Search Results (279)

Agentic artificial intelligence (AI) offers new opportunities for intelligent clinical decision support, but deployment in emergency and crisis settings remains challenging because time-critical recommendations must often be generated under incomplete patient information and system constraints. Conventional clinical decision support systems rely on rule-based workflows that degrade when structured data are absent, while standalone language models lack coordination mechanisms to enforce mandatory safety checks. We present TriAgent, a multi-agent framework that unifies adaptive orchestration, iterative retrieval, embedded safety verification, and end-to-end auditability within a single crisis clinical decision support workflow. An Orchestrator Agent dynamically selects specialist modules for clinical assessment, retrieval, treatment planning, safety verification, and system coordination, with routing determined by model reasoning rather than fixed execution paths. A retrieval sub-agent performs iterative query refinement and relevance grading over 49,000 MIMIC-IV discharge notes, while medication-conflict screening and allergy-risk assessment are invoked in parallel only when clinically indicated. A Critique Agent reviews the full reasoning trace before recommendation finalization. In a retrospective evaluation on 1000 real emergency presentations under synthesized incomplete-information inputs, TriAgent achieved 85.0% critical-case recall and 65.7% overall triage accuracy, versus at most 14.7% and 43.4% for matched single-model and retrieval-only baselines, with safety checks executed on every continuation pathway and adaptive routing invoking only the modules each case required. These results support multi-agent orchestration as a promising design pattern for transparent and auditable AI in healthcare. These gains are internal system properties; clinical-safety benefit remains to be established through prospective, clinician-involved validation. Full article

As High-Performance Computing (HPC) architectures evolve towards the Exascale, storage infrastructure reliability has emerged as a critical operational challenge, with traditional reactive and static preventive maintenance strategies proving increasingly insufficient. This study addresses this gap by proposing a comprehensive methodological framework for the transition from predictive to predictive-prescriptive maintenance in large-scale storage environments. By integrating the CRISP-DM industry standard with a multi-layered eXplainable Artificial Intelligence (XAI) suite, we develop a system capable of isolating hardware degradation signals amidst massive volumes of routine telemetry. To validate our approach, we leveraged a publicly available disk failure dataset to evaluate multiple Machine Learning configurations, addressing the challenge of severe class imbalance through optimized oversampling and Gradient Boosting algorithms. The methodology employs global and local XAI techniques, including Permutation Feature Importance, SHAP, and surrogate decision trees, to translate probabilistic risk assessments into auditable hardware engineering rules. Our results demonstrate that this hybridization of robust predictive modeling with multi-layered explainability provides a transparent, evidence-based decision support system. Ultimately, we conclude that converting opaque risk predictions into technical justifications enables infrastructure managers to optimize hardware lifecycle management and minimize system downtime in mission-critical environments, establishing a viable pathway toward more resilient and auditable storage management. Full article

Highway construction projects face interacting risks that affect time, cost, regulatory compliance, and delivery resilience, all of which are closely linked to sustainable infrastructure development. This study develops a hybrid decision-analytical and machine learning framework for sustainability-oriented risk assessment in highway construction under limited-data conditions. The framework combines (i) the Analytic Hierarchy Process (AHP) and tabular Generative Adversarial Networks (GANs) to structure and stress-test expert judgement, and (ii) Probability-Impact (P-I) scoring with a Bayesian Networks (BNs) to model dependencies and derive posterior weights for probability of occurrence, impact on time, and impact on cost across four headline risk factors: weather-related risks, lack of labour, design-related risks, and permitting/regulatory risks. AHP provides transparent and auditable priorities with consistency checks, while GAN-generated synthetic tables support diagnostics for central tendency (P50) and tail behaviour (P90) under data scarcity. The calibrated P-I scores parameterise BN conditional probability tables, enabling the updating of BN scores; and factor-level decomposition of expected contributions. The framework produces model-ready posterior weights that support early planning, contingency allocation, mitigation prioritization, scenario analysis, and subsequent simulation and optimization studies. In sustainability terms, the proposed approach helps project teams improve climate resilience, strengthen regulatory and environmental preparedness, and reduce inefficient use of time, cost, and project resources in data-constrained settings. The results show that permitting/regulatory risks have the highest contribution to probability of occurrence and time impact, while weather-related risks exert the greatest cost impact. The framework therefore offers a practical tool for supporting more resilient, transparent, and sustainable highway project delivery when large historical datasets or questionnaire surveys are unavailable. Full article

