PRISM-AI: An Explainable Machine-Learning Platform for Predictive Monitoring of Public Infrastructure Projects

Abstract

Public infrastructure projects generate high-consequence decisions under incomplete, heterogeneous, and changing information. This review examines how machine learning, explainable artificial intelligence, project controls, and digital monitoring can be combined into a platform for predictive oversight. The literature shows a progression from static classification of delay risk toward hybrid systems that combine schedules, earned-value indicators, risk registers, BIM, sensor streams, and probabilistic simulation. Ensemble and hybrid models often perform well in retrospective prediction, but reported accuracy is not equivalent to operational value. Explainability is increasingly implemented through feature attribution, SHAP, and counterfactual scenarios, although explanations usually describe model associations rather than causal effects. The strongest platform designs therefore connect predictions to risk bands, trigger conditions, responsible actors, intervention timing, and measurable outcomes while retaining professional accountability. Evidence remains concentrated in single programmes, small or expert-derived datasets, retrospective case studies, and simulated or semi-realistic environments. External validation, temporal drift, calibration, fairness, intervention effectiveness, and sustained use by public-sector teams are underdeveloped. PRISM-AI should consequently be designed as a human-supervised monitoring and decision-support system rather than an autonomous project controller. Its research contribution would be to evaluate whether transparent, calibrated, temporally updated predictions improve decisions and delivery outcomes across multiple public infrastructure contexts.

Introduction

Public infrastructure projects are difficult monitoring environments because schedule, cost, procurement, design, contractor performance, weather, safety, and regulatory conditions interact over time. Traditional project controls remain useful, but their forecasts depend on the quality of baseline schedules, risk assumptions, and update discipline. In large infrastructure programmes, the evidence base for schedule-risk estimation has often been subjective or too thin to represent correlations among activities and risk factors [citation_1x98dp]. A platform such as PRISM-AI is therefore best understood as an augmentation of project controls: it should convert heterogeneous project evidence into forecasts, explanations, and actionable warnings without displacing the accountable project team.

The recent literature spans three related but distinct problems. First, empirical machine-learning studies predict project-level delay, cost, duration, productivity, safety, or other risk outcomes from historical project attributes and expert assessments. Second, schedule and process-control studies use task networks, earned-value measures, BIM, sensors, or digital twins to update predictions during execution. Third, explainable-AI studies attempt to make model outputs intelligible through feature attribution or counterfactual reasoning. These strands have not yet converged into a mature, externally validated public-infrastructure platform. A 2026 scoping review of 19 empirical infrastructure studies identified machine learning, Bayesian networks, Monte Carlo simulation, hidden Markov models, earned-value forecasting, fuzzy systems, and hybrid analytics, but highlighted limited validation beyond single settings, weak integration with live control systems, and poor data quality [citation_13s8vb].

This review synthesizes the evidence around six questions: what should be monitored; which predictive methods are useful; how explanations can support decisions; how dynamic data should be incorporated; what platform and governance functions are required; and how such systems should be evaluated. The emphasis is on primary empirical studies for measured outcomes and on review-level evidence for field-wide patterns. The analysis treats high predictive scores cautiously because datasets, labels, validation designs, and deployment conditions differ substantially.

Infrastructure monitoring needs and risk targets

Public value and delivery exposure

The case for predictive monitoring is not limited to improving an internal project metric. Delays and cost deviations can affect public services, contractual relationships, resource allocation, and the timing of benefits from roads, rail, hospitals, utilities, and other assets. The infrastructure schedule study by Fitzsimmons, Lu, Hong, and Brilakis describes more than half of UK construction projects as finishing later than planned and frames schedule-risk analysis as a problem with economic and stakeholder consequences [citation_1x98dl]. This public-value framing supports monitoring systems that identify emerging risks early enough for a lawful and feasible response, rather than merely explaining a final failure.

A PRISM-AI design should distinguish between operational risk and public-impact risk. Operational variables might include float consumption, procurement delay, productivity deviation, schedule performance, or design-change frequency. Public-impact variables might include expected service delay, affected communities, safety exposure, contractual exposure, and the reversibility of an intervention. The distinction matters because a high-probability delay is not automatically the highest-priority public risk. Priority should combine likelihood, consequence, time to act, and the feasibility and legitimacy of proposed mitigation.

