatigue (AF) directly affect mean heart rate, SDSD, and the LF/HF ratio, which in turn influence SDNN and then respiration rate. The authors claim the joint distribution factorizes according to that graph, with continuous nodes modelled by conditional normal distributions whose parameters depend on the two mental-state indicators; for instance, mean heart rate given SDSD, ML, and AF follows a normal distribution with a mean that depends on the four combinations of ML and AF. Conditioning this fitted network on extreme physiological values yields the paper's headline probabilities: a 94% chance of mental load and a 78% chance of active fatigue when mean heart rate is above 100 beats per minute and respiration is above 20 breaths per minute.

Load-bearing premise

The claim rests on treating the 1,892 observations as 1,892 independent snapshots; if the repeated measurements from the same driver are correlated, the chosen network structure and the reported probabilities could be biased and look more certain than they are.

Editorial extensions

If this is right

A vehicle could query the network with live physiological readings and obtain an instantaneous probabilistic estimate of mental load or active fatigue without needing every sensor used in the study.
The learned graph indicates that mental states act on heart-function variables first and respiration only indirectly through SDNN, identifying heart-rate variability as the physiological bridge to mental state.
The affective module can supply probabilistic mental-state estimates to a larger driver behavioural model, so downstream nodes representing driving actions could consume these estimates.
The same network structure can be extended with additional markers such as facial expressions or with contextual variables such as road conditions and traffic density.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A test the paper does not report is leave-one-driver-out validation: train on 55 drivers and predict the held-out driver's observations; if the predicted 0.94 and 0.78 probabilities are not close to the observed rates, the estimates are specific to the training sample rather than to drivers generally.
The directed graph is a dependency structure, not a causal one; reading the arrows as causal would require additional assumptions, so the headline probabilities should be read as observational predictions rather than as evidence that changing heart rate changes mental state.
For a deployment that only monitors the two headline variables, much of the network could be marginalised away after fitting, leaving a small conditional model for the mental states given mean heart rate and respiration rate.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The paper is a straightforward application of Bayesian networks to a real driver dataset, and the new thing is the fitted model itself: the specific DAG and the conditional probability estimates for mental load and active fatigue from physiological signals. The writing is clear, the factorization is correct, and the authors deserve credit for reporting the correlation matrix and for using a standard, well-documented package (bnlearn) with BIC. If you want a quick example of how to fit a BN to physiological data and query posterior probabilities, this is a serviceable one.

The soft spot is real and load-bearing. The 1,892 observations come from 56 participants, about 34 per participant, and physiological and psychological measures within a person are usually correlated over time. The paper treats every observation as independent and identically distributed. That assumption is not defended anywhere. With clustering, the effective sample size is much closer to 56 than 1,892, so the BIC penalty is miscalibrated and the selected structure could be overfit or distorted. The headline numbers, P(ML=1 | HR>100, Resp>20)=0.94 and P(AF=1 | same)=0.78, are point estimates from that fit with no confidence intervals, no cluster-robust uncertainty, and no holdout or external validation. They are not independent predictions; they are posterior probabilities from the same model that was fitted to the same data. That doesn't make them wrong, but it does mean the paper provides no evidence they would generalize.

Two smaller issues add to the reproducibility problem. The binarization threshold for the 0-100 questionnaire responses is not reported; the paper just says the responses were "transformed into a binary format." And the model assigns an exact zero probability to one cell (ML=0, AF=1) under the query conditions, which suggests sparse data in the CPT and might be an artifact of the fitting rather than a genuine structural zero.

Despite these problems, the paper is honest about its scope and doesn't oversell the method. The network structure is physiologically plausible, and the authors frame the work as an affective module within a larger driver behaviour model, not as a finished product. The i.i.d. issue is fixable with a cluster bootstrap, participant-level random effects, or at least an ICC calculation. The binarization threshold and uncertainty intervals should also be reported.

Who is this for? Applied statisticians working on driver state estimation, and anyone who wants a clean worked example of BN structure learning on real physiological data. It deserves a serious referee: the approach is sound in principle, but the current evidence behind the central claims is thin. I would send it to peer review, with the expectation that the authors add a cluster-robust sensitivity analysis, disclose the threshold, and temper the real-time claims until validation appears.

Referee Report

4 major / 4 minor

Summary. The manuscript proposes a Bayesian network (BN) approach for modelling two binary driver mental states, mental load (ML) and active fatigue (AF), from physiological measurements (SDNN, SDSD, mean heart rate, LF/HF ratio, respiration rate) and demographic/questionnaire data. The network structure is learned using the bnlearn R package with BIC-based selection, and parameters are estimated by maximum likelihood. The paper's headline results are posterior probabilities computed from the fitted network, e.g., P(ML=1 | Mean HR>100, Resp rate>20)=0.94 and P(AF=1 | same condition)=0.78, along with the joint distribution of ML and AF under that evidence. The authors position this as an affective module for a driver behavioural model, with potential applications in traffic safety and autonomous vehicles.

