Predictive Analytics for Roadway Maintenance: A Review of Current Models, Challenges, and Opportunities

With the pressing need to improve the poorly rated transportation infrastructure, asset managers leverage predictive maintenance strategies to lower the life cycle costs while maximizing or maintaining the performance of highways. Hence, the limitations of prediction models can highly impact prioritizing maintenance tasks and allocating budget. This study aims to investigate the potential of different predictive models in reaching an effective and efficient maintenance plan. This paper reviews the literature on predictive analytics for a set of highway assets. It also highlights the gaps and limitations of the current methodologies, such as subjective assumptions and simplifications applied in deterministic and probabilistic approaches. This article additionally discusses how these shortcomings impact the application and accuracy of the methods, and how advanced predictive analytics can mitigate the challenges. In this review, we discuss how advancements in technologies coupled with ever-increasing computing power are creating opportunities for a paradigm shift in predictive analytics. We also propose new research directions including the application of advanced machine learning to develop extensible and scalable prediction models and leveraging emerging sensing technologies for collecting, storing and analyzing the data. Finally, we addressed future directions of predictive analysis associated with the data-rich era that will potentially help transportation agencies to become information-rich.


Introduction
United States is globally ranked among the top two countries for its excellence in financial systems, business dynamism, and innovation capability by the global competitiveness report of world economic forum [1]. However, this report ranked the United States 11th for road quality. Maintenance, repair, and rehabilitation (MR&R) strategies, significantly influence the condition of our roadway infrastructure, which was scored D, consistent with WEF report, on 2017 infrastructure report card of American Society of Civil Engineers [2]. During past decades, the importance of transportation infrastructure maintenance has significantly grown due to its contribution to economic growth. Wellmaintained transportation systems (a) better-connect geographical locations, (b) lower transportation and transaction costs-through decreased vehicle maintenance, reduced delays, and lowered fuel consumption, and (c) enhance the safety of transportation systems [1,2].
Among major maintenance strategies-predictive, preventive, and reactive-categorized by AASHTO (2011), predictive approaches have gained much importance in recent years. It is due to their potential in enhancing several aspects of maintenance objectives to lower cost over the life span of highways, increase the highway performance, provide optimal long-term planning capability, and integrate risk management into asset maintenance planning [3,4]. In addition, advanced sensing technologies with lower price points and higher computing power coupled with emerging prediction models pave the way for accurate condition prediction of highway assets.
To this end, not only increased investment is necessary to close the $543 billion capital investment gap for highways and bridges maintenance [2], but also smart investment is critical to take the optimal advantage of allocated budgets. Several state and federal level legislations such as MAP-21 [5], and FAST Act [6] were signed into law to increase investment and provide guidelines to enhance asset management. For example, state DOTs are required by FHWA to establish a process for conducting a network level Life Cycle Planning (LCP) for the National Highway System (NHS) pavements and bridges. FHWA defines LCP as the process of cost estimation for managing assets over their life span with consideration for minimizing cost while maximizing or preserving the condition. Identification of deterioration models that support predictive maintenance is one of the required elements of the LCP process as stated by FHWA. Life cycle planning will support prioritizing the improvement of various assets with minimum possible costs to achieve maximum possible return on investments on the roadway infrastructure [7][8][9]. While the current focus of most state DOTs is on project-level life cycle cost analysis, planning towards a network-level analysis is required to increase the return on maintenance investment. That is why data-driven predictive analytics of multiple highway asset classes is critical to LCP planning.
To investigate the potential of different predictive techniques in tackling the challenge of efficient and effective MR&R planning, we reviewed prediction models in the literature for a range of highway assets. We reviewed several methodologies, from simple subjective assessments based on experts' opinions to parsimonious probabilistic and deterministic approaches. We also discussed the limitations of reviewed techniques and how these shortcomings might impact the application and accuracy of the models. Furthermore, we explored the factors influencing the accuracy of predictive analytics and identified opportunities for future studies to take advantage of advanced sensing technologies, emerging predictive analytics paradigm, and computing power.
To obtain the objective of this study, in the next section, we will introduce the structure of our methodology. Next, we will explain the key terms in asset management used in this paper. Then, the categorized literature of the prediction methods will be reviewed along with the limitations and shortcomings of the methods. Finally, we will discuss our findings, opportunities for future studies.

