AI-enhanced Decision Support System

• AI-enhanced Decision Support Systems are frameworks that integrate domain knowledge, ML models, and explainable outputs to enhance decision-making in sectors like healthcare, manufacturing, and business intelligence.
• They employ modular layers—data ingestion, domain rule integration, ML-based inference, and explainability tools such as LIME and SHAP—to deliver reliable, regulatory-aligned insights.
• These systems continuously update using digital twins, hybrid metaheuristics, and case-based reasoning to improve performance, trust, and operational adaptivity in high-stakes environments.

AI-enhanced Decision Support System (DSS) refers to a class of computational frameworks that augment traditional decision-making processes across domains—healthcare, manufacturing, business intelligence, industrial optimization—by embedding artificial intelligence methods into data ingestion, modeling, inference, and explanation layers. These systems tightly couple curated domain knowledge, algorithmic learning, and explainable output, enabling higher reliability, personalization, and adaptive updating compared to non-AI DSS architectures.

1. System Architectures: Core Layers and Patterns

AI-enhanced DSS platforms typically organize processing into interlocking modular layers:

• Data Ingestion and Feature Engine: Aggregates multi-modal sources such as structured EHR, sensors, laboratory results, environmental parameters, and domain-specific entities. For example, the Digital Twin DSS in clinical settings standardizes a ten-dimensional vector from EHR and continuous sensor inputs as $$x = [\text{Age}, \text{Gender}, \ldots, \text{A/G Ratio}]^\top$$ (Rao et al., 2019). Manufacturing systems collect Cp, Cpk, NCR, and ENCR from the production floor (Oukhay et al., 2020). Marketing DSS fuse social media, financial, and behavioral logs with text embeddings or sentiment scores (Yin et al., 13 Nov 2025).
• Domain Knowledge Module: Encodes expert heuristics and regulatory rules, often using decision tables, threshold functions, or DMN (Decision Model and Notation) tables. These are instantiated either as explicit if–then rules or learned from expert inputs, yielding high baseline safety and regulatory alignment (Rao et al., 2019, Kovalchuk et al., 2020).
• Artificial Intelligence/ML Component: Implements supervised or hybrid models such as random forests, gradient boosting, SVMs, deep neural networks, graph neural networks, or personalized tree/rule ensembles. Training objectives vary—cross-entropy loss, regularized MSE, or metaheuristic optimization (e.g., for resource planning)—but consistently leverage large labeled datasets and cross-validation for model selection and hyperparameter tuning (Rao et al., 2019, Kovalchuk et al., 2020, Yin et al., 13 Nov 2025, Kumar et al., 2012, Rahimi, 23 Nov 2025).
• Explainability and Interpretation Layer: Agents such as LIME, PDP, SHAP, SkopeRules, and CLMs produce human-centric explanations for black-box predictions, joining both local feature attribution and counterfactual/visual “reason” generation. LLMs are now sometimes used to deliver these outputs in domain-adapted prose (Rao et al., 2019, 2503.06463, Lucieri et al., 2020).
• Software Orchestration & Continuous Updating: Real-time orchestration routes data through the three main layers (knowledge, ML, explanation), logs outputs, and (in digital twin frameworks) perpetually adapts domain boundaries/rules as update-worthy patterns emerge with new data (Rao et al., 2019, Yin et al., 13 Nov 2025).

2. Machine Learning Models and Integration with Domain Knowledge

AI-enhanced DSSs advance over purely rule-based or statistical systems by integrating learned models that generalize across heterogeneous expert knowledge and patient/user variation.