Cost, schedule, quality, and safety outcomes

The literature treats project delivery as a multi-outcome problem rather than a single delay-classification task. Sanni-Anibire, Zin, and Olatunji organize delay mitigation around cost estimation, duration estimation, and delay-risk assessment, arguing that a useful framework should cover these domains together [citation_1x9891]. Their empirical implementation used separate models for cost, duration, and risk, illustrating that a platform may need multiple linked targets rather than one generalized score. The broader predictive-analytics review likewise reports applications to project-level cost, schedule, and risk outcomes across roads, marine works, tunnels, highways, prefabrication, and buildings [citation_11gdi1].

Quality and safety should not be treated as optional secondary outputs. The BIM-XGBoost study encoded BIM data into schedule-deviation, structural-health, and environmental-stress features and reported cross-domain interactions between safety risk and schedule or cost pressure [citation_1xfz6y]. Such findings support a monitoring ontology in which schedule and cost alerts can be conditioned on safety and quality states. They do not, however, justify assuming that a model trained for delay will transfer to safety incidents or defects without separate labels, measurement processes, and validation.

Risk factors and leading indicators

Predictive monitoring is most useful when it uses leading indicators that precede an adverse outcome. In the large UK schedule dataset, the strongest signals included previous linked-task deviation and the level of concurrent activity; the authors also generated features for float, weather, economic conditions, and schedule quality [citation_1x98em]. In an Indonesian refinery-development case, cumulative deviation, current-period deviation, and previous deviation were reported as the principal predictors of delay risk [citation_1am6q6]. These studies point toward a design that monitors changes and trajectories, not only static project descriptors.

The likely leading indicators are context-dependent. A tall-building study based on 48 expert responses identified slow decision-making, subcontractor delay, late engineering instructions, and waiting for approvals as relevant variables, with an ANN classifier achieving 93.75% accuracy in that dataset [citation_11gdit]. The result is useful for feature discovery but is not equivalent to evidence of general performance in public infrastructure. PRISM-AI should preserve the source, timestamp, provenance, and measurement definition of each signal so that users can tell whether an alert reflects an observed operational change, an expert judgement, or a historical association.

From retrospective reporting to early warning

Retrospective dashboards describe what has happened; predictive monitoring estimates what may happen next and how much time remains to respond. This distinction is central to platform design. A 2026 early-warning study used monthly earned-value and earned-schedule records from 17 building projects to classify next-month performance, applying expanding-window validation and probability calibration; the model achieved a test macro-F1 of 0.760 and detected the poor-performance class with 91% recall [citation_1wxvze]. Although the study is recent and small, its time-aware design is closer to operational monitoring than a random train-test split.

The target should therefore be defined by decision horizon. PRISM-AI might forecast next-period performance, probability of exceeding a completion threshold, expected cost deviation, or the probability that an intervention window will close. Each target requires a timestamped label and a clear action threshold. A final project-outcome label cannot by itself demonstrate that the system provides useful early warning during execution.

Predictive modeling approaches for project risk

Classical supervised models

Early construction studies often compare familiar supervised models because they are feasible on small, tabular datasets. Gondia, Siam, El-Dakhakhni, and Nassar compiled multivariate project data on time performance and delay-inducing risk sources, then compared decision-tree and naïve-Bayes classifiers using cross-validation; naïve Bayes performed better for the dataset examined [citation_12xfeb]. The study is important because it models interdependent risk sources rather than treating delay as an unstructured outcome, but cross-validation within one compiled dataset does not establish transfer across owners, contract types, or infrastructure sectors.

Sanni-Anibire and colleagues used multiple linear regression, ANN, K-nearest neighbours, SVM, and ensemble methods in a framework spanning cost, duration, and delay risk. Their reported models achieved an R² of 0.69 for duration, 0.81 for cost, and 93.75% classification accuracy for delay risk [citation_1x989y] [citation_1x98a0]. These results show that tabular models can produce useful baselines, but the cost model’s reported MAPE was 80.95%, demonstrating why a single accuracy number can conceal practically important error. PRISM-AI should retain simple baselines and report outcome-specific error, calibration, and subgroup performance alongside any more complex model.

Tree ensembles and feature selection

Tree ensembles are attractive for infrastructure data because they can handle nonlinearities, mixed variables, missingness strategies, and interactions without requiring a fully specified functional form. In the Balikpapan refinery-development study, a genetic-algorithm-optimized random forest achieved 99.34% accuracy and an AUC of 1.000 on the study data, outperforming the unoptimized random forest and SVM [citation_1am6q6]. The result suggests that optimization can improve within-dataset discrimination, but the unusually high score also makes leakage, temporal separation, class imbalance, and external validation essential audit questions.