Key Terms in Asset Maintenance
Throughout this paper, the following terms are widely used. While a variety of definitions can be found in the literature, the authors comply with the following definitions and considerations.
Condition: The generally accepted use of the term "condition" refers to the physical characteristics of an asset which may affect its performance. The condition is assessed using physical factors.
Performance: To provide a satisfactory service, all transportation assets should work in an acceptable range of functional requirements. The performance of an asset denotes the degree to which these requirements are met, based on the expectations of users or owners [11,12].
Condition and performance are highly correlated and they have been interchangeably used in the literature. Since the focus of this paper is on the process of prediction, we also used these two terms interchangeably.
Levels of Service: The condition of an asset or services provided by the owner, which is evaluated according to the user's perception, is usually referred to its level of service. Levels of service are utilized to identify time and type of maintenance activities. This concept incorporates condition and performance into an easy to use indicator [13]. The NCHRP report provided a framework to measure levels of service based on quantitative grades using a letter-format classification [14].
Service Life: The period over which the performance of an asset is within an acceptable range is defined as its service life [11].
Remaining Service Interval: Because of different interpretations of Remaining Service Life (RSL) among stakeholders, and uncertain use of term "life" to represent various events of construction treatments in an asset's history, another term, Remaining Service Intervals (RSI) was recently introduced. This term defines the remaining time until a specific required MR&R event. RSI focuses on two questions: when a treatment is needed and what kind of treatment should be used. This process is created by taking into consideration the Life-Cycle Cost Analysis (LCCA) and prioritization of maintenance activities, in accordance with future construction events and user's benefits [15,16].
Reactive, Preventive and Predictive Maintenance: AASHTO (2011) categorized transportation asset maintenance into: (1) reactive or unplanned and (2) proactive or planned [3]. Reactive maintenance refers to activities performed after a failure occurs in a system [17]. But proactive maintenance is a strategy that is accomplished to prevent or postpone a failure from taking place [18]. Preventive and predictive are two types of proactive maintenance. The preventive techniques refer to a series of scheduled activities to prolong the life of an asset. However, the predictive methods are based on an inspection analysis to forecast the time of failures in a system. This type of maintenance also schedules necessary maintenance actions to prevent those failures [19]. A schematic representation of maintenance strategies including reactive, preventive and predictive is shown in Figure 3. Forecasting the future condition of assets is the most important aspect of the predictive strategy. Therefore, in the next section, we reviewed prediction models for a range of highway assets.

Prediction Models
Information about time, location and extent of foreseen failures, help agencies to be prepared for accomplishing associated treatments to those failures. One of the major drawbacks of using reactive maintenance is the lack of such information. Hence, the maintenance activities of the reactive method might be longer than expected and are relatively expensive. Unexpected failures result in unexpected times and costs of repairs. Therefore, budget allocation and coordination of maintenance teams in reactive strategy are poor. In contrast, proactive maintenance prevents most of the failures before happening and enables agencies to manage maintenance actions. In addition, the proactive strategy improves the reliability of the performance of highway assets by avoiding failures.
Preventive and predictive strategies, that are two types of proactive maintenance, perform maintenance works in two different ways that influence the efficiency of the maintenance plan differently. Since a preventive procedure implements a series of regular periodic actions with specified time intervals, sometimes it results in unnecessary maintenance works. But predictive procedures propose maintenance actions based on the forecasted condition of assets. Therefore, decision-making is relied on maintenance needs instead of predefined times of periodic actions. Consequently, in a predictive procedure the budget allocated to maintenance works is spent more efficiently.
Agencies rely on prediction models to coordinate maintenance teams, allocate required budget, assign resources, and prioritize and schedule maintenance works. Predictive analytics provide valuable information about location, time and type of failures that are likely to happen. Based on the outcomes of the utilized models, we investigated the selected papers in two groups: deterioration and life expectancy. In the next section, a summary of the investigation is provided.