AI-enabled cyber–physical systems (CPSs) are increasingly deployed in public governance contexts where they sense human populations, infer classifications or risks, and trigger interventions that can shape liberty, equality, and access to essential services. In these deployments, governance failures often arise not only from model error but from systems-level interactions across data generation, model updates, organizational practices, and downstream actuation. This paper introduces a Risk–Rights–Rules (3R) architecture that treats fundamental rights and legal rules as enforceable constraints on the sensing–inference–actuation loop, rather than as external ethical aspirations. Building on established risk-management baselines and safety engineering practice, we specify a testable assurance object, a structured 3R assurance case, that links rights claims to explicit assumptions, measurable evidence, and accountable control points across the lifecycle. The approach is designed to reduce “legitimacy drift” in stochastic decision pipelines by making uncertainty, demographic error, contestability, and procurement leverage auditable at the system level. The result is a governance blueprint for high-consequence public-sector AI deployments for governance failures, which is both technically robust and institutionally defensible. Full article

Background: Large language models and other generative artificial intelligence (AI) tools are increasingly being embedded in digital healthcare services, including mobile health applications, telerehabilitation, remote monitoring, and hybrid care pathways. In this review, digital healthcare refers to technology-mediated healthcare services in which digital platforms, mobile applications, wearables, remote communication, and AI-enabled interfaces support health assessment, self-management, rehabilitation, clinical decision support, or service delivery. When AI-generated exercise guidance moves from general education to individualized recommendations about dose, progression, contraindications, or rehabilitation, it may become directly actionable and safety-relevant. Objectives: This review aimed to clarify when AI-generated exercise guidance in digital healthcare may warrant safety-relevant governance attention and to outline implementation considerations for explainability, human oversight, and service-level governance. It addresses a gap in the literature: general AI-governance and exercise-prescription discussions rarely specify how point-of-use explanations, review thresholds, and escalation safeguards can be organized for directly actionable AI exercise guidance. Methods: We conducted a governance-oriented narrative review of peer-reviewed literature and representative regulatory or guidance documents. This review was not designed as a systematic review, scoping review, or exhaustive evidence map; transparent source mapping was used to support conceptual synthesis. Searches and source mapping focused on generative AI, large language models, explainable AI, clinical decision support, digital health, mobile health, exercise prescription, rehabilitation, trust, automation bias, and human oversight. Sources were included when they informed the safety, explainability, governance, or real-world implementation of patient-facing AI-generated exercise guidance. Extracted material was grouped by evidentiary role and synthesized through framework synthesis and governance mapping to distinguish literature-supported observations, author interpretation, and proposed implementation tools. Results: The included sources were first organized into five thematic groups: digital exercise delivery and exercise-prescription evidence; explainability, trust, and automation bias literature; professional responsibility, ethics, and patient disclosure literature; regulatory and policy documents; and digital literacy, patient/clinician attitudes, and equity literature. The synthesis then proceeded from safety relevance to explanation needs, human oversight and escalation needs, and selected regulatory and policy signals before translating these strands into conceptual and implementation-oriented outputs rather than empirically validated instruments. AI-generated exercise guidance was most safety-relevant in scenarios involving individualized dose, progression, contraindication-sensitive action, or rehabilitation strategy. Across the included sources, generic transparency alone was not sufficient to support reviewable use; relevant explanation elements included evidence sources, risk warnings, reasoning paths, and reasonable alternatives. Oversight considerations varied with embodied risk, clinical ambiguity, user vulnerability, and likelihood of direct enactment. Implementation considerations linked interface design, clinical review, escalation, auditability, and post-deployment monitoring. Conclusions: AI-generated exercise guidance in digital healthcare may warrant governance attention as a patient-safety and accountability issue when it influences actionable exercise decisions. The proposed framework offers a conceptual basis for designing more reviewable and accountable mobile and remote exercise-support services. Future work can validate these outputs in patient-facing services, clinician review workflows, usability studies, implementation pilots, and safety evaluations. Full article

