pith. sign in

arxiv: 2606.21154 · v1 · pith:TQMZVFEJnew · submitted 2026-06-19 · 💻 cs.CY · cs.SY· eess.SY· stat.AP

Virginia Tech Transportation Safety Index (VTTSI)

Jason Cusati , Cheng-Shun Chuang This is my paper

Pith reviewed 2026-06-26 12:55 UTC · model grok-4.3

classification 💻 cs.CY cs.SYeess.SYstat.AP
keywords real-time safety indexintersection safetyEmpirical Bayesmulti-criteria decision makingconnected vehicle telemetryexposure-adjusted scoresspeed varianceCRITIC weighting
0
0 comments X

The pith

VTTSI produces 0-100 safety scores for intersections every 15 minutes by fusing long-term crash data with real-time speed and conflict telemetry.

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

Traditional crash-based safety assessments depend on aggregated historical data that cannot track fast shifts in traffic conditions. VTTSI addresses this by stabilizing multi-year crash records through Empirical Bayes methods, then layering uplift adjustments drawn from current vehicle speed variance and conflict behavior. These inputs feed a CRITIC-weighted multi-criteria module that combines several decision techniques into a single exposure-adjusted score. The resulting 0-100 index updates every fifteen minutes and supports visualization of volumes, vulnerable road user exposure, and incidents.

Core claim

The Virginia Tech Transportation Safety Index integrates Empirical Bayes crash stabilization with uplift factors derived from speed and conflict behavior, processed through a CRITIC-weighted multi-criteria decision-making module that combines SAW, EDAS, and CODAS, to deliver interpretable, exposure-adjusted safety scores on a 0-100 scale every 15 minutes from multimodal connected-vehicle telemetry and multi-year crash history.

What carries the argument

Hybrid modeling approach that fuses Empirical Bayes crash stabilization, uplift factors derived from speed and conflict behavior, and a CRITIC-weighted multi-criteria decision-making module combining SAW, EDAS, and CODAS.

If this is right

  • Scores reflect both long-term crash risk and short-term behavioral dynamics at each update interval.
  • Validation across sites shows coherent diurnal patterns, consistency among the MCDM methods, and sensitivity to operational turbulence.
  • Sensitivity analysis indicates the index remains stable, with typical deviations below one point on the 0-100 scale under parameter changes.
  • The framework supports use by transportation agencies, traffic management centers, fleet operators, and autonomous vehicle systems through cloud visualization.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The 15-minute update cycle could allow traffic signals or variable message signs to respond directly to rising risk levels.
  • Extension to non-intersection segments would require only modest changes to the exposure and conflict detection layers.
  • Comparing VTTSI outputs against purely surrogate-safety indices over the same sites would quantify the contribution of the crash-stabilization component.
  • Deployment across additional cities with different crash-reporting practices would test whether the uplift factors transfer without retraining.

Load-bearing premise

Uplift factors derived from speed and conflict behavior, when combined via CRITIC-weighted MCDM, accurately capture rapidly changing operational conditions beyond what lagged crash data already provide.

What would settle it

If real-time 0-100 scores show no measurable correlation with observed near-miss rates or incident spikes during periods of elevated speed variance or conflict activity, the hybrid method would fail to demonstrate added value over crash history alone.

Figures

Figures reproduced from arXiv: 2606.21154 by Cheng-Shun Chuang, Jason Cusati.