Deterioration and Life Expectancy Models
Models that forecast the future condition, performance or level of service of an asset item are known as deterioration models. In contrast, predictive modeling that primarily deals with forecasting the time of failures, time of required maintenances or time between treatments are referred to as life expectancy models.
Methodologies that are used to forecast the life expectancy are classified in condition-based and interval-based classes. In a condition-based technique, the timings of necessary maintenance interventions are evaluated by considering trigger levels. Trigger levels are adequate values of condition, under which the asset is unable to provide its required service quality, and maintenance activities should be carried out. Interval-based techniques use times of previous maintenance interventions to evaluate the time of the required maintenance, without direct consideration of the condition value. Incorporating condition-based and interval-based methods, which is called condition-interval hybrid, increases the potential of accurate outcomes [20].
Based on the algorithms used in developing a prediction model, deterioration and life expectancy groups can be classified into deterministic, probabilistic and machine learning classes. Deterministic models are in the form of mathematical functions. Probabilistic models use probability distributions to forecast a range of possible conditions or the probability of a particular condition in the future [21][22][23]. However, in a machine learning model, Artificial Intelligence (AI) is leveraged to find relationships between contributing factors and the condition/life expectancy of assets, based on the learning process from historically recorded data. Figure 4 unveils the classification of reviewed prediction models.

Deterministic Models
Several parameters affect the condition of an asset. For example, pavement roughness, an index for the assessment of the pavement condition, is affected by initial roughness, age of pavement, climatic variables, structural characteristics of pavement, traffic load, subgrade specification, drainage type, drainage condition, and all historical treatment and maintenance activities [24][25][26][27][28][29][30][31]. Deterministic models introduce a relationship between the condition and factors that contribute to the deterioration. These models are vastly used by transportation agencies, because they are easy to understand and simple to use.
Reviewing the studies that developed deterministic models, revealed that they have focused only on a few factors that contribute to the degradation of assets. This has been due to the limitation of available data and insufficient knowledge about all factors that contribute to the deterioration. Consequently, considering a few numbers of contributors might limit the applicability, scalability, and extensibility of the models. Furthermore, in deterministic techniques, expert-based subjective assessments play a major role in the recognition of patterns in the data, which might impact the accuracy of the outcomes. Hence, to increase the reliability of the models, several studies reported the use of a combination of subjective engineering evaluation and statistical testing to find the relationships [32][33][34]. However, the subjectivity in making assumptions and simplifications are challenges of deterministic techniques.

Probabilistic Models
These models forecast the probability of the future condition/life expectancy of an asset or a range of possible condition/life expectancy by using the probability concept [35,36]. These models represent a better view of risks. Therefore, they may help asset managers to reduce the risks of their decision makings [23,37]. In this paper, reviewed probabilistic approaches are classified into Markov chain and distribution models.
Markov chain algorithms use the stochastic concept of the Markov process, in which the sequence of possible events is specified based on the probability of each event. This probability only depends on the state of the previous event. Transitions of the condition of an asset item, from one inspection time to the next one, create a transition matrix which is used to predict future condition values [38]. Markov chain is widely used to predict condition values, especially when there is not enough information on all contributing factors. The main focus of this model is on transition probabilities, and the cause of this transition, instead of factors that lead to the condition degradation. The analysis is merely based on the historical operation and maintenance data of roadways. A review of the literature on Markov chain approaches revealed they can be categorized into two types: homogenous and non-homogenous. Homogenous models assume that the condition of an asset item at a specified time is only related to its previous condition. Moreover, these models assume that transition probabilities are constant over time. The inaccuracy of outcomes may be resulted from such assumptions. Non-homogenous models use different transition probabilities at different times. Thus, these models consider all previous stages to predict the future condition of an asset. However, some studies reported inaccuracies of the results of non-homogenous models [39].
Distribution models use a specified probability distribution for the prediction purposes. The future condition and its probability are affected by a variety of factors, which remain unknown when available data are insufficient. Hence, using simplifications and selecting a probability distribution to specify condition probabilities during the lifetime of an asset help to forecast its future condition. However, the utilized probability distribution might not completely fit the actual observations and consequently leads to inaccuracies. Commonly probability distribution models that are addressed in the literature, are Weibull, Markov-Weibull, Cox, Kaplan-Meier, ordered-probit and Bayesian.
In a Weibull-based analysis, the survival time distribution of an asset item is assumed to follow the Weibull probability distribution. In the Markov-Weibull procedure, the Markov transition matrix and the Weibull survival distribution are incorporated to predict condition possibilities in the future. In this model, the Weibull distribution is used to find an age-based probability of failures. Cox is a well-recognized proportional hazard model. In the Kaplan-Meier model, the survival function is estimated from recorded data. In Ordered-probit, the probability concept is used to forecast the condition of an asset in discrete ordinal levels. Finally, Bayesian analysis incorporates existing knowledge of condition measurements with the information from historical observations to provide a probabilistic prediction model [22,23,40].