Artificial intelligence is reshaping warehouse work through algorithmic task allocation, scanner-based monitoring, KPI feedback, dynamic scheduling, and real-time performance control. Although these systems can improve coordination and operational visibility, they also create governance risks related to fairness, transparency, autonomy, privacy, workload pressure, trust, and employee resistance. This study develops a human-centered decision framework for prioritizing algorithmic management governance packages in third-party logistics (3PL) warehousing. The main contribution is to translate employee-level governance concerns into a scenario-sensitive decision model that helps managers select appropriate governance packages under different operational pressures. The study uses survey data from 380 warehouse employees to examine key psychological and behavioral mechanisms, including procedural fairness, transparency, system/information quality, autonomy, privacy concern, workload, trust, acceptance, and resistance/disengagement. These survey-supported constructs are then converted into six governance criteria: procedural fairness, transparency and contestability clarity, system and information quality, autonomy support, privacy boundary governance, and workload protection. A seven-expert panel evaluates five governance packages under three scenarios: peak season surge, labor shortage/high turnover, and audit pressure/compliance scrutiny. Methodologically, the framework combines Dynamic Multi-Facet Fuzzy Sets to capture membership, non-membership, hesitancy, engagement, and resistance; Bayesian Network weighting to reflect dependencies among governance criteria; and PCA-based ranking optimization to generate scenario-specific and robust rankings. Comparative validation with SAW and TOPSIS is also used to assess ranking consistency. The findings show that effective algorithmic management governance is not a fixed compliance solution. Transparency, workload protection, autonomy support, privacy boundary governance, and procedural fairness become more or less important depending on the operational scenario. A2, which combines transparency, workload protection, and autonomy support, emerges as the strongest robust package. A1 performs best under labor shortage/high turnover, while A3 performs best under audit pressure/compliance scrutiny. These results suggest that 3PL warehouses should adopt adaptive governance routines that combine explainability, contestability, workload safeguards, privacy boundaries, and employee voice mechanisms. The study contributes to the literature on AI in socio-technical systems by showing how human, organizational, and ethical concerns can be embedded into an interpretable decision framework for responsible algorithmic management in logistics work environments. Full article

Generative artificial intelligence (GAI) is increasingly embedded in personal finance, yet little is known about how models make recommendations using financial information and demographic cues. This study audits three frontier GAI models, GPT 5.5, Gemini 3.1 Pro, and Claude Opus 4.7, using a conjoint experiment in which each model evaluated the same hypothetical investor profiles and selected among standardized conservative, balanced, and aggressive portfolios. Investor profiles systematically varied attributes, including risk tolerance, time horizon, goal type, income, and age, gender, ethnicity, marital status, and employment type. Ordered logistic regressions and matched-profile comparisons show that all three models base recommendations primarily on financial attributes, especially risk tolerance and time horizon. Age and marital status shift recommendations towards conservatism in all models, conversely only Claude conditions on gender and employment type. Ethnicity exerts no detectable influence on the recommendations of ChatGPT or Claude, but is a small, statistically significant predictor for Gemini, with non-White profiles receiving slightly more conservative recommendations than otherwise identical White profiles. Overall, we find that the models are not interchangeable: they differ significantly in overall risk appetite and in how they translate risk tolerance, time horizon, goal type, and age into portfolio choices, with economically meaningful differences in predicted recommendations for identical clients. These findings suggest that contemporary GAI investment advice is driven mainly by financially relevant attributes, but that demographic sensitivity may appear in model-specific and statistically nuanced ways, alongside a distinct form of platform risk arising from model-specific advisory logic. Full article