Figure 1
Figure 1. Figure 1: Methodology overview To choose λ, we treat 2017–2024 as training data and 2025 as a hold-out set and minimize the Poisson log-loss L(λ) = X i,t∈2025  Yˆ (2025) i,t (λ) − Y (2025) i,t log Yˆ (2025) i,t (λ)  , (5) over a log-spaced grid λ ∈ {0.1, 0.3, 1, 3, 10, 30, 100, 300, 1000, 10000, 30000, 100000}. (6) After extending this grid, the optimal value was found to be λ ⋆ = 100000. In production we fix λ = … view at source ↗
Figure 2
Figure 2. Figure 2: Refined cloud-native system architecture of VTTSI. The left pathway shows the safety-index pipeline: Streamlit analysis pages call the FastAPI REST [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Final blended Safety Index at the Glebe–Potomac intersection [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: RT–SI scoring methods trend comparison at the Glebe–Potomac [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 9
Figure 9. Figure 9: Comparison of perturbed RT–SI trajectories with the baseline at the [PITH_FULL_IMAGE:figures/full_fig_p009_9.png] view at source ↗
Figure 8
Figure 8. Figure 8: Near-miss count at the Glebe–Potomac intersection [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗
Figure 10
Figure 10. Figure 10: Distribution of absolute deviations between baseline and perturbed [PITH_FULL_IMAGE:figures/full_fig_p010_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Correlation between parameter perturbations and RT–SI deviations [PITH_FULL_IMAGE:figures/full_fig_p010_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: All Variables Normalized to 0-100 scale for trend analysis on Birch_St W_Broad_St intersection [PITH_FULL_IMAGE:figures/full_fig_p015_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: All Variables Normalized to 0-100 scale for trend analysis on E_Broad_St N_Washinton_St intersection [PITH_FULL_IMAGE:figures/full_fig_p015_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Final Blended Safety Index on Birch_St W_Broad_ Stintersection [PITH_FULL_IMAGE:figures/full_fig_p015_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Final Blended Safety Index on E_Broad_St N_Washinton_St intersection [PITH_FULL_IMAGE:figures/full_fig_p016_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: MCDM methods trend comparison on Birch_St W_Broad_St intersection [PITH_FULL_IMAGE:figures/full_fig_p016_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: MCDM methods trend comparison on E_Broad_St N_Washinton_St intersection [PITH_FULL_IMAGE:figures/full_fig_p016_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: RT-SI scoring methods trend comparison on Birch_St W_Broad_St intersection [PITH_FULL_IMAGE:figures/full_fig_p017_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: RT-SI scoring methods trend comparison on E_Broad_St N_Washinton_St intersection [PITH_FULL_IMAGE:figures/full_fig_p017_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Comparison of perturbations with baseline RT-SI scoring on Birch_St W_Broad_St intersection [PITH_FULL_IMAGE:figures/full_fig_p017_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: Comparison of perturbations with baseline RT-SI scoring on E_Broad_St N_Washinton_St intersection [PITH_FULL_IMAGE:figures/full_fig_p018_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: Distribution of absolute deviations between baseline and perturbations on Birch_St W_Broad_St intersection [PITH_FULL_IMAGE:figures/full_fig_p018_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: Distribution of absolute deviations between baseline and perturbations on E_Broad_St N_Washinton_St intersection [PITH_FULL_IMAGE:figures/full_fig_p018_23.png] view at source ↗
Figure 24
Figure 24. Figure 24: Correlation between parameter perturbations and RT–SI score deviations on Birch_St W_Broad_St intersection [PITH_FULL_IMAGE:figures/full_fig_p019_24.png] view at source ↗
Figure 25
Figure 25. Figure 25: Correlation between parameter perturbations and RT–SI score deviations on E_Broad_St N_Washinton_St intersection [PITH_FULL_IMAGE:figures/full_fig_p019_25.png] view at source ↗
read the original abstract

The Virginia Tech Transportation Safety Index (VTTSI) is a real-time, cloud-native framework for quantifying intersection safety using multimodal connected-vehicle telemetry and multi-year VDOT crash history. Traditional crash-based methods rely on lagged, aggregated data and cannot reflect rapidly changing operational conditions. VTTSI addresses this gap through a hybrid modeling approach that fuses Empirical Bayes (EB) crash stabilization, uplift factors derived from speed and conflict behavior, and a CRITIC-weighted multi-criteria decision-making (MCDM) module combining SAW, EDAS, and CODAS. The system produces interpretable, exposure-adjusted safety scores on a 0--100 scale every 15 minutes. A cloud-deployed architecture built on FastAPI, PostgreSQL, PostGIS, and Streamlit supports interactive visualization of traffic volumes, VRU exposure, speed variance, and real-time incident activity. Validation across intersections demonstrates coherent diurnal patterns, consistency among MCDM methods, and sensitivity to observable operational turbulence. Sensitivity analysis further shows that the RT--SI is robust to parameter perturbations, with deviations typically remaining below one point on the 0--100 scale. By integrating long-term crash risk with short-term behavioral dynamics, VTTSI provides a transparent, adaptive, and proactive safety-monitoring framework suitable for transportation agencies, traffic management centers, fleet operators, and autonomous vehicle systems.%~\cite{persaud2007, montella2020systemic, schultz2025_surrogate, Amraji2025CombinedSafetyIndex}.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 1 minor

Summary. The manuscript presents the Virginia Tech Transportation Safety Index (VTTSI), a cloud-native framework that generates real-time (every 15 minutes) exposure-adjusted intersection safety scores on a 0-100 scale. It fuses Empirical Bayes (EB) crash stabilization with uplift factors derived from speed and conflict behavior, combined via CRITIC-weighted multi-criteria decision-making (MCDM) methods (SAW, EDAS, CODAS). The system is implemented with FastAPI, PostgreSQL, PostGIS, and Streamlit for visualization, and validation shows coherent diurnal patterns, MCDM consistency, sensitivity to operational turbulence, and robustness to parameter changes (deviations typically <1 point).

