Forecasting the fracture: a cost-driven machine learning framework for optimal bridge maintenance prioritization

Abstract

Aging civil infrastructure presents a critical economic and public safety challenge, with maintenance backlogs costing hundreds of billions of dollars. This study moves beyond simple condition prediction to develop and validate a comprehensive, cost-driven framework for optimizing bridge maintenance schedules. Leveraging a 5-year (2020–2024) longitudinal cohort constructed from the U.S. National Bridge Inventory (NBI), we train a tuned gradient-boosted model (XGBoost) to predict the one-year-ahead probability of a bridge transitioning into a “Poor” or “Failing” condition. This probabilistic forecast is then integrated into a decision-theoretic framework that explicitly weighs the expected cost of failure against the fixed cost of proactive maintenance. By simulating this framework on a held-out test set, we identify an optimal, risk-based decision threshold that maximizes net cost savings. The economic simulation reveals a positive but modest net savings, underscoring the critical dependence of such strategies on model precision. To ensure transparency and build stakeholder trust, Explainable AI (SHAP) is used to dissect the model’s logic, confirming that deck condition, traffic volume, and bridge age are the primary drivers of its predictions, aligning with established engineering principles. This work provides a rigorous, scalable, and fully articulated methodology for turning predictive insights into economically optimal, data-driven policy for critical infrastructure management, while also quantifying the profound impact of model precision on real-world economic viability.

1 Introduction

The state of American infrastructure has reached a critical inflection point, evolving from a subject of policy debate into a pressing national concern. The American Society of Civil Engineers’ (ASCE) 2021 Report Card for America’s Infrastructure assigned a grade of “C-” to the nation’s bridges, a stark assessment underscoring a systemic issue: 7.5% of the nation’s bridges are considered structurally deficient, and 42% have surpassed their 50-year design life (ASCE, 2021). The estimated cost to address this accumulated maintenance and replacement backlog has soared to $125 billion. This staggering figure represents not just a financial liability but a significant public safety risk. The core challenge for transportation agencies and infrastructure managers is not a lack of data—the National Bridge Inventory (NBI) contains millions of records—but the persistent difficulty of translating this vast repository of data into economically optimal, proactive decisions.

Traditional Bridge Management Systems (BMS), while foundational, have historically relied on deterministic, time-based inspection schedules and largely reactive maintenance strategies (Bektas and an, 2012). This approach, often treating all bridges within a class as deteriorating at a similar rate, can lead to a profound misallocation of scarce resources. Healthy bridges may be inspected with unnecessary frequency, while high-risk bridges may be overlooked until their condition becomes critical, leading to exorbitant repair costs or, in the worst case, catastrophic failure.

The central thesis of this paper is that the synthesis of modern machine learning, classic economic decision theory, and explainable AI can catalyze a paradigm shift from reactive to predictive, cost-optimized infrastructure management. Simply predicting that a bridge is likely to deteriorate is an academic exercise with limited practical utility. The critical question for policymakers and engineers is, “Given a probabilistic prediction of failure, what is the most cost-effective action to take?” This study addresses this question directly by developing and evaluating an integrated decision-support framework.

Our contribution is threefold. First, we develop a robust longitudinal binary classification model using XGBoost (Chen and Guestrin, 2016), a state-of-the-art algorithm, to predict the probability of a bridge entering a “Poor” or “Failing” state within the subsequent year. Second, we introduce and apply a formal decision-theoretic framework that uses these probabilistic outputs to identify an optimal “intervention threshold” by balancing the cost of proactive maintenance against the expected cost of failure. This moves the analysis beyond simple model accuracy and into the realm of economic utility. Third, we provide a clear, visual analysis of these economic trade-offs and leverage Explainable AI (XAI) techniques, specifically SHAP (SHapley Additive exPlanations) (Lundberg and Lee, 2017), to dissect the model’s logic. This final step is crucial for ensuring that the model’s recommendations are transparent, defensible, and ultimately trusted by the stakeholders responsible for implementing them.

