An explainable machine-learning framework designed for retrospective classification of duration-defined infrastructure-level workload status has demonstrated higher classification performance than baseline configurations in large-scale healthcare analytics. Developed to address the need for reliable and transparent operational performance when processing heterogeneous computing tasks, the framework was evaluated on a distributed Hadoop cluster executing National Early Warning Score (NEWS) workloads across four operational scales. In these evaluations, the complete optimisation pipeline achieved higher reported classification scores alongside shorter per-fit training times than baseline setups.
The experimental investigation relied on a partitioned infrastructure design that separated workload processing from model training. The distributed Hadoop cluster environment was utilized for executing the National Early Warning Score workloads and gathering operational infrastructure metrics, while all classifier development and performance evaluations were executed on a local workstation. Across the experimental execution runs, the researchers compiled a modelling dataset consisting of 24,000 task-level profiles extracted from SmartMonit monitoring records collected during cluster operations.
At the core of the framework is the retrospective classification of duration-defined workload status using extreme gradient boosting (XGBoost) and support vector machine (SVM) classifiers. To ensure transparency and refine classification accuracy, the pipeline combines classifier development with SHAP-informed feature-weight optimisation using the Covariance Matrix Adaptation Evolution Strategy (CMA-ES). The methodology additionally integrates Local Interpretable Model-agnostic Explanations (LIME), SHapley Additive exPlanations (SHAP), and partial dependence plots (PDP) to analyze how different infrastructure features influence model predictions.
The empirical findings demonstrated notable distinctions between training speeds, optimisation pipelines, and classification metrics across the tested setups. While per-fit training times were shorter under the complete optimisation pipeline than in baseline configurations, the authors noted that these specific duration measurements excluded the computational costs associated with model selection and optimisation procedures. Crucially, feature weighting performed via matched-budget random search achieved numerically higher mean F1-scores than the SHAP-informed CMA-ES optimisation strategy across both the XGBoost and SVM classifiers.
Model interpretability analyses conducted with SHAP and LIME identified data-splitting times and mapper-execution times as particularly influential workload indicators for duration status. Overall, the study demonstrates that these findings support a rigorous comparative analysis of machine-learning classification and post-hoc explanation techniques for retrospective, duration-defined task status within the evaluated Hadoop cluster environment, providing foundational insights for infrastructures tasked with handling large-scale, heterogeneous healthcare analytics workloads.
Subject of Research: Explainable workload classification in large-scale healthcare analytics infrastructures
Article Title: Interpretable infrastructure-level workload prediction for large-scale healthcare analytics
Article References: Arslan, S., Demirbaga, U., & Eken, S. (2027). Interpretable infrastructure-level workload prediction for large-scale healthcare analytics. Cluster Computing, 30(1), Article 51. https://doi.org/10.1007/s10586-026-06847-7
Image Credits: AI Generated
DOI: 10.1007/s10586-026-06847-7
Keywords: healthcare analytics, Hadoop cluster, workload prediction, extreme gradient boosting, support vector machine, LIME, SHAP, partial dependence plots, CMA-ES, National Early Warning Score, SmartMonit, feature-weight optimisation, data-splitting times, mapper-execution times, task-level profiling, random search, retrospective classification
News Source: Denise Maddox. (November 19, 2026). Explainable Models Classify Retrospective Healthcare Workload Durations in Cluster Computing Evaluation. Scienmag.