Machine Learning Models
Arthur Samuel introduced the machine learning concept in 1959. The learning process of a computer when it is not explicitly programmed is referred to as machine learning [41]. In recent years, the use of machine learning has attracted the attention of researchers in the prediction of maintenance activities [42]. A machine learning technique investigates deep inter/intra-correlations and patterns in a dataset with minimal human intervention. By learning from data, instead of using subjective assumptions and simplifications, machine learning improves predictive analytics.
Artificial Neural Network (ANN) is one of the most commonly used machine learning algorithms in highway asset management. Time-consuming training process and instability in local minimum points during training were addressed in the literature as shortcomings of the ANN. Additionally, selecting the optimal architecture of neural nets and choosing proper training algorithms have been addressed as challenging issues in an ANN model [22].
A summary of the reviewed prediction approaches in the literature for a range of transportation assets is provided in the next sections.

Pavement Prediction Models
Because of the huge length of roadways, the budget allocated to maintenance activities on pavements is considerably high. Furthermore, the condition of pavements influences the traffic and safety. Consequently, pavement is considered as a high capital asset in the highway asset management. We reviewed prediction models for pavements under two categories: deterioration and life expectancy models.

Pavement Deterioration Models
In this section, pavement deterioration models are reviewed in three main groups: deterministic, probabilistic and machine learning.

Deterministic Models
Since these models are easy to understand and simple to use, they are widely used by transportation agencies. Deterministic models for pavements are classified into three groups: mechanistic, empirical and mechanisticempirical.
In a mechanistic approach, the process of condition prediction is accomplished based on the mechanical and structural characteristics of pavements [21]. Mechanistic models define the condition of assets as a function of mechanistic responses, including stresses, strains, and deflections. For example, the load of vehicle wheels is utilized to estimate stresses and strains in a pavement layer. Then, these estimations are utilized in mechanistic functions to quantify condition values [35,43]. Hence, these models provide a deep understanding of the relationships between mechanical responses and the condition. However, the accuracy of results is questionable due to the complexity of considering all parameters that potentially influence the condition [22].
Empirical methods deploy statistical analysis and consider main factors that contribute to the deterioration, such as traffic load and pavement age [21]. Literature review unveils widely use of empirical procedures to predict the future condition of pavements. However, shortages in the available data made these models be valid only in limited situations. Furthermore, a key problem with most of the empirical methods is that they consider only a few determinant factors in the degradation of assets due to the insufficient available data or the limitations of the utilized analysis algorithms. The complexity of equations that define the relationship between contributing factors and the condition is addressed as another challenge [25].
Linear and non-linear regression is commonly used in empirical methods. In contrast with linear regression, nonlinear approaches have a better estimation because of their capability of testing multiple arrangements of condition curves [32]. When regression analysis is conducted with one variable, it is called the univariable regression method. If multiple variables are used, it is known as a multivariable regression technique. Due to a variety of factors that might impact the degradation of assets, univariable regression methods are usually unable to provide accurate outcomes.
Mechanistic-empirical procedures use mechanistic functions and historical observations to come up with a prediction model. These models determine functional forms and descriptive variables by leveraging mechanistic methods [21,22]. These models are often applied at the project level and rarely at the network level. However, in contrast with empirical methods, these models can be used in more situations [44]. Nevertheless, calibration is a vital process to make the models be valid in various situations [21]. In addition, the main challenge of mechanisticempirical models is the need for more structural data that are usually unavailable. Table 1 shows a summary of reviewed papers that developed deterministic models for the prediction of the pavement condition.  Table 2 reveals a summary of reviewed papers associated with probabilistic models for the prediction of the pavement condition. The review revealed that the Markov chain is the most commonly used probabilistic method to forecast the performance of pavements.