A cost-risk stacking model trained on 458 construction projects combined random forest, gradient-boosting decision trees, and SVM, with logistic regression as a meta-learner. The study reports improved test accuracy relative to its best individual model and identifies material-price volatility, construction complexity, and design changes as influential variables through SHAP analysis [citation_1n95w3]. For PRISM-AI, ensemble models can serve as high-performing candidates, but the system should expose the training cohort, feature availability date, missing-data handling, and validation split whenever a prediction is displayed.

Hybrid probabilistic and simulation models

Hybrid approaches are particularly relevant to project controls because they can combine predictive learning with schedule logic and uncertainty simulation. Fitzsimmons and colleagues trained a model on 293,263 tasks from 302 completed UK infrastructure projects, combining topic modelling, an empirical Bayesian network, SVM, and Monte Carlo simulation. In ten held-out projects, the hybrid approach had a standard prediction error of 16.7 weeks versus 36.6 weeks for standard Monte Carlo simulation, a reported 54.4% improvement [citation_1x98es] [citation_1x98et]. This is stronger evidence than an in-sample classification score because the method was tested on projects withheld from training, although the case-study count remained small and the authors identify schedule quality and hindsight as limitations [citation_1x98fd].

The methodological lesson is not that every platform should combine every algorithm. It is that prediction should respect the structure of the decision problem. A probabilistic layer can represent uncertainty; a schedule-network layer can preserve dependencies; and a machine-learning layer can learn patterns not captured by subjective assumptions. PRISM-AI should therefore compare a transparent baseline, a discriminative model, and a probabilistic or simulation-linked model rather than assuming that a larger ensemble is automatically more useful.

Transferability and model choice

Model rankings vary with outcome, sample size, data provenance, and validation design. A review of 64 studies reports that hybrid and ensemble methods tend to perform well for cost estimation, tree ensembles for duration forecasting, and metaheuristic hybrids for delay prediction, but this is a synthesis of heterogeneous comparative studies rather than a universal benchmark [citation_10urbu]. Similarly, the 19-study infrastructure scoping review reports broad gains over conventional regression but identifies limited external validation and poor data quality as barriers to adoption [citation_13s8vb].

The platform should make model selection conditional on use. A project owner may prefer a calibrated model with stable explanations and low update latency over a marginally more accurate model that cannot be audited. A planner may need probabilistic completion ranges, whereas a procurement lead may need a ranked list of controllable risks. PRISM-AI should store model cards describing intended population, outcome, horizon, training period, known exclusions, expected data freshness, and failure modes. Model choice should be revisited when project type, geography, procurement route, or data-generating process changes.

Explainability and actionable interpretation

Transparent models and structured explanations

Explainability is not a decorative layer added after prediction. In project management, an alert is useful only if a responsible actor can assess whether the signal is credible, relevant, and actionable. The construction delay-mitigation framework uses a CRISP-DM cycle that moves from business understanding and data preparation through modelling, evaluation, and deployment [citation_1x989w]. That structure suggests a practical interpretation of explainability: the output should be tied to a defined management question and a documented intervention path.

Transparent models remain valuable even when they are less flexible than black-box learners. In the schedule-risk literature, SVM is described as more interpretable and suitable for small samples than ANN, although its nonlinear kernels still require explanation [citation_1x98dw]. A PRISM-AI implementation should begin with interpretable features and baseline models, then justify additional complexity through out-of-sample improvement, calibration, or decision benefit. Explanations should include the prediction target, forecast horizon, baseline risk, major contributing features, data freshness, and uncertainty rather than only a feature-ranking plot.

SHAP and attribution-based evidence

SHAP is increasingly used to quantify how features contribute to individual or aggregate predictions. Santos, Pereda, Ahedo, and Galán propose a SHAP layer for machine-learning models fitted to Monte Carlo project-network simulations, distinguishing prospective analysis for execution or replanning from retrospective analysis for identifying contributors to project status [citation_12oa86]. This separation is directly relevant to PRISM-AI: an explanation of why a current prediction is high is not the same as a recommendation about what intervention will reduce it.