2 Literature review

This research is situated at the confluence of three distinct but increasingly intertwined fields: traditional bridge management, machine learning for prognostics, and decision science in engineering. Our work seeks to synthesize advances in these areas to address a persistent gap in the literature.

2.1 Evolution of bridge management and structural health monitoring

The systematic management of bridge infrastructure has its roots in the development of early BMS like PONTIS and BRIDGIT in the 1990s. These systems primarily employed Markov chain models to forecast future bridge conditions (Madanat et al., 1997; Sobanjo, 2010). In these models, a bridge’s condition is represented by a discrete state, and its deterioration is modeled as a probabilistic transition from one state to a worse one over time. While foundational, these Markovian models possess inherent limitations, such as the memoryless property, which assumes future deterioration depends only on the current state, not its history, and their difficulty in capturing complex, non-linear interactions between a multitude of predictive variables (e.g., traffic, climate, material) (Wellalage et al., 2021).

Concurrently, the field of Structural Health Monitoring (SHM) emerged, focusing on the deployment of physical sensors on high-value bridges to provide continuous, high-fidelity data on structural response (Farrar and Worden, 2012; Sony et al., 2019). While SHM provides unparalleled insight for individual structures, its instrumentation and data management costs are prohibitive for system-wide application across an entire national inventory, confining its use to a small number of iconic or uniquely critical bridges (Catbas et al., 2008). This leaves a vast “data gap” for the hundreds of thousands of conventional bridges that a BMS must manage.

Recent bridge Structural Health Monitoring (SHM) research has increasingly paired continuous sensing with advanced data-driven algorithms to automate modal identification and improve damage detection reliability. For example, Civera et al. (2023) propose an automated operational modal analysis workflow that uses DBSCAN clustering to robustly extract modal parameters from ambient vibration data for bridge monitoring, reducing manual intervention and improving outlier handling. Complementarily, Martucci et al. (2023) demonstrate a vibration-based SHM approach that couples Gaussian Process Regression with Extreme Function Theory for mode-shape damage detection on the I-40 bridge case study, explicitly targeting reduction of false alarms in operational conditions. Together, these works illustrate how sensor-rich SHM pipelines can enable high-fidelity, structure-specific diagnostics, while also underscoring practical scalability constraints (instrumentation, deployment, data stewardship) that motivate the present study’s system-level, inspection-data-driven alternative using the National Bridge Inventory.

2.2 The application of machine learning to infrastructure prognostics

The last decade has seen a surge in the application of machine learning (ML) to overcome the limitations of traditional models. Unlike Markov chains, which require pre-defined transition probabilities, ML models can learn complex, non-linear patterns directly from historical data (Flint et al., 2020). Researchers have successfully applied a range of algorithms to predict bridge element conditions, including artificial neural networks (ANNs) (Ali et al., 2021), support vector machines (SVMs) (Nguyen and Dinh, 2022), and, increasingly, ensemble methods like Random Forests and Gradient Boosted Machines (GBMs) (Assaad & El-adaway, 2020).

GBMs, and particularly their efficient implementation in XGBoost (Chen and Guestrin, 2016), have consistently demonstrated state-of-the-art performance on tabular datasets like the NBI. Their ability to handle missing data, reduce overfitting through regularization, and capture intricate feature interactions makes them exceptionally well-suited for this problem domain (Tobias and Máté, 2020). While these studies have proven ML’s predictive superiority, many stop at reporting classification metrics (e.g., accuracy, AUC), failing to translate these statistical measures into actionable economic policy.

2.3 Explainable AI (XAI) and decision-theoretic frameworks

The adoption of complex ML models in high-stakes public domains like infrastructure management is often hampered by their “black box” nature (Ribeiro et al., 2016). For an engineer or policymaker to trust a model’s recommendation to spend millions of dollars on a bridge intervention, they must understand why the model made that prediction. This has given rise to the field of Explainable AI (XAI). SHAP Lundberg and Lee (2017) has emerged as a leading XAI technique due to its solid foundation in cooperative game theory and its ability to provide both global explanations (which features are most important overall) and local explanations (why a specific prediction was made for a single bridge). The use of XAI is critical for model validation, debugging, and building stakeholder trust necessary for real-world implementation (Molnar, 2020).