Crash severity prediction is critical for emergency response, infrastructure spending, and risk communication. Although machine learning has been widely applied to this problem, three gaps prevent practical deployment: uncalibrated probability scores, SHAP-based explanations whose faithfulness has not been verified, and models never tested outside their training jurisdiction. The proposed framework, SAE-XCrash (Safety-Aware and Explainable Crash Severity Prediction), addresses all three using two public datasets—US-Accidents (7.0 million records, 2016–2023) and UK STATS19 (approximately 1,010,000 records, 2016–2022)—with strict temporal splits throughout. Notably, the US-Accidents severity label measures traffic disruption duration, not injury outcome; results should be interpreted accordingly. Previously unknown label-schema drift led to a revised binary target with Severity 4 as the only positive class. Five classifiers are compared. Post hoc isotonic calibration reduces Expected Calibration Error by 97.3% at negligible discrimination cost. A four-step quantitative SHAP audit confirms statistically significant deletion faithfulness; however, explanation stability fails at realistic perturbation levels (54.3% low-stability fraction at sigma = 0.05), driven by spatial data sparsity in sparse geohash cells—a negative result that carries direct operational implications for deployment. A three-tier cross-dataset transfer experiment (zero-shot, recalibration, full retrain) shows that temporal features transfer robustly across jurisdictions, while spatial memorization is the primary generalization barrier. All code, split indices, and model artifacts are publicly available. Full article

Background: Acute kidney injury (AKI) is one of the more serious complications following cardiac surgery, consistently linked to prolonged mechanical ventilation and higher in-hospital mortality. This study examined whether preoperative anaemia and impaired renal function are associated with AKI and death in a high-risk cardiac-surgery cohort requiring extended postoperative ICU monitoring and how these associations behave after adjustment for procedure type and intraoperative variables. Methods: In this single-centre retrospective cohort study, we screened 950 patients admitted to a cardiothoracic ICU between January 2018 and January 2024. After standard exclusion criteria and an audit of operative records, 553 cardiac-surgery patients formed the principal analysis cohort. AKI was defined by KDIGO criteria using serial postoperative serum-creatinine measurements during the first 7 days. Multivariable logistic regression for AKI and in-hospital mortality was built sequentially: Model A (baseline only); Model B (+procedure type); and Model C (+intraoperative variables: aortic cross-clamp time, intraoperative RBC units, and intraoperative inotrope use). Calibration was assessed by the Hosmer–Lemeshow test. Total cardiopulmonary bypass duration was not separately captured in the institutional database and is disclosed as a limitation. Results: AKI occurred in 174 of 553 patients (31.5%), and in-hospital mortality was 16.6% (92/553). Patients with AKI were older (median 77 vs. 68 years, p < 0.001), with lower preoperative haemoglobin (11.4 vs. 12.3 g/dL, p < 0.001) and lower eGFR (38.1 vs. 63.7 mL/min/1.73 m², p < 0.001). The aortic cross-clamp time was shorter in AKI patients (56 vs. 70 min, p = 0.043), a counterintuitive finding likely reflecting residual confounding by case-mix and procedure selection rather than a protective operative effect. In the fully adjusted multivariable model, the haemoglobin–AKI association attenuated and was no longer independently significant (OR 0.89 per 1 g/dL, 95% CI 0.73–1.08, p = 0.24), while intraoperative RBC transfusion emerged as an independent predictor (OR 1.12 per unit, p = 0.046). For mortality, AKI remained an independent predictor after full adjustment for procedure type and intraoperative variables (OR 7.14, 95% CI 1.45–35.13, p = 0.016), with cross-clamp time (OR 1.30 per 10 min, p = 0.010) and intraoperative RBC units (OR 1.48 per unit, p < 0.001) also independently associated. Both fully adjusted models showed acceptable calibration (Hosmer–Lemeshow p = 0.48 for AKI, p = 0.56 for mortality). Conclusions: In cardiac-surgery patients with a prolonged ICU stay, AKI is independently associated with in-hospital mortality even after adjustment for operative variables. The univariable association between preoperative haemoglobin and AKI is attenuated after adjustment for procedure type and intraoperative transfusion exposure, suggesting confounding or mediation by operative and case-mix factors rather than an independent direct association. The contribution of this analysis is aetiological/analytical rather than predictive (modest discrimination, AUROC 0.67 for AKI), and findings should be interpreted within the selected high-risk ICU ≥ 72 h population. Full article