Construction studies use SHAP to identify signals such as material supply, contractor performance, design changes, and safety-related factors [citation_1aoyi8] [citation_1xfz6y]. These attributions can support triage, but they should be labelled as model-attributed contributions. The infrastructure explainable-delay study explicitly warns that its SHAP signals are not causal effects [citation_1aoyi8]. PRISM-AI should therefore avoid wording such as “design change caused the delay” unless supported by a separate causal design. A safer interface would say that design-change frequency contributed to the model’s estimated risk under the training distribution.

Counterfactual and recourse explanations

Feature attribution answers “what influenced this prediction?” Counterfactual explanation asks “what would need to be different for the prediction to change?” Zhan, Fang, Love, and Luo applied a KPCA-DNN model and counterfactual explanations to tunnel-induced ground settlement in a Wuhan road-tunnel project. Their approach identified a risk path involving bubble-chamber pressure, cutter-head speed, and equipment inclination, and was intended to help site managers, engineers, and operators understand mitigation choices [citation_10fa2w].

Counterfactuals are promising for infrastructure decisions because managers act on controllable conditions, not merely on rankings. They are also easy to misuse. A mathematically sufficient change may be physically infeasible, contractually unavailable, unsafe, or outside the authority of the user. PRISM-AI should constrain counterfactuals with engineering limits, schedule dependencies, procurement lead times, safety rules, and accountable ownership. Each suggested change should show whether it is observed, feasible, reversible, and supported by a validated intervention model. Otherwise, the platform should present it as a scenario rather than a recommendation.

Trust, uncertainty, and professional judgement

Explainability does not remove uncertainty or accountability. In a 470-observation infrastructure-programme study, SHAP was used with a weighted RF-SVM ensemble and the resulting signals were translated into risk bands, trigger conditions, responsible actors, intervention timing, and measurable outcomes. The authors state that the framework should support rather than replace professional judgement and stakeholder accountability [citation_1aoyi8]. This is a useful governance principle for public projects.

The quality of explanations also depends on model and feature structure. A perspective study comparing SHAP and LIME warns that both methods are affected by the chosen model and by feature collinearity [citation_108ze6]. Correlated project variables can divide attribution among several features or make a ranking unstable without changing predictive performance. PRISM-AI should report explanation stability, correlated-feature groups, confidence or credible intervals where available, and disagreement between explanation methods. Users should be able to inspect the underlying observations and challenge an alert rather than treating a ranked list as an objective account of causation.

Dynamic data and predictive monitoring architecture

Schedule networks and project-control data

Schedule files are a rich but difficult source because task descriptions, dependencies, baselines, actual dates, revisions, and corrective actions evolve during execution. Fitzsimmons and colleagues converted short task descriptions into vector and topic features, combined them with concurrency, float, weather, economic, and schedule-quality variables, and linked the resulting predictions to Monte Carlo simulation [citation_1x98dz] [citation_1x98e9]. Their work demonstrates how unstructured schedule text can be integrated with network-derived features, but it also shows that data preparation and schedule quality are central modelling decisions.

Earned-value and earned-schedule metrics provide a more standardized monitoring layer. The near-term early-warning study used monthly EVM/ES records and acceleration features with chronological validation, finding that schedule indices were the main drivers of next-month performance predictions [citation_1wxvze]. PRISM-AI should preserve both levels: schedule-network features for activity dependencies and project-control indicators for regular executive monitoring. The platform should not collapse these into one score when their horizons or meanings differ.

Sensors, site monitoring, and temporal inference

Dynamic monitoring can support more timely signals than periodic manual reporting. Giretti and colleagues developed probabilistic models for real-time construction-progress estimation from monitoring networks, using dynamic Bayesian structures and hidden Markov models to propagate observations through time. Their feasibility study combined worker, equipment, material, and contextual state variables with BIM and visualization in a monitoring framework, while noting that accuracy depends on data resolution and quality [citation_1xfz8m].

A monitoring system must distinguish measurement frequency from predictive validity. High-frequency sensor data may improve situational awareness but can also create missingness, drift, false alarms, and privacy concerns. The evidence supports using temporal models when the process itself is dynamic, but it does not support assuming that a sensor stream automatically produces better managerial decisions. PRISM-AI should evaluate detection latency, false-alert burden, data completeness, and whether users can act within the forecast horizon.

BIM, digital twins, and multimodal integration

BIM can provide structured representations of assets and tasks, while sensors, UAV scans, positioning systems, weather, and administrative records add observed execution states. A digital-twin and BIM study collected UWB, environmental-sensor, and UAV-scanning data from three structural projects, used an LSTM to predict progress, and linked dynamic alarms to resource-allocation optimization [citation_1wuit6]. The study reports a closed loop from deviation detection to control, but its project count and setting limit generalization to public infrastructure portfolios.