Parallel to the rise of ML, operations research has a long history of developing decision-theoretic models for maintenance and replacement (Jardine and Tsang, 2013). These models formalize the trade-off between the certain, lower cost of preventative action and the uncertain, higher cost of failure. Recent work has sought to integrate probabilistic failure models into these frameworks for bridges (Orcesi and Frangopol, 2011; Sobanjo and Thompson, 2012), but often relies on simpler statistical methods for the probability estimates.

A critical body of work in bridge management has long recognized the limitations of traditional Markovian deterioration models and sought to overcome them using more sophisticated, non-Markovian approaches. For example, Madanat et al. (1997) introduced transition probability estimation techniques that explicitly relax the memoryless property, while Orcesi and Frangopol (2011) embedded performance-based deterioration models into life-cycle cost optimization frameworks. More recent work has applied hazard models and condition-based Markov decision processes (Bektas and Smilowitz, 2013) to accommodate historical deterioration paths. This framework extends this line of research by (a) learning complex transition probabilities directly from longitudinal data using a state-of-the-art machine learning model (XGBoost), (b) bypassing the memoryless limitation entirely, and (c) integrating cost optimization directly into threshold-based intervention policies. Unlike prior models that derive deterioration transitions from pre-specified statistical formulations, our XGBoost-based approach learns these dynamics directly from observed data, capturing non-linear and time-dependent effects. This allows a unified decision rule that directly balances predictive probabilities and economic trade-offs within a transparent, data-driven framework. Recent studies have similarly explored coupling ML-based deterioration models with life-cycle cost analysis, underscoring the growing importance of this integrative approach. While previous studies often rely on simplified or statistical deterioration models, our approach fuses predictive precision with economic decision logic, offering a scalable alternative. A detailed comparison of these methodological distinctions is presented in Appendix A.

The significant research gap, which this paper directly addresses, lies in the complete synthesis of these fields: a cohesive framework that uses state-of-the-art, explainable machine learning predictions to drive an economically optimal decision policy on a national scale.

3 Methodology

Our methodology is designed as a replicable, four-stage process, moving from raw data to an economically optimized decision strategy.

3.1 Data acquisition and longitudinal cohort construction

The study utilizes the National Bridge Inventory (NBI) datasets for the years 2020 through 2024, as provided by the U.S. Federal Highway Administration (FHWA). After combining these annual files, a unique identifier (BRIDGE_ID) was created for each bridge by concatenating its state code and structure number. Key features, including YEAR_BUILT_027, AVERAGE_DAILY_TRAFFIC_029, and condition ratings, were converted to numeric formats, with non-convertible entries treated as missing data. A bridge’s AGE was calculated from its construction or reconstruction year relative to the inspection year. Predictor observations from 2020 to 2023 were paired with outcome labels from 2021 to 2024 using a 1-year forward shift within each BRIDGE_ID. Observations from 2024 were excluded as predictors due to the absence of a subsequent inspection year within the dataset window.

To ensure the generalizability of the model and mitigate the risk of data leakage, the dataset was partitioned using a Group-wise shuffle split based on unique BRIDGE_ID. This ensures that all historical observations for a single structure are isolated within either the training or test set, preventing the model from “memorizing” structure-specific deterioration patterns. Furthermore, a selection bias analysis was conducted to account for the reduction of the modeling cohort (see Table 1). Comparison of physical descriptors (e.g., age) between excluded and included bridges suggested no systemic survival bias within the final longitudinal cohort.

3.2 Problem formulation and feature engineering

The research problem was formulated as a longitudinal binary classification task. The target variable, IS_AT_RISK_NEXT_YEAR, was derived for each observation at year t based on its condition at year t + 1. A bridge was labeled as 1 (positive class) if its STRUCTURAL_EVAL_067 rating in the following year was 4 or below, corresponding to the NBI definitions of “Poor” or “Failing.” All other bridges were assigned a target value of 0.