Fraud detection models often achieve a strong ranking performance through black-box ensembles, but operational deployment also requires calibration, low explanation cost, and auditable scoring logic. This study develops an interpretable fraud-detection pipeline for IEEE-CIS by combining a 63-variable causal concept bank with teacher-guided additive Explainable Boosting Machine (EBM) students. The concept bank summarizes the temporal state, entity history, novelty/reuse, identity missingness, and aggregate deviation. Experiments use a chronological out-of-time split and a stricter pseudo-entity-disjoint holdout. In the main three-seed evaluation, the CatBoost predictive ceiling and XGBoost teacher achieved PR-AUC 0.489 ± 0.001 and 0.478 ± 0.003, respectively. Among interpretable models, concept-only EBM reached 0.189 ± 0.000, raw-only EBMs reached 0.372 ± 0.005 (top-k = 8) and 0.383 ± 0.002 (top-k = 12), and hybrid EBMs reached 0.407 ± 0.003 (top-k = 8) and 0.407 ± 0.004 (top-k = 12), consistently improving over matched raw-only additive baselines. The final top-k = 8 hybrid reduced input features from 154 to 71, achieved about 9.7× faster inference than XGBoost, remained close to XGBoost in ECE-15 calibration (0.01587 vs. 0.01611) while having a higher Brier score, and produced native local explanations far faster than XGBoost + SHAP. The results position CatBoost as the predictive ceiling and hybrid EBM as a benchmark-supported, deployment-relevant interpretable compromise for applied financial risk-screening workflows, rather than as a production-validated fraud-monitoring system. Full article

Accurate viral protein function annotation is essential for genomic surveillance, yet conventional retrieval-augmented generation (RAG) pipelines often fragment biological evidence into fixed-length text chunks, disrupting relationships among ORFs, annotations, structural domains, sequence motifs, residue mappings, and model-derived attention evidence. We propose ViroBioTree, a tree-structured biological evidence retrieval framework for downstream viral protein evidence review rather than a new primary annotation classifier. Built as an evidence organization layer on ViralMultiNet-derived ORF-level predictions and annotations, ViroBioTree converts sequence, annotation, structure, and attention evidence into typed biological nodes and traceable edges, then performs deterministic multi-channel recall, evidence-aware reranking, balanced TopK selection, rule-based verification, and node-cited report generation. In a demo benchmark, ViroBioTree achieved its strongest deterministic proxy performance on structure-explanation tasks, with Precision@K = 1.0, Recall@K = 1.0, and diversity = 0.52; these values reflect expected node-type and tag agreement rather than independent biological correctness. A bounded full-scale SARS-CoV-2 index contained 39,800 ORF rows, 80,000 attention records, 199,418 nodes, and 495,886 edges. In a stratified full20k diagnostic evaluation, ViroBioTree showed task-dependent advantages over LlamaIndex vector retrieval for conflict detection, evidence retrieval, and structure explanation, while LlamaIndex remained competitive or stronger for annotation-rich function annotation. A cross-family Influenza A Virus (IAV) diagnostic audit showed that the schema can represent IAV evidence namespaces while explicitly exposing missing formal ORF inputs, missing attention evidence, and unavailable residue/PDB assertions. Supplementary robustness, external sanity-check, diversity-risk, expert-evaluation, domain-tool positioning, and cross-family audit analyses supported traceability, report quality, and conservative evidence handling, but also showed that stable Precision@K under query perturbation does not necessarily imply stable retrieved evidence sets. ViroBioTree operates offline and deterministically, but does not address raw-read assembly, base calling, primary ORF prediction, or wet-lab validation. Its results should be interpreted as proxy and expert-reviewed evidence for traceable viral protein evidence retrieval and report generation rather than as direct validation of biological function annotation. Full article