The architecture should therefore treat BIM as a semantic backbone rather than as a guarantee of interoperability. A feature should retain its object, location, time, unit, source system, and transformation history. Cross-project transfer requires stable definitions for activities, milestones, quantities, cost codes, risk categories, and responsible organizations. A PRISM-AI data model should support versioned mappings so that a change in a schedule coding system is visible to users and evaluators.

Data quality, lineage, and update loops

Data quality is a recurring constraint across the evidence. The 2021 framework’s professional validators rated applicability, appropriateness, and practicality more highly than reliability, and attributed concerns about reliability to the credibility and documentation of historical data [citation_1x98a2]. The authors also state that a real-life deployment with a robust dataset would be needed to determine performance against existing approaches [citation_1x98a7].

A production platform should make data quality operational. Each prediction should carry a freshness timestamp, completeness status, source lineage, and a warning when key inputs are imputed or stale. Retraining should be triggered by documented drift criteria, not simply by calendar time. The system should retain previous model versions and predictions so that later users can reconstruct what information was available when an alert was issued. These functions are essential for public accountability and for learning from false positives and missed warnings.

Platform deployment, governance, and decision use

Integration with existing controls

The literature repeatedly treats integration as the divide between a promising model and a usable platform. The MLCPM systematic review proposes a four-layer architecture covering data, machine-learning processing, integration, and output, with feedback and retraining loops [citation_10urbu]. The earlier delay-mitigation study similarly describes deployment as either a standalone tool or an addition to existing project-management tools [citation_1x98a1]. These findings support embedding PRISM-AI within established scheduling, cost-control, risk-register, document-management, BIM, and reporting workflows rather than creating an isolated analytical portal.

Integration should be evaluated at the level of work practice. A technically correct alert that arrives after the procurement decision, cannot be traced to an accountable owner, or duplicates an existing control is not operationally useful. The platform should support APIs and batch imports, but also define a minimum viable monitoring cadence when live integration is unavailable. It should allow users to acknowledge, defer, contest, or close alerts with a recorded rationale and outcome.

Dashboards, alerts, and intervention pathways

An alert should be more than a probability. The productivity-monitoring study by Pourrahimian and colleagues combined expert knowledge, historical data, machine-learning forecasts, and a warning dashboard for semi-real-time decision-making. The authors report early identification of productivity issues but also note constraints related to data availability, expert-pool limitations, implementation, and long-term evaluation [citation_15qdpi].

PRISM-AI should present a layered interface. Executives may need portfolio-level trend and exposure summaries; programme managers may need risk trajectories and cross-package dependencies; package managers may need local explanations and controllable actions; and site teams may need immediate operational conditions. Every alert should show threshold, horizon, confidence or calibration information, top contributing evidence, affected work package, responsible role, recommended review date, and closure status. An alert should not imply that the model has authorized an intervention.

Public-sector governance and accountability

Public infrastructure adds obligations that are not captured by predictive accuracy. Decisions may affect public funds, access to services, safety, contractors, communities, and the distribution of risk. A scientometric review of construction-risk ML identifies insufficient attention to worker-centred and organizational factors, limited theorization of how outputs inform safety decisions, and fragmented treatment of data governance and ethics [citation_1xfz75]. These gaps are directly relevant to a public platform.

Governance should cover lawful data access, role-based permissions, procurement and vendor accountability, record retention, auditability, cybersecurity, privacy, and contestability. Model explanations should not expose confidential commercial information beyond the user’s authority. The platform should distinguish evidence from recommendation and preserve the human decision record. When a model is used to prioritize inspection or intervention, affected parties should be able to understand the basis of prioritization and challenge erroneous data.

Human-in-the-loop operation

Construction planning research emphasizes that human reasoning remains necessary even when machine learning is introduced, particularly for responsibility in action-taking and accountability in decision-making [citation_1xfz8p]. This is not an argument against automation. It is an argument for allocating automation to evidence processing, anomaly detection, forecasting, and scenario generation while retaining human authority over interpretation, trade-offs, and action.

A PRISM-AI workflow could require a qualified reviewer to confirm a high-severity alert, record whether the leading signal is credible, select or reject a feasible response, and document the observed result. The system should learn from these dispositions only under controlled governance because user actions may reflect organizational bias, resource constraints, or changing policy. Human feedback is valuable, but it is not automatically ground truth.