Machine Learning Models
A review of the literature on machine learning prediction models for pavements unveiled that most studies used Artificial Neural Network (ANN) technique. Table 3 provides a summary of the reviewed papers.

Pavement Life Expectancy Models
A variety of factors affect the life of an asset: Pavement material, traffic, age, climatic parameters, subgrade specifications, maintenance, and operation history are some of these factors that are addressed for pavements [20]. All contributing factors to the deterioration of an asset influence the length of its life or remaining service intervals. Life expectancy models consider these factors to predict time-based terms, such as remaining time to failure or remaining time to the next maintenance need. These terms are usually achievable by using deterioration models. But, in this section, we reviewed life expectancy models separately because they specifically focus on the quantification of timebased terms. We reviewed life expectancy models in three groups of deterministic, probabilistic and machine learning.

Deterministic Models
The reviewed papers that leveraged deterministic approaches for the prediction of life expectancy of pavements, utilized linear and non-linear regression techniques for explaining the relationship between variables. Table 4 provides a summary of these papers.

Probabilistic Models
Probabilistic techniques are used to forecast the life expectancy of pavements. Table 5 shows a summary of the reviewed studies that developed probabilistic models. Enabling highway agencies to determine when a particular pavement needs to be smoothed to obtain a given (desired) life extension Labi [94] 2001 General Investigating impacts of maintenance treatments in short-term and entire life of pavements, and examining the trade-off relationships between rehabilitation intervals and maintenance, traffic loading and weather to improve pavement management system

Machine Learning Models
Some studies were carried out to implement machine learning to forecast life expectancy of pavements. ANN models were commonly used for this purpose. Table 6 reveals a summary of the reviewed papers that leveraged machine learning to predict the life expectancy of pavements.

Markings Prediction Models
Pavement markings are generally used to convey messages to drivers. Markings provide helpful information such as which part of the road should be used, what conditions are ahead or where passing is allowed. There are various types of markings associated with their role. The expected roles of markings in a transportation system indicate their color, pattern, shape, and location. As markings are usually in contact with corrosive and destructive phenomena, they deteriorate over time. Some of the causal factors of deteriorations are weather and climatic parameters, debris and vehicle transit. In this paper, the review of markings prediction models was conducted in two groups: deterioration and life expectancy.

Markings Deterioration Models
Retroreflectivity is a general definition used to evaluate the performance of markings. Retroreflectivity is the capability of materials to reflect light back to its source [89]. For markings, the retroreflectivity reveals the ability of which for reflecting light from vehicle headlights to the eye of drivers [40]. To predict the future condition of markings, most studies focused on the retroreflectivity forecasts. Deterministic methods are the most common procedure used in the reviewed studies. However, there were a few studies that developed machine learning models. A summary of the literature on the prediction modeling for the deterioration of markings is given in Table 7.

Markings Life Expectancy Models
Most research studies on the life expectancy of markings used deterministic approaches. A summary of the review of life expectancy models for markings is outlined in Table 8.

Signs Prediction Models
In this paper, signs prediction models were reviewed in deterioration and life expectancy classes.