• In healthcare digital twins, a random forest classifier is trained to mimic aggregated physician labeling: $f(x;\theta) = P(y=1|x;\theta)$, with loss function $$L(\theta) = -\frac{1}{N} \sum [y_i \log p_i + (1-y_i) \log(1-p_i)] + \lambda \|\theta\|^2$$, where $p_i=f(x_i;\theta)$ (Rao et al., 2019).
• On industrial lines, recommender DSSs dynamically choose between multi-criteria optimization—using Analytic Hierarchy Process weights $w_i$ and Choquet integrals $C_\mu(v_k)$—and case-based reasoning via nearest-neighbor retrieval on quality-indicator vectors, continuously updating a scenario base for adaptation (Oukhay et al., 2020).
• Marketing DSS architecture applies dual-channel hybrid deep learning: GNNs structurally encode user-content-ROI networks while Temporal Transformers capture engagement over time; causal inference modules simulate counterfactual campaigns to optimize spend and forecast market growth (Yin et al., 13 Nov 2025).
• Benchmarking studies show that choice of classifier impacts system performance on trade-off axes: SVMs offer rapid, high accuracy; genetic algorithms achieve maximum accuracy and interpretable rules given longer training times; decision trees (e.g., QUEST, C4.5) excel in comprehensibility and speed for certain data regimes (Kumar et al., 2012).

Table: Example Model Integration Patterns

Domain	Knowledge Integration	ML Component	Explanation Layer
Clinical	DMN tables from guidelines	Random Forest (RF), DNN, GBM	LIME, PDP, SHAP
Manufacturing	AHP + Choquet + CBR	Nearest-case retrieval, ensemble	Scenario rationale
Marketing	ROI formulas, causal rules	GNN+Transformer, Causal module	Attention mapping, ACE
Heritage	Material damage functions, ISO	PyCaret ensemble, expert rules	Justification snippets

3. Explainability, Trust, and Human-Centered Output

Transparency in AI-driven DSS is operationalized through formalized, multi-perspective explanation modules and empirically tested frameworks for trust and calibration.

• Local Surrogates and Feature Attributions: LIME fits sparse linear models $g(z) = w^\top z$ around a target $x_0$ by optimizing fidelity loss $L(f,g,\pi_{x_0}) + \Omega(g)$. This reveals the lab values or features contributing most to a risk prediction (Rao et al., 2019). PDPs and SHAP values afford both local and global decomposability.
• Spatial Concept Localization: In imaging, CLMs map concept activations to spatial domains via gradient or perturbation methods, supporting experts in verifying that the DSS recognizes semantically meaningful regions (“biomarker” on synthetic or facial features on real images) (Lucieri et al., 2020).
• Causal and Counterfactual Explanations: Causal inference layers provide intervention/ACE estimates (e.g., ROI under changed ad spend), while counterfactual generators highlight minimal changes to flip DSS outputs (RCC explanation paradigm) (Dorsch et al., 23 Sep 2024, Yin et al., 13 Nov 2025, 2503.06463).
• Trust Calibration and User Acceptance: The Multisource AI Scorecard Table (MAST) evaluates DSS transparency and user trust across nine dimensions—data provenance, uncertainty quantification, distinguishing, analysis of alternatives, and logical reasoning—predicting trust and perceived credibility but not always improving joint human+AI accuracy (Salehi et al., 2023). Interaction studies confirm that DSS recommendations increase decision correctness and are most useful when designed to prompt critical, evaluative reflection rather than passive acceptance (Mastrianni et al., 17 May 2025, Yin et al., 13 Nov 2025).

4. Continuous Learning, Adaptivity, and Optimization

Advanced DSS platforms emphasize closed-loop adaptation, operationalized at both algorithmic and workflow levels:

• Digital Twin Updating: Each patient twin is instantiated, analyzed, and compared between ML and domain decision boundaries. Over time, systemic, consistent ML deviations induce upstream threshold updates (e.g., lowering ALP cutoff), yielding continual improvement (Rao et al., 2019).
• Hybrid Metaheuristics and Scalable Optimization: In resources planning (e.g., open-pit mining), DSSs blend genetic algorithms, large neighborhood search, simulated annealing, and reinforcement learning, under $\varepsilon$-constraint relaxation for tractable, uncertainty-resilient optimization—enabled by deep generative models (VAEs) for scenario sampling and GPU-parallel evaluation for runtime speedups of $10^6$–$10^7$x compared to classical MILP solvers (Rahimi, 23 Nov 2025).
• Case-Based Updating: CBR modules in manufacturing DSSs grow their case base by retention/adaptation, shrinking decision time and improving quality compliance as the experience corpus expands (Oukhay et al., 2020).