Evaluation and evidence quality

Metrics and decision-relevant performance

The reviewed studies use accuracy, precision, recall, F1, AUC, R², RMSE, MAPE, correlation, and simulation error. These metrics answer different questions. The refinery case reports very high classification discrimination [citation_1am6q6], whereas the Sanni-Anibire framework reports distinct regression and classification performance across cost, duration, and delay risk [citation_1x98a0]. Neither result alone establishes whether managers receive useful warnings or whether interventions improve outcomes.

PRISM-AI evaluation should report discrimination, calibration, lead time, false-alert rate, missed-event rate, and error by project type and subgroup. For continuous outcomes, absolute and percentage errors should be interpreted relative to decision thresholds. For probabilistic outcomes, calibration should be assessed because a well-calibrated 0.70 risk estimate is more useful for planning than an overconfident 0.95 score. Decision-curve or cost-sensitive analysis can connect predictions to the consequences of acting and not acting, but any assumed costs must be explicit.

Validation designs and leakage control

Random splits are inadequate when repeated observations from the same project appear in both training and test data or when future information leaks into features. Fitzsimmons and colleagues held out ten projects from a much larger infrastructure-task dataset, offering a stronger project-level test than a purely random split [citation_1x98er]. The early-warning study used expanding-window cross-validation and a chronological holdout, which better matches temporal deployment [citation_1wxvze]. These designs should be standard for PRISM-AI.

External validation should span owners, regions, contract types, delivery methods, asset classes, and economic conditions. A model trained on one large programme or one country may identify local documentation practices rather than transferable risk mechanisms. The explainable infrastructure study explicitly limits transferability because it uses one programme and partly expert-informed observations [citation_1aoyi8]. The platform should expose performance by deployment site and permit a “not enough evidence” state when the current project falls outside the validated population.

Primary-study comparison

The studies below illustrate the diversity of data, targets, validation designs, and interpretive claims. Their scores should not be ranked directly because the outcomes and testing conditions differ.

Study	Design and data	Model or platform function	Main finding and limitation	Citation
Gondia, Siam, El-Dakhakhni, and Nassar (2019)	Primary empirical study using a multivariate dataset of project time performance and delay-inducing risk sources; cross-validation was used [citation_12xfeb].	Decision-tree and naïve-Bayes classification for delay extent [citation_12xfeb].	Naïve Bayes performed better for the examined dataset; external transfer was not established [citation_12xfeb].	[citation_12xfeb]
Fitzsimmons, Lu, Hong, and Brilakis (2022)	Primary empirical study trained on 293,263 tasks from 302 completed UK infrastructure projects and tested on 10 held-out projects [citation_1x98dl] [citation_1x98er].	Topic modelling, empirical Bayesian network, SVM, and Monte Carlo schedule-risk simulation [citation_1x98dz].	Hybrid prediction had 16.7 weeks standard error versus 36.6 weeks for standard Monte Carlo in the case studies; schedule quality and hindsight limited interpretation [citation_1x98es] [citation_1x98fd].	[citation_1x98es]
Sanni-Anibire, Zin, and Olatunji (2021)	Primary framework study using 48 professional risk assessments and 35 duration/30 cost project cases; 10 professionals assessed deployment potential [citation_1x989m] [citation_1x989s].	Regression, ANN, KNN, SVM, and ensemble models within an adapted CRISP-DM framework [citation_1x9891] [citation_1x989n].	Reported R² values were 0.69 for duration and 0.81 for cost, with 93.75% delay-risk accuracy; experts rated reliability lower than applicability and practicality [citation_1x98a2].	[citation_1x98a2]
Sanni-Anibire, Zin, and Olatunji (2020)	Primary empirical tall-building study using 48 usable subject-matter-expert responses to construct the dataset [citation_11gdit].	KNN, ANN, SVM, and ensemble models for delay-risk assessment [citation_11gdit].	ANN achieved 93.75% classification accuracy and highlighted decision-making, subcontractors, instructions, and approvals; the expert-derived sample limits generalization [citation_11gdit].	[citation_11gdit]
Tasa, Golabchi, and Ravanshadnia (2026)	Primary case-grounded study using 470 programme-level risk-state observations from one infrastructure programme, cross-checked with documents, interviews, and schedules [citation_1aoyi8].	Weighted RF-SVM ensemble, genetic-algorithm optimization, SHAP explanations, risk bands, and trigger-based decision support [citation_1aoyi8].	Accuracy was 92.2%; material/equipment supply, contractor performance, and design changes were salient model signals, explicitly not causal effects [citation_1aoyi8].	[citation_1aoyi8]
Zhan, Fang, Love, and Luo (2024)	Primary case study of tunnel-induced ground settlement in the San-yang road tunnel project in Wuhan [citation_10fa2w].	KPCA-DNN with counterfactual explanations and analysis of factor importance, necessity, and sufficiency [citation_10fa2w].	The model outperformed AdaBoost and random forest in the reported case and produced an operational risk path; multi-project validation was not reported in the retrieved passage [citation_10fa2w].	[citation_10fa2w]