Significance. The application is timely and relevant, and the BN framework is a natural fit for modelling uncertainty in driver mental states. The paper is transparent in presenting the factorization and the conditional parameter estimates, which aids reproducibility. However, the central quantitative claims are not currently supported: the repeated-measures structure of the data (1,892 observations from 56 participants) is ignored, the binarization threshold for the questionnaire responses is not disclosed, and the headline probabilities are in-sample posterior summaries with no validation or uncertainty quantification. As a proof-of-concept, the paper shows how a BN could be used, but the specific evidence for the 0.94 and 0.78 estimates is statistically insufficient in its present form.

major comments (4)

[Section IV-A] The 1,892 observations are drawn from only 56 participants, with roughly 34 observations per participant, yet the model treats each observation as independent and identically distributed in both BIC-based structure learning and maximum-likelihood parameter estimation. Within-participant correlation in physiological and self-report measures is typical in repeated-measures studies and, if present, invalidates the effective sample size used in the BIC penalty (closer to 56 than 1,892), potentially biasing the selected DAG and all derived probabilities. The paper should report intraclass correlations for the physiological variables, or fit a model with a participant-level random effect, and should compute cluster-robust confidence intervals for the headline probabilities.
[Section IV-A] The text states that the original questionnaire responses on a 0-100 continuous scale were transformed into a binary format (1 = state present, 0 = absent), but the threshold used for this binarization is never stated. This threshold is a modeling choice that directly determines the values of ML and AF, and thus affects the learned network structure and every reported conditional probability, including the 0.94 and 0.78. The threshold must be reported explicitly, and a sensitivity analysis over plausible threshold values should be provided.
[Section IV-B] The quantities Prob(ML=1 | Mean HR > 100, Resp rate > 20) = 0.94 and Prob(AF=1 | same evidence) = 0.78 are posterior probabilities computed from the same maximum-likelihood fit that was used to select the network structure. They are therefore in-sample descriptions of the fitted model, not out-of-sample predictions. To support the central claim that the model can estimate a driver's mental state from physiological conditions, the paper needs a validation step, such as a held-out test set or k-fold cross-validation, with predictive metrics (e.g., AUC, Brier score) and ideally a comparison against a null model.
[Section IV-B and Table III] All reported probabilities are point estimates with no measure of uncertainty. The cluster design (56 participants) and the large differences in estimated standard deviations across cells in Table II (e.g., σ = 4.432 for ML=0, AF=1 versus σ = 26.755 for ML=1, AF=1) suggest that some conditional distributions are estimated from small numbers of observations. A nonparametric bootstrap clustered by participant, or another resampling scheme that treats participant as the sampling unit, should be used to provide confidence intervals for the probabilities in Table III.

minor comments (4)

[Section II] The sentence describing the Bayesian approach to parameter estimation is redundant: 'requires assigning a prior distribution to θ which models uncertainty in probabilistic terms and requires a prior probability distribution to initiate the inferential process' repeats the same idea twice. Please rephrase.
[Figure 1 caption] The caption reads 'Basic direct acyclic graph example'; it should be 'Directed acyclic graph'.
[Table III] The probability P(ML=0, AF=1 | Mean HR > 100, Resp rate > 20) is reported as exactly 0.000. Since the model uses continuous conditional densities, an exact zero is surprising and may be a rounding artifact; if so, report the value as <0.001 or explain why the probability is exactly zero.
[Section IV-A] The paper motivates including all physiological variables by stating that correlations are 'not very high', but the correlation between SDNN and Mean HR is 0.412. A formal multicollinearity check, such as variance inflation factors, would be more convincing.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the reported probabilities are explicit in-sample queries of the fitted Bayesian network, not independent predictions used as evidence.

full rationale

The paper's derivation chain is: collect data → learn a Bayesian network structure via BIC → estimate conditional distributions by maximum likelihood → compute conditional probabilities such as P(ML=1 | Mean HR>100, Resp rate>20) from the fitted joint distribution. This is a standard statistical inference pipeline. The reported values (0.94 and 0.78) are not presented as out-of-sample predictions or as confirmations of the model; they are explicitly described as outputs of the estimated model ('The estimated model also allows us to assess the probability...'). No load-bearing step reduces to its own input by construction: the probabilities are plug-in calculations from the fitted parameters, not definitions of those parameters. There are no self-citations that carry the argument, no imported uniqueness theorems, and no ansatz smuggled via citation. The paper's substantive limitations—treating 1,892 observations from 56 participants as i.i.d. and lacking an external holdout benchmark—are important statistical validity concerns, but they are not circularity. Therefore the circularity score is 0.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The model is entirely data-driven: all conditional distributions are fitted by maximum likelihood to the same dataset, and the binary mental-state definitions depend on an unreported threshold. No external constraints or independent benchmarks are used.

