Source-linked AI summary
Interpretable Synthetic Medical Tabular Data Generation for Clinical Decision Support Using Fuzzy Cognitive Maps
Michael Vasilakakis, Dimitris K. Iakovidis
TL;DR
Synthetic medical tabular generation is framed as a need for clinical interpretability and causal coherence beyond purely statistical density estimation. The paper uses fuzzy feature concepts and FCM reasoning to generate records; across benchmarks, it reports competitive utility, fidelity, and privacy versus baselines under CPU-only constraints.
Problem
The paper frames synthetic medical tabular generation as requiring clinical interpretability and causal coherence beyond purely statistical density estimation.
Source: §III · p. 3Method
The framework represents features with fuzzy concepts and generates records by propagating random linguistic activations into causally stabilized concept states and numerical feature vectors.
Source: §III · p. 3, C. Medical Data Generation · p. 4, B. Results · p. 6Results
Across multiple benchmarks, results are competitive in utility, fidelity, and privacy versus baselines under CPU-only constraints; under TSTR, the method outperforms CTGAN and is competitive with TVAE and Gaussian Copula.
Source: B. Results · pp. 5–6Takeaways & Limitations
Causally grounded fuzzy modeling provides a practical, resource-efficient alternative for synthetic medical tabular data generation.
Source: B. Results · p. 6
Abstract
from arXiv · showhide
Synthetic medical tabular data generation has become essential for developing and validating computer-based medical systems (CBMSs) when real clinical data is restricted due to privacy, ethical, or data availability limitations. Existing probabilistic and deep generative models often lack interpretability and fail to preserve clinically meaningful dependencies, limiting their suitability for safety-critical applications. This paper proposes a novel application of Fuzzy Cognitive Maps (FCMs) in a framework for synthetic medical tabular data generation with explicit causality and privacy preservation. Clinical features are described using linguistically interpretable fuzzy sets, and inter-feature dependencies are encoded as FCM edge weights computed from fuzzy set intersections. Synthetic patient records are generated by propagating randomly initialized linguistic activation vectors through the FCM until convergence, followed by defuzzification to produce clinically coherent numerical values. The approach natively handles mixed data types, and domain constraints common in health records. Experimental evaluation on UCI medical benchmark datasets demonstrates competitive performance under a Train-on-Synthetic-Test-on-Real (TSTR) protocol. The proposed method achieves accuracy of up to 0.81 and AUROC of up to 0.90 on the Heart Disease dataset, matching or exceeding TVAE and Gaussian Copula baselines while running exclusively on CPU. Fidelity metrics including KS Complement (up to 0.91) and Correlation Similarity (up to 0.95) confirm strong statistical coherence, and DCR Baseline Protection scores consistently exceed those of TVAE, confirming adequate privacy guarantees. These results demonstrate that causally grounded, interpretable fuzzy modeling offers a computationally efficient and transparent alternative to deep generative models for trustworthy synthetic data generation in CBMSs.
II. FUZZY COGNITIVE MAPS
An FCM represents concepts as a directed graph whose weighted edges encode causal influence. It reasons by iteratively updating concept activations toward a steady state.
II. FUZZY COGNITIVE MAPS
- FCMs are directed graphs of concepts connected by weighted edges that represent the strength and direction of causal influence.Edges may express positive, negative, or neutral relationships between concepts.Source: §II · pp. 2–3
- FCM reasoning iteratively updates concept activations through a sigmoid-based formulation until a steady state, whose output activations are interpreted as inferred outcomes.Convergence is not guaranteed in general.Source: §II · p. 3
- The initial state vector represents the concepts’ initial activation values at iteration t=0.Source: §II · p. 3
III. MEDICAL DATA GENERATION FRAMEWORK
The framework treats synthetic medical tabular data generation as clinically interpretable, causally coherent reasoning rather than purely statistical density estimation.