The final modeling dataset was created by dropping any rows containing missing values for the target variable or any of the selected features. This resulted in a high-quality but focused dataset of 1,559 bridge-year observations suitable for prognostic modeling.

3.3 Predictive model: XGBoost for probability prediction

We selected XGBoost (eXtreme Gradient Boosting) as our predictive engine, implemented via the Python library of the same name (Chen and Guestrin, 2016). XGBoost is an ensemble of decision trees, which it builds sequentially, with each new tree correcting the errors of the previous ones. This boosting mechanism makes it a powerful and often state-of-the-art algorithm for structured, tabular data (Shwartz-Ziv and Armon, 2022).

The model was configured to predict the probability of a bridge belonging to the IS_AT_RISK_NEXT_YEAR = 1 class. A critical parameter for this type of problem is scale_pos_weight. Given the significant class imbalance in the data (far fewer bridges become at-risk in any given year than remain healthy), this parameter was set to the ratio of negative-to-positive class instances. This effectively increases the penalty for misclassifying the rare but critical positive class, forcing the model to pay closer attention to it. To enhance robustness, the model was tuned with standard regularization parameters (max_depth = 5, subsample = 0.8, colsample_bytree = 0.8, and a learning_rate = 0.1).

Recognizing that gradient-boosted models often produce uncalibrated probabilities, Isotonic Regression was applied as a post-processing step. This calibration aligns the model’s predicted risk probabilities with actual observed failure frequencies, which is a prerequisite for ensuring that the intervention thresholds are mathematically consistent with economic cost-ratios (k).

3.4 Cost-driven decision framework

An intervention is economically rational only if the expected cost of failure is greater than the certain cost of maintenance. This yields the following decision rule: Intervene if P(Failure) × C_f > C_m, which simplifies to intervening when the model’s predicted probability exceeds the maintenance-to-failure cost ratio. We simulate the total net savings of this strategy across a range of decision thresholds on the test set to identify the optimal threshold that maximizes economic benefit.

4 Results

The performance of the XGBoost classifier was evaluated on a held-out test set comprising 30% of the available data. The model demonstrated strong predictive power, and the detailed performance metrics are presented in Table 2.

As Table 2 shows, the model achieves a precision of 0.34 and recall of 0.67 for the minority “At Risk” class. This indicates that while the model correctly identifies roughly two-thirds of bridges that will deteriorate (recall = 0.67), a precision of 0.34 means that only about one-third of the bridges flagged as “At Risk” truly were at risk, demonstrating a substantial number of false positives—a common result of severe class imbalance (450 “Not at Risk” vs. 18 “At Risk” samples). This asymmetry directly influences the economic optimization: low precision inflates unnecessary maintenance costs, reducing net savings. Initial cost simulations identified an optimal threshold of 0.82; however, diagnostic evaluation via a Reliability Diagram revealed that the underlying XGBoost model was significantly uncalibrated in the lower probability ranges (Figure 3). Following the application of Isotonic Regression to align predicted risk with actual observed failure frequencies, the optimal intervention threshold shifted to 0.01. This calibrated threshold is mathematically consistent with the theoretical decision-theoretic rule (P > k), where k = 0.1 represents the maintenance-to-replacement cost ratio (Figure 4). This result confirms that proactive maintenance is economically rational at low probability levels when the relative cost of failure is significantly higher than the cost of intervention. To understand the drivers behind these predictions, we analyzed the model using SHAP. Figure 1 illustrates the global feature importances, ranking each feature by its mean absolute impact on the model’s predictions across the test set. Deck condition emerged as the most influential factor, followed by traffic volume and age.

FIGURE 1

The central result of our study is the cost-benefit analysis. By simulating the economic outcomes of applying different decision thresholds to the test set, we identified the optimal intervention strategy. The results of this simulation, which plots the total net savings as a function of the decision threshold, are presented in Figure 2.