free parameters (2)

Binarization threshold for ML and AF questionnaire responses = not reported
Continuous 0-100 mental state responses are converted to binary 1 or 0, but the cutoff is not stated. All downstream probability estimates depend on this threshold.
Conditional probability parameters of all BN nodes = partially reported in Table II for Mean HR
All node parameters are estimated by maximum likelihood from the same 1,892 observations; the headline probabilities 0.94 and 0.78 are functions of these fitted values.

assumptions (4)

standard math Bayesian network joint distribution factorizes as the product of local conditional distributions.
Invoked in Section II to justify the model.
domain assumption ML and AF are binary latent states that directly influence physiological variables.
Section IV-B places ML and AF as root nodes with edges to SDSD, Mean HR, and LF HF ratio; this structure is learned or assumed but not independently validated.
domain assumption All 1,892 observations are independent draws, ignoring clustering by 56 participants.
Section IV-A describes repeated measures from 56 participants; no random effects or cluster structure is included in the analysis.
domain assumption The BIC-selected network structure adequately approximates the true dependency structure.
The paper selects one DAG by BIC without reporting structure uncertainty or validating the chosen graph.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Bayesian network approach to building an affective module for a driver behavioural model." pith.science (2026). https://pith.science/paper/TX36UJLM

@misc{pith2026250203254,
  author       = {Pith},
  title        = {Pith review of: Bayesian network approach to building an affective module for a driver behavioural model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TX36UJLM}},
  note         = {Machine review of arXiv:2502.03254}
}

read the original abstract

This paper focuses on the affective component of a Driver Behavioural Model (DBM), specifically modelling some driver's mental states, such as mental load and active fatigue, which may affect driving performance. We used Bayesian networks (BNs) to explore the dependencies between various relevant variables and estimate the probability that a driver was in a particular mental state based on their physiological and demographic conditions. Through this approach, our goal is to improve our understanding of driver behaviour in dynamic environments, with potential applications in traffic safety and autonomous vehicle technologies.

Figures

Figures reproduced from arXiv: 2502.03254 by the authors.

**Figure 1.** Basic direct acyclic graph example. Dynamic Bayesian Networks (DBNs) extend the framework of static BNs by introducing a temporal dimension into the network structure [5]. This enhancement allows DBNs to model time-varying systems more effectively, capturing the evolution of variables and their dependencies over time. While DBNs are particularly well-suited to systems with temporal or sequential dynamics, the inclus… view at source ↗

**Figure 2.** Histograms illustrating the basic characteristics of the study partici [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 4.** Mean (µ) of the conditional distribution of the Mean HR variable with regard to SDSD for the four possible scenarios defined by ML and AF. The estimated model also allows us to assess the probability that somebody is in a particular mental state given the values of a subset of the other variables. For example, the probability that a person is under mental load given that their mean heart rate is over 100 and their r… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

13 extracted references · 13 canonical work pages

[1]

From Bayesian Networks to Causal Networks

Pearl, J., “From Bayesian Networks to Causal Networks”. In *Coletti, G., Dubois, D., Scozzafava, R. (eds) Mathematical Models for Handling Partial Knowledge in Artificial Intelligence*. Springer, 1995

work page 1995
[2]

Bayesian Networks,

J. Pearl and S. Russell, “Bayesian Networks,” in *Handbook of Brain Theory and Neural Networks*, M. A. Arbib, Ed., MIT Press, 1995

work page 1995
[3]

The BATmobile: Towards a Bayesian automated taxi,

J. Forbes, T. Huang, K. Kanazawa, and S. Russell, “The BATmobile: Towards a Bayesian automated taxi,” in *IJCAI*, vol. 95, pp. 1878– 1885, 1995

work page 1995
[4]

Driver Behavior Modeling Toward Au- tonomous Vehicles: Comprehensive Review,

Negash, N. M. and Yang, J.,“Driver Behavior Modeling Toward Au- tonomous Vehicles: Comprehensive Review,” *IEEE Acces*, vol. 11, pp. 22788–22821, 2023

work page 2023
[5]

A model for reasoning about persistence and causation,

T. Dean and K. Kanazawa, “A model for reasoning about persistence and causation,” *Computational Intelligence*, vol. 5, no. 2, pp. 142–150, 1989

work page 1989
[6]