Deployment, impact, and design criteria

Most evidence evaluates models rather than sustained use. The delay-mitigation framework’s ten professional interviews found favorable views of applicability and practicality but lower confidence in reliability, and the authors state that real-life deployment is needed for definitive validation [citation_1x98a1] [citation_1x98a7]. The productivity-dashboard study also identifies the need for longer-term evaluation and broader project application [citation_15qdpi]. These limitations matter because a prediction system can change reporting behavior, intervention timing, and the data later used for retraining.

A rigorous PRISM-AI evaluation should therefore use staged evidence. Retrospective validation tests whether the model predicts; silent prospective deployment tests whether it remains calibrated without influencing decisions; pilot deployment tests usability and alert burden; and a controlled or quasi-experimental evaluation tests whether the system changes decisions or outcomes. Measures should include intervention lead time, accepted-alert proportion, reviewer disagreement, mitigation completion, schedule or cost outcomes, and unintended effects such as risk displacement or excessive escalation.

The literature supports six design criteria. First, the platform should be multi-outcome and preserve the distinction among cost, schedule, quality, safety, and public impact [citation_1x9891]. Second, it should combine static project attributes with temporal control and execution data where available [citation_1x98e9] [citation_1xfz8m]. Third, it should compare transparent baselines with ensemble and hybrid candidates rather than optimizing a single score [citation_10urbu]. Fourth, explanations should distinguish association, uncertainty, and feasible intervention [citation_1aoyi8] [citation_10fa2w]. Fifth, the platform should integrate with project-control workflows and maintain feedback and retraining loops [citation_10urbu]. Sixth, deployment evaluation should test external validity, reliability, human use, and outcome impact rather than stopping at retrospective accuracy [citation_13s8vb].

Methodological Comparison

The evidence base contains primary predictive studies, hybrid schedule simulations, expert-derived risk models, dynamic monitoring feasibility studies, and review-level syntheses. Primary studies provide direct measurements but are often narrow in setting; reviews reveal recurring field-level constraints but cannot replace validation of the individual models they summarize. Fitzsimmons and colleagues provide one of the more substantial infrastructure datasets and a project-level holdout, but the evaluation remains limited to ten case projects and is affected by retrospective schedule information [citation_1x98es] [citation_1x98fd]. Sanni-Anibire and colleagues connect modelling to a deployment framework and professional review, but their reliability assessment was weaker than their applicability assessment and did not constitute field deployment [citation_1x98a2] [citation_1x98a7].

The expert-derived studies illustrate a different trade-off. Tall-building risk assessment can encode professional knowledge when operational data are scarce, but the 48-response dataset and ANN result cannot be assumed to represent observed project trajectories [citation_11gdit]. The 2026 explainable infrastructure study improves decision translation by linking predictions to risk bands and actors, yet it remains based on one programme with partly expert-informed observations [citation_1aoyi8]. These designs are appropriate for feature discovery and framework development, but weaker for estimating population-level intervention effects.

Dynamic monitoring studies move closer to operational use. Giretti and colleagues use sequential probabilistic inference from monitoring networks and explicitly relate accuracy to data resolution and quality [citation_1xfz8m]. The digital-twin study links sensors, LSTM prediction, alarms, and resource optimization in three projects [citation_1wuit6]. Such closed-loop architectures are promising, but their evaluation should be extended beyond technical feasibility to independent replication, alert burden, data drift, and decision outcomes.