FIGURE 2

FIGURE 3

5 Discussion

This study successfully developed and validated an end-to-end framework for translating predictive insights into economically optimized maintenance decisions. The results, while modest, provide a realistic and valuable perspective on the application of machine learning to real-world infrastructure data, highlighting both the potential for economic benefit and the critical challenges that must be addressed.

5.1 Interpretation of key findings

The primary finding of this research is that a proactive, data-driven strategy can yield a positive financial return. The achievement of a non-zero net savings on the test set validates the core hypothesis that this framework can create economic value compared to a purely reactive approach. However, the small magnitude of this savings and the high optimal decision threshold of 0.82 are equally important findings. These results are a direct and quantifiable reflection of the model’s performance trade-offs, as detailed in Table 2. The model’s predictive capabilities, particularly its precision-recall balance for the minority class, directly dictate the economic outcome. The economic simulation correctly concluded that to remain profitable, the intervention strategy must be highly conservative, only acting when the model is overwhelmingly confident. This minimizes the cost of unnecessary maintenance on false positives, but consequently limits the total number of interventions and the overall potential savings. This reveals a critical insight: the decision framework acts as an “economic amplifier” for model quality, where gains in predictive precision would directly translate into more aggressive, and thus more profitable, maintenance policies.

The SHAP analysis, summarized in Figure 1, provides crucial transparency into the model’s decision-making process. The model learned a policy that aligns with established engineering principles, identifying deck condition (DECK_COND_058) as the most significant predictor of next-year risk. This is logical, as the deck is directly exposed to traffic and environmental stressors. Average daily traffic and age were ranked as the second and third most important factors, respectively. The model’s ability to learn these intuitive relationships demonstrates that it is not a “black box,” but a tool that has successfully captured domain-relevant patterns, increasing its credibility for use by transportation agencies.

5.2 Implications for policy and practice

The implications of this framework for transportation agencies are significant. It provides a quantitative and objective methodology for resource allocation. Rather than relying solely on age-based or time-based schedules, an agency can use the probabilistic output of the model to rank its entire bridge inventory by risk, creating a “triage list” for inspections and maintenance planning. The optimal threshold provides a data-driven cutoff point for action, allowing for the justification of budget expenditures. Perhaps most importantly, the framework itself is a tool for self-evaluation. An agency can use it to test the economic value of different predictive models or to quantify the return on investment of improving their data collection practices. For instance, they could measure the increase in net savings generated by a model trained with more granular climate or material data, thereby justifying the cost of acquiring that data.

5.3 Limitations and directions for future research

This study, while providing a complete framework, has several limitations that offer clear directions for future research. The most significant limitation is the data sparsity inherent in creating a longitudinal cohort from the public NBI dataset. The requirement for consecutive-year inspection records to create a longitudinal target variable significantly reduced the size of the available dataset. This data sparsity may limit the model’s ability to learn more complex deterioration patterns and is a likely contributor to its modest precision. Future work could explore advanced techniques for handling sparse time-series data or seek access to more complete, non-public state-level datasets.

Second, while the features used—age, deck condition, and traffic volume—are strong proxies for overall deterioration, they do not fully represent the temporal evolution of degradation. Future work could introduce lagged condition features (e.g., values at t–1 or t–2) to capture deterioration velocity and better encode historical patterns, though this remains constrained by data completeness.

Third, the model presently produces point-estimate probabilities without explicit uncertainty quantification. Future research could apply bootstrapping or Bayesian gradient-boosting variants to quantify predictive confidence intervals and assess how uncertainty propagates into economic outcomes.

The cost model, while effective, is a simplification…

The omission of indirect societal costs (e.g., traffic detours, lost productivity, environmental externalities) likely underestimates the true cost of failure. Including such externalities would increase the relative value of proactive maintenance and thus lower the economically optimal intervention threshold.

Furthermore, the assumed 10 percent maintenance-to-failure cost ratio is heuristic. A brief sensitivity consideration suggests that lower ratios (cheaper maintenance) would shift the optimal threshold downward, while higher ratios would push it upward, demanding greater predictive confidence before intervention.