The increasing volume of digital health data generated through Electronic Health Records (EHRs), emergency care systems, and real-time monitoring technologies has intensified the need for secure cross-institutional healthcare analytics. However, privacy concerns, regulatory restrictions, institutional mistrust, and risks associated with centralized data aggregation continue to limit large-scale healthcare data sharing. This paper presents OPAL4Health, a governance-oriented and privacy-preserving distributed healthcare analytics framework grounded in the MIT Open Algorithms (OPAL) paradigm. The framework integrates federated analytics, blockchain-based auditability, explainable artificial intelligence (XAI), and institutional governance mechanisms within a unified computation-to-data healthcare ecosystem. Unlike conventional federated healthcare systems that primarily focus on decentralized computation alone, OPAL4Health emphasizes governance, transparency, auditability, and policy-aligned distributed analytics while preserving institutional data sovereignty. The privacy protections supported by OPAL4Health are primarily architecture-based and governance-oriented, relying on local institutional data retention, controlled query execution, and blockchain-auditable analytical workflows rather than formally provable cryptographic privacy guarantees. The framework was evaluated through a real-world urgent care pilot across seven hospitals in Riyadh, Saudi Arabia, using 184 anonymized patient cases collected between May 2015 and September 2016. Analytical findings identified a median onset-to-arrival delay of 285 min (95% Confidence Interval (CI): 270–302), low ambulance utilization (18.5%), and hospital bypass behavior in 42% of cases. Peak Emergency Department (ED) congestion periods were also identified. Scenario-based modeling projected potential long-term healthcare savings of approximately $602 million over 15 years through improved Emergency Medical Services (EMS) allocation and reduced disability-adjusted life years (DALYs). The findings demonstrate the feasibility of governance-oriented, privacy-preserving distributed healthcare analytics within OPAL4Health while generating actionable operational and policy-relevant insights without centralizing sensitive patient-level records. The proposed framework provides a transferable model for secure, transparent, and accountable digital health collaboration across healthcare ecosystems. Full article

As cloud computing proliferates, outsourced data faces severe security threats, yet existing searchable encryption (SE) schemes rely on classical hardness assumptions, centralized trust authorities, and static access control, leaving critical gaps in quantum resistance, single-point-of-failure prevention, and dynamic permission management. To address these limitations, we propose BL-ABSE, a blockchain-enhanced, lattice-based attribute-based searchable encryption framework. BL-ABSE employs the Ring Learning With Errors (RLWE) problem as its security foundation and applies the Number Theoretic Transform (NTT) to reduce polynomial multiplication from $O\left(n^2\right)$ to $O(nlogn)$. To eliminate single-point trust risks, the framework further integrates a $(t,n)$ threshold key protocol across an edge-node consortium governed by Practical Byzantine Fault Tolerance (PBFT) consensus. A smart-contract-maintained on-chain revocation list enables permission withdrawal via a single blockchain transaction without re-encryption. Experimental evaluation demonstrates that commitment generation requires approximately 23 ms at $n=1024$, search latency scales linearly at roughly 29 µs per record, and revocation completes in approximately 2 s regardless of system scale. Formal security proofs under the quantum polynomial-time (QPT) adversary model reduce six security properties—index indistinguishability, query privacy, threshold key security, Byzantine fault tolerance, audit immutability, and revocation immediacy—to the hardness of RLWE and the Short Integer Solution (SIS) problems. To the best of our knowledge, BL-ABSE is the first framework to simultaneously achieve post-quantum security, attribute-based access control, decentralized key management, instant revocation, and immutable auditing within a single unified framework. We further conduct threshold parameter verification, end-to-end revocation latency decomposition, blockchain throughput stress testing, search-pattern leakage quantification, and communication/storage overhead analysis, providing a comprehensive evaluation of both performance and security trade-offs. We explicitly characterize the search-pattern leakage inherent in the deterministic commitment design as a correctness–privacy trade-off and discuss mitigation directions. Full article