Significance. If the hybrid model can be shown to add predictive value beyond EB alone, VTTSI would provide a transparent, proactive tool for real-time safety monitoring that integrates long-term crash risk with short-term behavioral dynamics, with potential utility for agencies, traffic centers, and AV systems.

major comments (1)
  1. [Validation] Validation description (abstract and results): The reported checks (coherent diurnal patterns, MCDM consistency, sensitivity to turbulence, parameter robustness <1 point) are internal consistency tests only. No quantitative out-of-sample comparison is provided demonstrating that the hybrid model (EB + uplift factors via CRITIC-MCDM) outperforms an EB-only baseline on held-out crash data or other external criterion. This is load-bearing for the central claim that uplift factors capture rapidly changing conditions beyond lagged crash data.
minor comments (1)
  1. [Abstract] Abstract: The final sentence contains a LaTeX citation remnant ('%~\cite{persaud2007, montella2020systemic, ...}') that should be removed or properly formatted.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the detailed review and constructive feedback on the VTTSI manuscript. We address the single major comment below and outline planned revisions to strengthen the validation section.

read point-by-point responses

  1. Referee: Validation description (abstract and results): The reported checks (coherent diurnal patterns, MCDM consistency, sensitivity to turbulence, parameter robustness <1 point) are internal consistency tests only. No quantitative out-of-sample comparison is provided demonstrating that the hybrid model (EB + uplift factors via CRITIC-MCDM) outperforms an EB-only baseline on held-out crash data or other external criterion. This is load-bearing for the central claim that uplift factors capture rapidly changing conditions beyond lagged crash data.

    Authors: We agree that the current validation emphasizes internal consistency, sensitivity to operational turbulence, and parameter robustness rather than a direct quantitative out-of-sample comparison of predictive performance against an EB-only baseline on held-out crash data. The hybrid design is motivated by the need to integrate short-term behavioral signals (speed variance and conflicts) that historical crash data alone cannot reflect in real time; however, we acknowledge that demonstrating incremental predictive value on an external criterion would provide stronger support for this claim. In the revised manuscript we will add a dedicated subsection performing such a comparison where data permits (e.g., using temporal or spatial hold-outs of recent crash records) or, if data constraints limit direct comparison, will explicitly discuss this limitation and its implications for the framework's claims. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected in derivation chain

full rationale

The paper's hybrid approach fuses established Empirical Bayes crash stabilization with uplift factors computed from speed and conflict telemetry plus standard CRITIC-weighted MCDM (SAW/EDAS/CODAS) to produce 15-minute safety scores. No equations, parameter-fitting steps, or self-citation chains are shown that reduce the final index to its own inputs by construction (e.g., no case of fitting uplift factors to crash data then claiming the same factors as an independent prediction). Validation consists of internal consistency checks rather than external benchmarks, but the central modeling claim retains independent content from the fusion of lagged crash data with real-time behavioral signals. No load-bearing self-citations or ansatzes smuggled via prior work appear in the provided text.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review yields insufficient detail to enumerate specific free parameters, axioms, or invented entities; the hybrid fusion and weighting steps are described at a high level without explicit fitting procedures or background assumptions listed.

pith-pipeline@v0.9.1-grok · 5806 in / 1256 out tokens · 21599 ms · 2026-06-26T12:55:18.733709+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

27 extracted references · 6 canonical work pages

  1. [1]

    Sidey-Gibbons, Charlotte Sun, Alina Schneider, et al

    Ahmed Almutairi et al. “Deep Learning Based Pre- dictive Models for Real-Time Accident Prevention in Autonomous Vehicle Networks”. In:Scientific Reports 15.1 (2025), p. 20844.doi: 10.1038/s41598- 025- 04867-8.url: https://doi.org/10.1038/s41598-025- 04867-8

    work page doi:10.1038/s41598- 2025

  2. [2]

    Combined Safety Index Using CRITIC Weighting

    A. Amraji. “Combined Safety Index Using CRITIC Weighting”. In:Journal of Transportation Research 15.2 (2025), pp. 200–210

    2025

  3. [3]

    Jason Cusati.Agentic Knowledge Graphs for Traffic Safety. Tech. rep. Unpublished manuscript. Virginia Tech, 2025