Signs Deterioration Models
To measure the condition of a sign, retroreflectivity is the most common characteristic which is identified as a condition indicator [130]. The retroreflectivity identifies how a sign is distinguishable during day and night, and the measurement of this parameter has a main role in preparing maintenance plans. A review of the literature showed various contributing factors to the retroreflectivity degradation. Most of the previous studies used deterministic procedures to find a relationship between determinant parameters and the condition of a sign. There are a few research studies on other types of prediction models. For example, Swargam et al. (2004) utilized an ANN method to predict the sign retroreflectivity based on the Louisiana DOT database and compared its results with linear regression models [131]. The reviewed papers that used deterministic models for signs are addressed in Table 9.

Signs Life Expectancy Models
The review of previous studies showed that models developed for the prediction of life expectancies of signs are mainly deterministic. A summary of the reviewed research studies is shown in Table 10.

Barriers and Culverts Prediction Models
To gain better perspective on prediction models that are used in various asset items we reviewed some of the studies on barriers and culverts. Whether deterioration or life expectancy models are considered, most studies utilized deterministic and probabilistic approaches for their prediction. A summary of the reviewed studies for culverts and barriers is shown in Table 11.

Discussion
Prediction algorithms are one of the key elements in making optimal life cycle decisions for maintenance management. Accurate predictions will help to develop decision support systems for improving the performance of assets while decreasing their life cycle maintenance costs. The authors of this paper investigated the literature of prediction models for transportation asset management and highlighted some of their shortcomings that might lead to inaccuracy of the outcomes. The level of inaccuracies is different across asset classes, and prediction methods. In addition, considering the amount of data, which has recently become available through new sensing technologies, the accuracy of some of these models still seems unsatisfactory. In this section, some of the important limitations and challenges of the reviewed prediction methods are discussed.
 In deterministic and probabilistic methods, some simplifications are made to define and formalize relationships between variables. In addition, the process for recognition of patterns is based on subjective assumptions and simplifications. For example, in the Weibull distribution models, it is assumed that the distribution of survival time for an asset follows Weibull probability distribution, which may not completely fit the actual historical observations. Therefore, these assumptions may lead to meaningful discrepancies between the forecasted and observed values.
 Since the 90s, ANN is a widely used machine learning algorithm in the prediction problems including condition or life expectancy of highway assets. In traditional ANNs, not only selection of training algorithms, and optimal architecture of the model are challenging; But also, time-consuming training process and instability of the model in local minimum points limit the performance of the models [22,144]  Most prediction models were built upon limited historical data. Costly data collection procedures in the past and limited available sensing technology to facilitate extensive data collection are some of the reasons for inadequate historical data. The robustness of data processing and developing a prediction model highly depends on the quantity and accuracy of the collected data. Sometimes, the inadequacy of the data is a major challenge. Consequently, most prediction models are not scalable and extensible, and the inaccuracy of outcomes is likely.
 Addressed predictive procedures considered limited factors that have a probable contribution to the deterioration. It was due to the limited available data and/or challenges of big-data analysis at the time of the previous studies.
 The condition of asset items might be affected by the condition of other assets. For example, pipe erosion of underway culverts has the potential of impacting the pavement condition. Most prediction models focused only on one asset item or analyzed the behavior of assets independently. Therefore, the interrelation between neighboring asset items and their mutual impacts has not been the focus of most studies.
 Advancements in material science and the utilization of new materials in the construction and maintenance of transportation assets will impact operation and maintenance decisions. For example, using lightweight structural materials, including foamed structures, magnesium-based components, and advanced aluminum alloys [145], self-healing concrete [146], nano-materials in concrete [147], and nano-modified asphalts [148].
Utilizing new materials in construction or production of highway assets results in a change of characteristics and the enhancement of their properties. There are limited studies on the deterioration pattern of these new materials to be used in prediction models.
 Most studies focus on primary roads and interstate highways, and there are a few studies on secondary and rural roadways. Hence, the accuracy of the prediction in these roadways is still a challenge.