Finally, the current framework operates at the single-period, single-asset level. Real-world maintenance policies are multi-year and network-oriented, constrained by budgets, crew logistics, and corridor interdependencies. A natural extension of this research is to integrate the risk probabilities generated here into multi-asset, multi-period optimization models that maximize total network reliability under resource limits.

5.4 Economic sensitivity analysis

A sensitivity analysis was performed to evaluate the framework’s scalability across varying economic landscapes. By varying the maintenance-to-failure cost ratio (k) from 0.02 to 0.50, we demonstrated that the framework acts as an economic amplifier. At lower k values (e.g., k = 0.02), the optimal strategy is highly aggressive, while higher k values (e.g., k = 0.50) demand conservative thresholds (e.g., 0.98) to remain profitable (see Figure 5). This ensures the framework is adaptable to different agency budget constraints.

FIGURE 4

FIGURE 5

6 Conclusion

This research presented and validated a comprehensive, cost-driven machine learning framework for the optimal prioritization of bridge maintenance. By systematically integrating a high-performance predictive model, a transparent explainability module, and a rigorous economic decision rule, we have demonstrated a methodology that moves beyond simple condition forecasting to provide actionable, economically-grounded policy recommendations. The results indicate that such an approach can lead to positive cost savings, but that the scale of these savings is fundamentally tethered to the precision of the underlying predictive model. The principal contribution of this work is not the specific monetary figure, but the framework itself: a robust and scalable methodology that allows any transportation agency to evaluate the real-world economic utility of its predictive models and data assets. This study provides a practical blueprint for shifting from traditional reactive maintenance to a proactive, data-driven, and economically optimized paradigm for managing the nation’s critical infrastructure, while also highlighting the critical importance of high-quality, longitudinal data in realizing the full potential of machine learning in this domain.

References

• 1

  AliI.El-adawayI. A.AssaadR. (2021). A novel approach to enhance bridge service life prediction using recurrent neural networks. J. Infrastructure Syst.27 (4), 04021035.

  • Google Scholar
  • View reference in article

• 2

  American Society of Civil Engineers (ASCE) (2021). 2021 report card for america's infrastructure. Available online at: https://infrastructurereportcard.org/.

  • Google Scholar
  • View reference in article

• 3

  AssaadR.El-adawayI. A. (2020). Bridge infrastructure asset management system: Comparative computational machine learning approach for evaluating and predicting deck deterioration conditions. J. Infrastruct. Syst.26 (3), 04020032. 10.1061/(ASCE)IS.1943-555X.0000572

  • CrossRef
  • Google Scholar
  • View reference in article

• 4

  BektasB. A.anM. C. (2012). A markov decision process-based model for bridge maintenance and replacement. Transp. Res. Rec.2292 (1), 20–28.

  • Google Scholar
  • View reference in article

• 5

  BektasB. A.SmilowitzK. (2013). Modeling infrastructure deterioration and maintenance decisions using semi-markov decision processes. J. Infrastructure Syst.19 (1), 45–52.

  • Google Scholar
  • View reference in article

• 6

  CatbasF. N.SusoyM.FrangopolD. M. (2008). Structural health monitoring and reliability estimation: Long span truss bridge application with environmental monitoring data. Engineering Structures30 (9), 2347–2359. 10.1016/j.engstruct.2008.01.013

  • CrossRef
  • Google Scholar
  • View reference in article

• 7

  ChenT.GuestrinC. (2016). “XGBoost: a scalable tree boosting system,” in Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining (New York, NY: ACM), 785–794. 10.1145/2939672.2939785

  • CrossRef
  • Google Scholar
  • View reference in article

• 8

  CiveraM.SibilleL.Zanotti FragonaraL.CeravoloR. (2023). A DBSCAN-based automated operational modal analysis algorithm for bridge monitoring. Measurement208, 112451. 10.1016/j.measurement.2023.112451

  • CrossRef
  • Google Scholar
  • View reference in article

• 9

  FarrarC. R.WordenK. (2012). Structural health monitoring: a machine learning perspective. John Wiley & Sons.