    2025

  4. [4]

    FHWA, 2003

    Douglas Gettman and Larry Head.Surrogate Safety As- sessmentModelandValidation:FinalReport.Tech.rep. FHWA, 2003

    2003

  5. [5]

    High- way Crash Detection and Risk Estimation Using Deep Learning

    Tingting Huang, Shuo Wang, and Anuj Sharma. “High- way Crash Detection and Risk Estimation Using Deep Learning”. In:Accident Analysis & Prevention135 (2020), p. 105392.doi: 10.1016/j.aap.2019.105392

    work page doi:10.1016/j.aap.2019.105392 2020

  6. [6]

    A New Combina- tive Distance-Based Assessment (CODAS) Method for Multi-Criteria Decision-Making

    Mehdi Keshavarz-Ghorabaee et al. “A New Combina- tive Distance-Based Assessment (CODAS) Method for Multi-Criteria Decision-Making”. In:Economic Com- putation and Economic Cybernetics Studies and Re- search50 (2016), pp. 25–44

    2016

  7. [7]

    Multi-Criteria In- ventory Classification Using a New Method of Eval- uation Based on Distance from Average Solution (EDAS)

    Mehdi Keshavarz-Ghorabaee et al. “Multi-Criteria In- ventory Classification Using a New Method of Eval- uation Based on Distance from Average Solution (EDAS)”. In:Informatica26 (2015), pp. 435–451.doi: 10.15388/Informatica.2015.57

    work page doi:10.15388/informatica.2015.57 2015

  8. [8]

    Systemic Approach to Improve Safety of Urban Unsignalized Intersections: Develop- ment and Validation of a Safety Index

    Alfonso Montella et al. “Systemic Approach to Improve Safety of Urban Unsignalized Intersections: Develop- ment and Validation of a Safety Index”. In:Accident Analysis & Prevention141 (2020), p. 105523.doi: 10. 1016/j.aap.2020.105523

    arXiv 2020

  9. [9]

    IoT-Enabled Adaptive Traffic Management: A Multiagent Framework for Urban Mo- bility Optimisation

    Iqbal Mutambik et al. “IoT-Enabled Adaptive Traffic Management: A Multiagent Framework for Urban Mo- bility Optimisation”. In:Sensors25.13 (2025), p. 4126. doi: 10.3390/s25134126

    work page doi:10.3390/s25134126 2025

  10. [10]

    Empirical Bayes Before–After Safety Studies: Lessons Learned from Two Decades of Experience and Future Directions

    Bhagwant Persaud and Craig Lyon. “Empirical Bayes Before–After Safety Studies: Lessons Learned from Two Decades of Experience and Future Directions”. In: Accident Analysis & Prevention39.3 (2007), pp. 546– 555.doi: 10.1016/j.aap.2006.09.009

    work page doi:10.1016/j.aap.2006.09.009 2007

  11. [11]

    Assessing the Safety Effect Through Google Maps Usage: FMEA Approach (Case Study: Indonesia)

    A. F. D. Pribadi, A. D. Oktaviani, and A. S. Utomo. “Assessing the Safety Effect Through Google Maps Usage: FMEA Approach (Case Study: Indonesia)”. In: Case Studies on Transport Policy10.1 (2022), pp. 42– 53

    2022

  12. [12]

    A Study on Surrogate Safety Indicators

    John Schultz. “A Study on Surrogate Safety Indicators”. In:Journal of Transportation Safety. Vol. 12. 3. 2025, pp. 123–134

    2025

  13. [13]

    A High-Resolution Trajectory Data Driven Method for Real-Time Evaluation of Traffic Safety

    Y. Wang et al. “A High-Resolution Trajectory Data Driven Method for Real-Time Evaluation of Traffic Safety”. In:Accident Analysis & Prevention160 (2021), p. 106323

    2021

  14. [14]

    Evaluating the Use of Google Maps as Navigation Application by Identifying Hazards and Assessing Risks Using HIRA Matrix

    S. H. Wicaksono, A. Nugroho, and R. R. R. Putra. “Evaluating the Use of Google Maps as Navigation Application by Identifying Hazards and Assessing Risks Using HIRA Matrix”. In: 1064.1 (2022), p. 012020

    2022

  15. [15]