Core Statistics

Wood, S. N., “ Core Statistics”, Cambridge University Press, 2015

work page 2015
[7]

Quantification of uncertainty and its applications to complex domain for autonomous vehicles percep- tion system,

K. Wang, Y . Wang, B. Liu, and J. Chen, “Quantification of uncertainty and its applications to complex domain for autonomous vehicles percep- tion system,” *IEEE Trans. Instrum. Meas.*, vol. 72, pp. 1–17, 2023

work page 2023
[8]

The analysis of driver’s behavioral tendency under different emotional states based on a Bayesian network,

Y .-Q. Liu and X.-Y . Wang, “The analysis of driver’s behavioral tendency under different emotional states based on a Bayesian network,” *IEEE Trans. Affect. Comput.*, vol. 14, no. 1, pp. 165–177, 2020

work page 2020

Show all 13 references

[9]

Driver fatigue evaluation model with integration of multi-indicators based on dynamic Bayesian network,

Q. He, W. Li, X. Fan, and Z. Fei, “Driver fatigue evaluation model with integration of multi-indicators based on dynamic Bayesian network,” *IET Intell. Transp. Syst.*, vol. 9, no. 5, pp. 547–554, 2015

2015
[10]

A driver fatigue recognition model based on information fusion and dynamic Bayesian network,

G. Yang, Y . Lin, and P. Bhattacharya, “A driver fatigue recognition model based on information fusion and dynamic Bayesian network,” *Inf. Sci.*, vol. 180, no. 10, pp. 1942–1954, 2010

1942
[11]

and Pivik, K., ”Age and gender differences in risky driving: the roles of positive affect and risk perception,” *Accid

Rhodes, N. and Pivik, K., ”Age and gender differences in risky driving: the roles of positive affect and risk perception,” *Accid. Anal. Prev.*, vol. 43, no. 3, pp. 923–931, 2011

2011
[12]

Learning Bayesian Networks with the bnlearn R Package,

M. Scutari, “Learning Bayesian Networks with the bnlearn R Package,” *Journal of Statistical Software*, vol. 35, no. 3, pp. 1–22, 2010

2010
[13]

The Bayesian information criterion: Background, derivation, and applications,

Neath, A. A. and Cavanaugh, J., “The Bayesian information criterion: Background, derivation, and applications,” Wiley Interdiscip. Rev. Com- put. Stat., vol. 4, no. 1, pp. 99–106, 2012

2012

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.

[1] [1]

From Bayesian Networks to Causal Networks

Pearl, J., “From Bayesian Networks to Causal Networks”. In *Coletti, G., Dubois, D., Scozzafava, R. (eds) Mathematical Models for Handling Partial Knowledge in Artificial Intelligence*. Springer, 1995

work page 1995

[2] [2]

Bayesian Networks,

J. Pearl and S. Russell, “Bayesian Networks,” in *Handbook of Brain Theory and Neural Networks*, M. A. Arbib, Ed., MIT Press, 1995

work page 1995

[3] [3]

The BATmobile: Towards a Bayesian automated taxi,

J. Forbes, T. Huang, K. Kanazawa, and S. Russell, “The BATmobile: Towards a Bayesian automated taxi,” in *IJCAI*, vol. 95, pp. 1878– 1885, 1995

work page 1995

[4] [4]

Driver Behavior Modeling Toward Au- tonomous Vehicles: Comprehensive Review,

Negash, N. M. and Yang, J.,“Driver Behavior Modeling Toward Au- tonomous Vehicles: Comprehensive Review,” *IEEE Acces*, vol. 11, pp. 22788–22821, 2023

work page 2023

[5] [5]

A model for reasoning about persistence and causation,

T. Dean and K. Kanazawa, “A model for reasoning about persistence and causation,” *Computational Intelligence*, vol. 5, no. 2, pp. 142–150, 1989

work page 1989

[6] [6]

Core Statistics

Wood, S. N., “ Core Statistics”, Cambridge University Press, 2015

work page 2015

[7] [7]

Quantification of uncertainty and its applications to complex domain for autonomous vehicles percep- tion system,

K. Wang, Y . Wang, B. Liu, and J. Chen, “Quantification of uncertainty and its applications to complex domain for autonomous vehicles percep- tion system,” *IEEE Trans. Instrum. Meas.*, vol. 72, pp. 1–17, 2023

work page 2023

[8] [8]

The analysis of driver’s behavioral tendency under different emotional states based on a Bayesian network,

Y .-Q. Liu and X.-Y . Wang, “The analysis of driver’s behavioral tendency under different emotional states based on a Bayesian network,” *IEEE Trans. Affect. Comput.*, vol. 14, no. 1, pp. 165–177, 2020

work page 2020