A dual-stream deep learning architecture for business impact scoring and alert escalation

Clinical Snapshot

Key Findings

• P Value: Not reported

• Effect Size: Precision 0.94, Recall 0.92, F1-score 0.93 on combined KPI dataset; Impact-Weighted F1 (IW-F1) 0.85, BIS accuracy 0.88; claimed 31.6% reduction in operational costs

• Primary Outcome: Multi-class alert escalation and anomaly detection performance on combined multivariate KPI dataset and AI4I_2020 predictive maintenance dataset

• Nnt Or Sensitivity: ROC-AUC 0.94, PR-AUC 0.95 on combined KPI dataset; no sensitivity/specificity breakdown or NNT equivalent reported; no threshold-specific operating point analysis described

• Confidence Interval: Not reported for any metric

Clinical Application

Feasibility in healthcare settings is entirely undemonstrated. Implementation would require: integration with existing hospital IT monitoring platforms (e.g., SIEM systems, clinical network management tools), validation on healthcare-specific telemetry data, compliance with data governance frameworks (Australian Privacy Act, My Health Records Act), and clinical IT governance approval. Computational infrastructure requirements are not specified. The absence of real-world deployment data makes feasibility assessment speculative. This paper has no direct relevance to PBS listings, TGA regulatory pathways, or RACGP clinical guidelines. In the Australian healthcare context, potential indirect relevance exists for: (1) Australian Digital Health Agency initiatives around My Health Record infrastructure reliability; (2) hospital network operations within state health departments (e.g., NSW Health, Queensland Health digital infrastructure); (3) HIMAA and ACHI health informatics governance frameworks for clinical IT system resilience. Any deployment in Australian healthcare IT would require compliance with the Australian Government's Essential Eight cybersecurity framework and relevant ACSC guidelines. The TGA's Software as a Medical Device (SaMD) regulatory pathway would apply if the system were used to influence clinical decision-making directly. No PBS or TGA considerations apply to the current study as presented. This study has no direct clinical patient population. Potential indirect applicability exists in healthcare IT infrastructure management — specifically hospital NOC environments, clinical network monitoring, electronic health record system uptime management, and medical device telemetry monitoring. The framework concept may be relevant to health informatics teams managing critical clinical IT systems where alert fatigue and escalation prioritisation are recognised operational problems.

Abstract

Modern network monitoring systems generate massive volumes of telemetry data, yet most existing anomaly detection models fail to prioritize alerts according to their operational urgency and business impact. This limitation results in delayed incident responses and inefficient alert management in Network Operations Centers. To address this gap, this study proposes a Dual-Stream Predictive Alert Escalation Framework that integrates temporal failure pattern learning with business impact-aware alert prioritization. The proposed architecture consists of two key components: a bidirectional temporal encoder for modeling multivariate Key Performance Indicator (KPI) time-series data, and an auxiliary severity encoder that captures contextual metadata related to operational risk and service criticality. The outputs of these two learning streams are combined through an attention-based fusion mechanism, and a Business Impact Scoring (BIS) layer generates impact-weighted escalation decisions for proactive incident management. Experimental evaluations using real-world KPI datasets and the AI4I_2020 predictive maintenance dataset demonstrate the superior performance of the proposed framework compared to baseline methods such as LSTM, GRU, CNN-LSTM, and BiLSTM-VAE. On the combined multivariate KPI dataset, the model achieved a precision of 0.94, a recall of 0.92, and an F1-score of 0.93, along with a PR-AUC of 0.95 and a ROC-AUC of 0.94. Under impact-aware evaluation, the framework attained the highest Impact-Weighted F1 (IW-F1) of 0.85 and BIS accuracy of 0.88, resulting in an estimated 31.6% reduction in operational costs through earlier and more accurate escalation of critical events. The suitability of the selected datasets is justified by their complementary roles: publicly available KPI time-series datasets represent real-world network telemetry behavior, while the AI4I_2020 dataset provides structured severity and operational context, enabling joint evaluation of failure prediction accuracy, escalation timeliness, and business impact modeling. By prioritizing alerts based on impact-aware severity rather than raw anomaly scores, the proposed framework directly supports operational cost reduction through earlier mitigation and improved decision-making in network operations. The proposed approach bridges the gap between anomaly detection and intelligent alert management by incorporating business relevance into predictive modeling. This dual-stream architecture offers a scalable and proactive solution for AIOps-driven network reliability and automated service resilience.

References

1. 1.Javeed, M. S., Khatun, M. M., Alom, J., Islam, R., & Shoaib, H. A. (2026). A dual-stream deep learning architecture for business impact scoring and alert escalation. PLOS ONE. https://doi.org/10.1371/journal.pone.0350676