Conclusions
Different deterministic, probabilistic, and machine learning models have been developed for pavements, however, developing prediction models for other assets has often been based on deterministic or probabilistic methods. The current review has raised several shortcomings of existing models that calls for further investigation. The critical issues and challenges, discussed in the previous section, suggest the following opportunities for the future research studies.
 By improving data modeling, state-of-the-art machine learning and deep learning approaches tackle challenges of subjective assumptions and simplifications in deterministic and probabilistic models. Machine learning methods, by minimal human intervention, use fewer assumptions about inter/intra correlations in data. Additionally, they provide insights into deep correlations and patterns that might not be apparent to a human. Furthermore, advanced machine learning algorithms and training methods mitigate several shortcomings and limitations of traditional ANNs and offers the potential for beneficial developments in prediction models. Therefore, it is recommended that future studies investigate the potential use of advanced machine learning and deep learning models in condition predictions.
 To overcome the challenge of limited available data, more frequent and extensive data collection are needed. Advanced sensing technologies facilitate inspection and data collection processes. For example, collecting 3D surface data by using Stereovision and Light Detection And Ranging (LIDAR) [149], utilizing Unmanned Aerial Vehicles (UAV), and robots such as Pioneer P3-AT and Pioneer 3-DX for inspection purposes [150]. Future studies should be conducted to investigate these technologies and their impacts on the cost, safety, and accuracy of data collection to identify high impact technologies. In addition, requirements for the implementation of a data management system for storing and processing the collected data should be investigated.
 The integration of big-data analytics and advanced sensing technologies delivers an enormous amount of data for developing a robust predictive analysis. To handle the huge volume of data a platform is required for connecting, securing and analyzing the data [151]. Research studies should be conducted to identify and propose the structure and application of the platform in TAM, including collecting, storing and manipulating the data.
 Recently, specialized hardware productions make the complex computation of massive data more possible than before. Furthermore, using state-of-the-art methods of data analysis provides more opportunities for researchers to examine big-data in their studies on predictive maintenance. Therefore, new investigations are recommended to examine the potential of various factors that have probable contributions to the deterioration but are not considered in the previous studies.
 To respond to the challenge of considering the mutual impacts of multiple nearby assets, further studies are needed to examine and consider interrelationships between conditions of multiple assets in neighboring locations. Consideration of the mutual impacts helps to improve the robustness of forecasting models.
 The effects of extreme climate events lead to situations that are not considered in the design of many infrastructures. Drought, extreme flood, warm weather, changing wind speed, hurricanes and intense freezingthawing cycles are some of climate change consequences [145]. Climate change impacts roadway assets. For example, more frequent buckling might occur because of the rising temperature, unpaved shoulders might be washed-out due to increasing of intense precipitations, and intense freezing-thawing cycles lead to more cracks on pavements. Hence, MR&R plans based on the current prediction models should be reassessed to improve the resiliency of assets to climate change. More studies are needed to consider climate change, examine its impacts on the long-term performance of assets, and develop new prediction models.
 The investigation of how advancements in treatment technologies and the use of new materials impact predictive maintenance is another challenge for future studies. New prediction models are needed to consider the impacts of new materials on the performance or useful life of asset items.
 There are many studies on new methods of commuting to make greener, safer and more efficient transportation systems. For example, innovations in transportation methods and state-of-the-art Autonomous Vehicle (AV) technologies are expected to make a revolution in the transit system. On one hand, this change has significant effects on traffic patterns and consequently on the condition and performance of asset items and their useful life. Hence, future studies should deal with the effects of new methods of commuting on the prediction of the deterioration and life expectancy. On the other hand, using AVs will change the extent of needs for some of assets. For example, the width of lanes and shoulders, using median barriers and guardrails are some of the issues that will be affected by the use of AVs. This emerging paradigm of transportation leads to a new and different viewpoint in the design, construction, operation, and maintenance of the transportation system. Therefore, further studies are recommended to examine the impacts of new commuting methods on long-term conditions of various highway assets.

Conflicts of Interest
The authors declare no conflict of interest.