    Dynamic Spatiotemporal Graph Network for Traffic Accident Risk Predic- tion

    Pengcheng Zhang et al. “Dynamic Spatiotemporal Graph Network for Traffic Accident Risk Predic- tion”. In:GIScience & Remote Sensing62.1 (2025), p. 2514330.doi: 10.1080/15481603.2025.2514330. url: https://doi.org/10.1080/15481603.2025.2514330. Appendix A Backend Service Descriptions TABLE II Core Backend Services Service Description MCDM Safety In- dex Ser...

    work page doi:10.1080/15481603.2025.2514330 2025

  16. [16]

    We adopt fixed policy- driven weights: wfatal = 10, w injury = 3, w PDO = 1, which match the implementation used to build severity- weighted crash counts

    Severity Weightsws:Severity weighting is applied only in the historical crash-rate computation. We adopt fixed policy- driven weights: wfatal = 10, w injury = 3, w PDO = 1, which match the implementation used to build severity- weighted crash counts

  17. [17]

    A grid search overλ∈{0.1,...,100000} minimizes Poisson negative log-likelihood when predicting 2025 outcomes

    Empirical Bayes Shrinkage Parameterλ:The historical crash rate for a 15-minute bin is stabilized using a no-exposure Empirical Bayes (EB) estimator: ˆri = Yi +λr0 1 +λ, whereYi is the severity-weighted crash count in bini,r0 is the pooled mean over 2017–2024, andλis selected using temporal cross-validation. A grid search overλ∈{0.1,...,100000} minimizes P...

    2017

  18. [18]

    a) Speed deficit uplift.:LetFFSdenote the free-flow speed andvthe observed mean speed: Fspeed = min ( 1,1.5 FFS−v FFS )

    Real-Time Uplift Factors(Fspeed,Fvar,Fconf):The real- time Safety Index multiplies the EB baseline by three bounded uplift factors derived from speed, turbulence, and VRU– vehicle conflict exposure. a) Speed deficit uplift.:LetFFSdenote the free-flow speed andvthe observed mean speed: Fspeed = min ( 1,1.5 FFS−v FFS ) . b) Speedvarianceuplift.:Letσ 2 bethe...

  19. [19]

    These values are policy parameters, not learned

    VRU–Vehicle Index Blendω:The real-time index blends VRU-specific and vehicle-specific sub-indices using fixed weights: SIRT = 0.6 SIVRU + 0.4 SIVEH. These values are policy parameters, not learned

  20. [20]

    MCDM Criterion Weightswj (CRITIC Only):For the MCDM component, each 15-minute intersection–time pair is treated as an alternative, and the following five real-time criteria form the decision matrix: C={vehicle_count,vru_count,avg_speed, speed_variance,incident_count}. (43) After min–max normalization, CRITIC weighting is ap- plied: Cj =σj ∑ k (1−ρjk), w j...

  21. [21]

    Method weights are recalculated dynamically for each evalu- ation window

    Hybrid Method WeightsWE,WC,WS (CRITIC Again): After computing SAW, EDAS, and CODAS scores (all nor- malized to[0,100]), these three method outputs form a matrix M= [EDAS a,CODAS a,SAW a], from which CRITIC weights are again derived: Wj = σj ∑ k(1−ρjk)∑ jσj ∑ k(1−ρjk), j∈{EDAS,CODAS,SAW}. Method weights are recalculated dynamically for each evalu- ation wi...

  22. [22]

    (42),SIFinal i,t =αSIRT i,t + (1−α)SI MCDM i

    Final Blend:The released backend computes both the real-time (RT–SI) and MCDM indices and combines them into the blended Final Safety Index of Eq. (42),SIFinal i,t =αSIRT i,t + (1−α)SI MCDM i . The blend coefficient defaults toα= 0.7and is exposed both as a request parameter on the safety-index API endpoint and as an interactive slider in the Streamlit fr...

  23. [23]

    Cross-Validation:Only the Empirical Bayes shrinkage parameterλuses temporal cross-validation:

  24. [24]

    Use 2017–2024 severity-weighted crash counts as train- ing data

    2017

  25. [25]

    Compute pooled meanr 0

  26. [26]

    b) Predict 2025 counts

    For eachλin a predefined grid: a) Compute EB-stabilized rates for training years. b) Predict 2025 counts. c) Compute Poisson log-loss

    2025

  27. [27]

    Weights for uplift factors, VRU–vehicle blending, and MCDM are not tuned via cross-validation

    Selectλ⋆minimizing log-loss; hardcode into real-time service. Weights for uplift factors, VRU–vehicle blending, and MCDM are not tuned via cross-validation. Validation additional images from other intersection Results for Birch and Broad intersections exhibit the same behavior and are included in Appendix figures. Figure 12 and figure 13 are all variables...