Decision-making under uncertainty often depends on combining information from multiple competing forecasting models. Although forecast aggregation can improve predictive performance, existing approaches optimise aggregation decisions using uncertainty representations that are estimated separately and then treated as fixed throughout the decision process. This separation limits the ability of aggregation strategies to adapt as forecasting environments evolve. This paper introduces Adaptive Uncertainty-Regularised Optimisation for Robust Aggregation (AURORA), a sequential learning–decision architecture that integrates structural admissibility, adaptive uncertainty learning, and robust optimisation (RO) within an unified decision framework. Instead of treating uncertainty as a fixed modelling input, AURORA learns and updates a time-varying uncertainty representation from realised forecast-error behaviour, allowing uncertainty, model admissibility, and aggregation decisions to evolve jointly over time. An empirical evaluation across heterogeneous forecasting environments demonstrates that no aggregation rule dominates uniformly across settings. The results also show that aggregation performance depends on how uncertainty is represented and updated over time, with different uncertainty-learning mechanisms producing distinct decision behaviours under varying forms of predictive multiplicity and dependence. More broadly, the findings demonstrate the value of integrating uncertainty learning directly into the decision process, establishing a unified sequential learning–decision architecture for decision-making under evolving predictive uncertainty.
Chronic Kidney Disease (CKD) affects more than 10% of the global population and often remains asymptomatic until advanced stages, making early detection and reliable severity assessment important for clinical decision support. This study develops a hybrid decision-analytics framework that integrates fuzzy clustering with an improved adaptive neuro-fuzzy inference system (IFC-ANFIS) for CKD detection and severity stratification. The framework is evaluated using the University of California, Irvine (UCI) Chronic Kidney Disease dataset, comprising 400 patient records and 25 attributes. During data preprocessing, missing numerical and categorical values are imputed using the mean and mode, respectively, followed by Min-Max normalization. Multiple feature-selection techniques are evaluated to reduce dimensionality and computational requirements while preserving predictive information. Fuzzy clustering, supported by validated cluster-quality criteria, stratifies patients into six distinct severity groups, achieving a clustering-based classification accuracy of 95.75%. The resulting cluster-informed data are subsequently used to train the IFC-ANFIS model for binary classification of patients as CKD or non-CKD. The proposed model achieves an accuracy of 98.38%, a precision of 96.82%, a recall of 100%, and an F1-score of 98.38%. The results demonstrate that combining unsupervised patient stratification with neuro-fuzzy classification can provide an interpretable analytical framework for identifying CKD and differentiating disease severity. The proposed approach offers a data-driven decision-support mechanism that can assist in early risk identification and patient stratification while reducing the dimensionality and computational burden of CKD classification.
Variation in disease burden across age and sex is central to targeted healthcare planning, equitable insurance design, and preventive strategy, yet the age structure of diagnostic utilization is often flattened in conventional analyses. This study develops an interpretable framework that clusters cumulative age-structured diagnostic profiles to support patient stratification and resource allocation. From 583,178 visit records from Nong Han Hospital, Udon Thani, Thailand (2020–2025), primary International Classification of Diseases, Tenth Revision (ICD-10) diagnoses at first encounters were converted into cumulative age-specific probability profiles stratified by sex. Hierarchical agglomerative clustering was applied under five distance metrics; the number of clusters was selected by the Gap Statistic with the one-standard-error rule, and cluster quality was assessed with Cluster Centroid Variance, Cluster Weighted Variance, and Davies–Bouldin Index. Cubic spline smoothing identified age intervals of concentrated demand. Among 146 high-frequency sex-diagnosis profiles, Euclidean distance, Dynamic Time Warping with a window constraint, and Shape-Based Distance produced the most compact, well-separated clusters, highlighting the value of constrained temporal alignment. Eight clinically meaningful clusters formed early-onset, working-age, and late-onset regimes, with demand peaking in early childhood and older adulthood. Sex stratification revealed systematic disparities, including higher anemia and asthma utilization among young females. Requiring only age, sex, and ICD-10 codes, the framework provides a transparent decision-support tool for preventive planning, capacity design, and insurance risk modeling.
This paper proposes PRO-MAMCA, a probabilistic PROMETHEE-based extension of Multi-Actor Multi-Criteria Analysis (MAMCA) for nested multi-actor decision support under preference uncertainty. The framework represents decision makers as nested within actor groups and separates three computational layers: individual probabilistic PROMETHEE evaluation, intra-group consensus diagnosis through PRO-GDM, and inter-group compromise analysis. Criteria weights are elicited using the Simos-Roy-Figueira procedure and represented through compatible weight spaces rather than fixed point estimates. Uncertainty in performances, PROMETHEE preference-function types, and preference-function parameters is propagated through Monte Carlo simulation to obtain distributions of flows, rankings, and acceptability indices. Within actor groups, PRO-GDM combines support, agreement, inverse-optimization feasibility, and adjustment effort to assess whether a coherent group position can be justified. Across actor groups, PRO-MAMCA evaluates compromise through average and weakest-group support, opposition risk, balance, consensus-adjusted acceptability, and maximum regret. Decision-maker and actor-group weights are treated as explicit modeling choices. A reproducible last-mile delivery case shows that the leading compromise remains stable across alternative preference-function and weighting specifications, while the framework reveals substantial differences in the distribution and reliability of support across actor groups. PRO-MAMCA therefore supports deliberation by indicating not only which alternatives perform well, but also why they are or are not collectively acceptable.
Survey-based evaluations inform high-stakes ranking and performance decisions, yet scores often reflect non-instructional influences. This paper introduces PROMSET (Preference Ranking Organization Method for Student Evaluation of Teaching), a bias-mitigated outranking method that integrates robust multiple linear regression for bias identification with a comparison-dependent extension of PROMETHEE II. The method is validated on Student Evaluation of Teaching (SET) data from a Spanish university (45,357 surveys; 879 course evaluations; four knowledge areas). PROMSET incorporates a comparison-dependent activation mechanism that applies bias-calibrated indifference thresholds only when pairwise comparisons involve bias-favored alternatives. The results show that grade leniency bias and professor gender bias are no longer statistically detectable at conventional significance levels (p > 0.05) under PROMSET, while the discriminating power of the ranking increases substantially. Benchmarks against the traditional method and PROMETHEE II confirm that bias remains statistically significant under conventional approaches but not under PROMSET. PROMSET improves the informational quality and fairness of survey-based rankings for institutional decision-making.