5. Empirical Performance, Applications, and Limitations

AI-enhanced DSSs now match or exceed traditional metrics while providing higher-level interpretability and workflow compatibility:

• Healthcare: Digital Twin DSS achieved 72% accuracy, AUC ≈0.75, sensitivity 0.70, specificity 0.74 for liver disease risk, surpassing domain-only rules and reducing false positives by 12% (Rao et al., 2019). Three-stage pipelines improved on FINDRISK T2DM detection sensitivity and specificity, with focus on UI compatibility and explainability (Kovalchuk et al., 2020).
• Benchmarking Decision Algorithms: SVMs or DNNs achieved perfect accuracy on simple domains; genetic algorithms led on more complex, multiclass data; decision trees maximized comprehensibility. Clear trade-off: speed/accuracy vs. rule transparency (Kumar et al., 2012).
• Marketing: GNN–Transformer DSS produced RMSE=0.063, F1=0.884, ATE Error=0.015—outperforming LSTM, GRU, and TCN baselines; ablation showed each architectural layer substantially contributes to performance/stability (Yin et al., 13 Nov 2025).
• Expert Reviews: User studies confirm increased accuracy and satisfaction with AI-aided DSSs in time-critical medicine (odds ratio OR=1.80 for correctness vs. baseline) but spotlight accuracy–time trade-offs and divergent attitudes toward automated recommendations (Mastrianni et al., 17 May 2025).

Identified limitations: Data scarcity and lack of external validation (clinical), continuous integration costs, explainability granularity (LIME surrogates may not reflect global model), challenge of mining unstructured data (e.g., free-text in oncology), and generalization limits when expanding to new modalities or materials (Rao et al., 2019, Kovalchuk et al., 2020, Grüger et al., 12 Mar 2025, Kuchař et al., 21 May 2025).

6. Design Recommendations and Future Directions

Best practices and forward-looking guidance coalesce around several themes:

• Hybrid Pipelines: Combine deterministic rule bases anchored in regulatory or domain knowledge with ML-driven prediction and XAI-driven interpretation, keeping interfaces compatible with existing workflows (Kovalchuk et al., 2020, 2503.06463).
• Explanations Aligned with Human Reasoning: Emphasize the RCC (Reasons/Counterfactuals/Confidence) triad; avoid anthropomorphizing AI roles and always triangulate uncertainty (Dorsch et al., 23 Sep 2024).
• Transparency and Assessment: Track trust metrics (e.g., via MAST), instrument feedback loops for ongoing explanation quality assessment, and implement retrainable mechanisms for model and domain knowledge updating (Salehi et al., 2023).
• Scalable and Uncertainty-Resilient Computation: Integrate generative models and parallelized architectures (GPUs) for tractable real-time scenario handling in resource-intensive domains (Rahimi, 23 Nov 2025).
• Data Readiness and Documentation: Begin with deep data readiness audits, ensure structured mapping of key decision variables, and develop minimal-burden documentation interfaces to close capture/extractability gaps—equally applicable to clinical, industrial, and heritage preservation domains (Grüger et al., 12 Mar 2025, Kuchař et al., 21 May 2025).
• Modular, Loosely-Coupled Pipelines: Keep predictive, explanation, causal, and adaptation layers modular to allow for targeted upgrades and task-specific tuning (2503.06463).
• Ethical and Regulatory Considerations: Embed fairness (via pre/in/post-processing) and privacy-preserving (DP, FL, encryption) approaches given domain risks (especially in clinical and sensitive market domains) (Alkan et al., 16 Jan 2025).

Future directions include patient- or user-facing explanations, real-time closed-loop adaptation (digital twins and automation), quantification of over-reliance and behavioral economics of DSS usage, and regulated inclusion of validated AI models into clinical and industrial standards (Kovalchuk et al., 2020, Kuchař et al., 21 May 2025, Dorsch et al., 23 Sep 2024).