The rapid proliferation of Narrowband Internet of Things (NB-IoT) devices necessitates robust, privacy-preserving intrusion detection systems. While Federated Learning (FL) mitigates data privacy risks through localized training, it introduces vulnerabilities to model poisoning and computational bottlenecks on edge devices. To address these challenges, we propose a secure, hardware-optimized Blockchain-Federated Learning (BC-FL) framework. Deploying a lightweight Hybrid CNN-RNN model on Edge Gateways, we relieve end-sensors of heavy computational tasks. To overcome the ‘cold-start’ problem, we introduce a Domain-Adaptive Transfer Learning strategy, dynamically adapting a pre-trained binary classifier to a multi-class task (Normal, Mirai, Bashlite). Furthermore, a lightweight blockchain ledger provides an immutable audit trail and a reputation-based isolation mechanism to penalize malicious nodes. Evaluated on the N-BaIoT dataset, the proposed 3-class CNN-RNN model achieves 95.62% overall accuracy, with precision/recall/F1-scores of 0.99/0.91/0.95 for Mirai and 0.93/0.99/0.96 for Bashlite attacks. The framework reduces communication bandwidth by 96% compared to centralized learning. During simulated Byzantine attacks, the reputation mechanism successfully banned malicious nodes, maintaining a robust 95.62% global accuracy. This framework offers a highly scalable, secure, and computationally feasible solution for real-time anomaly detection in resource-constrained IoT edge environments. Full article

Accurate volatility forecasting and timely risk early warning are foundational requirements of financial risk management: Value-at-Risk estimates, portfolio risk limits, derivative hedging ratios, and stress-test scenario calibrations all depend on forward-looking volatility signals that remain reliable when market conditions depart from average. This paper develops a public-data causal multiscale wavelet spillover learning (CMWSL) framework that jointly addresses stock-index volatility forecasting and high-volatility early warning under strict walk-forward evaluation. CMWSL integrates three components: a heterogeneous autoregressive (HAR) persistence block as the dominant linear baseline, causal stationary wavelet transform (SWT) summaries that encode within-index multiscale market dynamics, and a cross-index spillover layer that tests whether medium- and long-scale wavelet energy from peer indices carries incremental risk-relevant information. The empirical analysis covers the S&P 500, Nasdaq-100, and Dow Jones Industrial Average over a 2513-step out-of-sample evaluation period from 2016 to 2025, with forecast horizons $h\in \{1,5,10\}$ and OHLC-based volatility targets. All preprocessing, wavelet decomposition, calibration rules, and warning thresholds are re-estimated inside each rolling training window to eliminate look-ahead bias. HAR remains the strongest average model in the main Rogers–Satchell specification, confirming that daily index volatility risk is highly persistence-driven. The multiscale extension delivers statistically significant improvements at longer horizons, in richer public macro-financial information environments, and under the Parkinson target. Clark–West tests detect significant spillover gains in five of nine index–horizon cells (CW $=4.83$, $p<0.001$ for S&P 500 at $h=10$). Critically, tail-conditioned and rolling-window diagnostics show that multiscale and cross-index gains concentrate in upper-volatility regimes and synchronized stress episodes—precisely the conditions in which risk management decisions are most consequential. For market-risk early warning, a logistic classifier built on the same causal feature pipeline delivers the most stable precision–recall performance across all settings, providing an interpretable and operationally auditable alert mechanism suitable for practical risk monitoring. Full article