Across designs, the most defensible evidence hierarchy for PRISM-AI is: externally validated prospective prediction; project-level temporal or leave-one-project-out validation; retrospective project-level validation; within-dataset cross-validation; and expert or simulated demonstration. This hierarchy does not make lower-level studies useless. It determines which claims they can support. A small expert study can identify candidate features; it cannot establish generalizable risk probabilities. A simulation can test architecture; it cannot prove public-project impact. A high retrospective score can justify further testing; it cannot by itself justify automated intervention.

Research Gaps

PRISM-AI should address the field’s gaps through a research programme rather than by adding more algorithms to a single dashboard. The evidence is concentrated in individual programmes, countries, sectors, or project types. The explainable-delay study uses one large-scale infrastructure programme [citation_1aoyi8], the counterfactual study uses one tunnel project [citation_10fa2w], and the UK hybrid study evaluates ten held-out projects within a UK infrastructure database [citation_1x98es]. This concentration limits claims about transfer to public portfolios that contain different asset classes, procurement arrangements, institutional capabilities, and reporting practices. Multi-owner, multi-region validation is a central gap.

Many studies use historical or expert-derived data, while real monitoring systems change the process they observe. New procurement rules, market conditions, design practices, contractors, and reporting systems can alter feature-outcome relationships. The early-warning study’s expanding-window design is a useful direction [citation_1wxvze], but the field needs repeated prospective evaluations with drift monitoring, model recalibration, and explicit treatment of interventions. A model that predicts risk before corrective action may perform differently after teams begin acting on its alerts.

SHAP and related methods explain model behavior, not necessarily the causes of project outcomes. The infrastructure XAI study makes this distinction explicit [citation_1aoyi8], while methodological work warns that SHAP and LIME depend on model choice and feature collinearity [citation_108ze6]. Research should test explanation stability, user comprehension, calibration of confidence, and whether explanations improve decisions. Counterfactual systems also need feasibility and engineering-constraint checks before they can be treated as recourse recommendations [citation_10fa2w].

Construction risk models may inherit uneven documentation, unequal access to information, and organizational biases. The construction-risk review identifies fragmented treatment of data governance and ethical risks and insufficient attention to worker and organizational factors [citation_1xfz75]. Public-sector research should examine whether alerts systematically over-prioritize certain contractors, locations, communities, or work types because of historical reporting practices. Governance studies should connect model records to procurement, audit, privacy, and public-accountability requirements.

The literature commonly reports predictive performance, but much less often reports whether model use improves schedule reliability, cost control, safety, quality, or public-service delivery. The 2021 framework authors explicitly call for real-life deployment to compare performance against existing approaches [citation_1x98a7]. Future work should measure decisions and outcomes longitudinally, including the cost of false alarms and the consequences of missed warnings. PRISM-AI’s strongest contribution would be evidence that explainable monitoring changes the quality and timing of decisions, not simply that one algorithm achieves a high score.

A practical research programme should construct a multi-project, versioned data model spanning schedules, cost controls, procurement, risk registers, design changes, quality, safety, BIM, and selected sensor sources. It should define outcomes and horizons with project teams before training, including what counts as a warning and what actions are feasible. Transparent statistical and rule-based baselines should be compared with tree ensembles and hybrid probabilistic models using project-level and chronological validation. Explanations should be evaluated as interventions: users should identify relevant evidence, distinguish association from causation, select feasible responses, and recognize uncertainty. Counterfactuals should be filtered through engineering, contractual, safety, and authority constraints [citation_10fa2w]. Prospective monitoring should include model cards, drift checks, calibration monitoring, audit logs, and explicit human override. Impact evaluation should measure lead time, mitigation completion, alert burden, schedule and cost outcomes, and equity-relevant consequences. The platform should be allowed to abstain when data quality or domain shift makes a forecast unreliable.

Synthesis/Conclusion

The literature supports PRISM-AI as a human-supervised platform that joins project controls, machine learning, explainability, and operational governance. The strongest technical direction is hybrid: use structured project and temporal data, preserve schedule or process relationships, quantify uncertainty, and connect predictions to feasible decision pathways. The strongest governance direction is equally clear: treat explanations as evidence about model behaviour, retain professional and institutional accountability, and make data lineage and uncertainty visible.

Current evidence is sufficient to justify a carefully staged platform study, not an autonomous public-project controller. PRISM-AI should be judged by external validity, calibration, timely and credible explanations, user decisions, and delivery impact across multiple infrastructure contexts. Its distinctive contribution would be to close the gap between predictive performance and accountable action.

Generated Outputs

[SANDBOX_OUTPUT:7e06635a86dca407:prism-ai-literature-review.md]