  • Google Scholar
  • View reference in article

• 10

  FlintM.GoT.ShapoB. (2020). Machine learning for bridge condition assessment. Automation Constr.119, 103348.

  • Google Scholar
  • View reference in article

• 11

  JardineA. K.TsangA. H. (2013). Maintenance, replacement, and reliability: theory and applications. Boca Raton, FL: CRC Press. 10.1201/b15856

  • CrossRef
  • Google Scholar
  • View reference in article

• 12

  LundbergS. M.LeeS. I. (2017). A unified approach to interpreting model predictions. Adv. Neural Inf. Process. Syst.30, 4765–4774. 10.48550/arXiv.1705.07874

  • CrossRef
  • Google Scholar
  • View reference in article

• 13

  MadanatS.MishalaniR.Wan IbrahimW. H. (1997). Estimation of infrastructure transition probabilities from condition rating data. J. Infrastructure Syst.3 (3-4), 120–125. 10.1061/(ASCE)1076-0342(1997)3:3/4(120)

  • CrossRef
  • Google Scholar
  • View reference in article

• 14

  MartucciD.CiveraM.SuraceC. (2023). Bridge monitoring: application of the extreme function theory for damage detection on the I-40 case study. Eng. Struct.279, 115573. 10.1016/j.engstruct.2022.115573

  • CrossRef
  • Google Scholar
  • View reference in article

• 15

  MolnarC. (2020). Interpretable machine learning: a guide for making black box models explainable. Munich, Germany: Lulu.com. Available online at: https://christophm.github.io/interpretable-ml-book/.

  • Google Scholar
  • View reference in article

• 16

  NguyenT. H.DinhT. C. (2022). A hybrid machine learning model for predicting bridge deck deterioration. Structures35, 103–114.

  • Google Scholar
  • View reference in article

• 17

  OrcesiA. D.FrangopolD. M. (2011). Optimization of bridge maintenance strategies based on structural performance and life-cycle cost. Eng. Struct.33 (4), 1239–1247. 10.1016/j.engstruct.2010.12.031

  • CrossRef
  • Google Scholar
  • View reference in article

• 18

  RibeiroM. T.SinghS.GuestrinC. (2016). “Why should I trust you?: Explaining the predictions of any classifier,” in Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, 1135–1144.

  • Google Scholar
  • View reference in article

• 19

  Shwartz-ZivR.ArmonA. (2022). Tabular data: deep learning is not all you need. Inf. Fusion81, 84–90. 10.1016/j.inffus.2021.11.011

  • CrossRef
  • Google Scholar
  • View reference in article

• 20

  SobanjoJ. O. (2010). Markov-chain models for transportation asset management: case study of highway bridge systems. J. Infrastructure Syst.16 (2), 99–108. 10.1061/(ASCE)IS.1943-555X.0000024

  • CrossRef
  • Google Scholar
  • View reference in article

• 21

  SobanjoJ. O.ThompsonP. D. (2012). Development of a risk-based approach for bridge project selection. Transp. Res. Rec.2292 (1), 11–19. 10.3141/2292-02

  • CrossRef
  • Google Scholar
  • View reference in article

• 22

  SonyS.LaventureS.SadhuA. (2019). A literature review of next-generation smart sensing technology in structural health monitoring. Struct. Control Health Monit.26 (3), e2321. 10.1002/stc.2321

  • CrossRef
  • Google Scholar
  • View reference in article

• 23

  TobiasN.MátéK. (2020). A review of gradient boosting-based methods for predicting bridge condition. J. Civ. Eng. Manag.26 (8), 754–766.

  • Google Scholar
  • View reference in article

• 24

  WellalageN.BerndtM.MaierD. (2021). A review of deterioration models for reinforced concrete bridges. Constr. Build. Mater.280, 122473.

  • Google Scholar
  • View reference in article