III. MEDICAL DATA GENERATION FRAMEWORK
- The framework frames synthetic medical data generation as a clinically interpretable and causally coherent reasoning process rather than purely statistical density estimation.Source: §III · p. 3
- It combines semantic feature abstraction, causal dependency modeling, and expert-guided reasoning to generate synthetic patient records.Source: §III · p. 3
- The framework is designed to preserve clinical plausibility, interpretability, and privacy in generated records.Source: §III · p. 3
A. Fuzzy Set Construction for Medical Data Presentation
The framework represents medical features with data-derived fuzzy sets that express interpretable, relative clinical states and support subsequent causal reasoning.
A. Fuzzy Set Construction for Medical Data Presentation
- Fuzzy sets act as feature-specific linguistic concept nodes for relative states, not fixed clinical labels, and clinicians may relabel, merge, or adjust them.The sets are derived from data and aligned with medical semantics.Source: A. Fuzzy Set Construction for Medical… · pp. 3–4
- Features are normalized to [0,1] and analyzed by class using clustering, usually fuzzy c-means, to identify representative value regions.Source: A. Fuzzy Set Construction for Medical… · p. 3
- Triangular membership functions are centered on cluster values and extend toward neighboring centers, with boundary cases assigned trapezoidal functions to support gradual transitions.Membership degrees provide a soft association between measurements and linguistic states.Source: A. Fuzzy Set Construction for Medical… · p. 3
- Fuzzy cardinality aggregates class-specific membership degrees to estimate how representative each linguistic state is, ranking dominant and marginal states and giving high states greater influence in later reasoning.Source: A. Fuzzy Set Construction for Medical… · pp. 3–4
- The illustration places normalized medical values on the horizontal axis and membership degree on the vertical axis; cross-feature overlap visualizes potential co-activation, not jointly defined membership functions.Features share a normalized domain for overlap analysis, while their fuzzy sets are defined independently.Source: A. Fuzzy Set Construction for Medical… · pp. 3–4
B. Definition of FCM Interconnections
The framework estimates interpretable FCM connections from overlaps between fuzzy feature representations, then allows clinicians to review and refine the resulting graph.
B. Definition of FCM Interconnections
- Directed FCM edges represent clinically interpretable influence between medical features, with initial weights estimated from interactions between their fuzzy representations.Source: B. Definition of FCM Interconnections · p. 4
- The influence between feature concepts is defined through fuzzy-set intersection, whose co-activation strength informs directed weights computed using fuzzy center of gravity and normalized to [−1,1].Source: B. Definition of FCM Interconnections · p. 4
- Clinicians can validate dependencies, remove spurious associations, enforce known causal directions, and constrain edge signs to refine the FCM into a medical causal representation.Source: B. Definition of FCM Interconnections · p. 4
C. Medical Data Generation
Synthetic records begin as randomly initialized linguistic activations, which are propagated through the FCM to a stable state and defuzzified into numerical feature values. Experts can inspect and revise the process to enforce clinical rules.
C. Medical Data Generation
- Randomly initialized linguistic feature states are propagated through the FCM to produce synthetic samples with causal consistency.The activation vector is initialized in fuzzy concept space, then iteratively propagated until convergence.Source: C. Medical Data Generation · p. 4
- At convergence, the most activated fuzzy set guides a crisp value sampled from its support, preserving controlled variation across plausible values.The method sets a maximum of 100 iterations; when multiple fuzzy sets are activated, it uses a weighted combination of their centers.Source: C. Medical Data Generation · pp. 4–5
- Experts can revise concept values and edge weights, enforce clinical rules, and reject implausible stabilized states before defuzzification.The paper describes this as a human-in-the-loop capability for aligning samples with established medical knowledge.Source: C. Medical Data Generation · p. 5
A. Experimental Setup
The evaluation compares the proposed method with representative tabular synthesizers on three UCI medical datasets, using CPU-only execution and a consistent train–test split.
A. Experimental Setup
- Experiments used CPU-only execution to compare the proposed method with Gaussian Copula, CTGAN, and TVAE under comparable computational conditions.The baseline models used default SDV configurations, and each generated as many samples as the real training set contained.Source: A. Experimental Setup · p. 5
- The benchmarks were Pima Indians Diabetes, South African Heart Disease, and Statlog Heart Disease, with 768, 462, and 270 instances, respectively.The datasets contain mixed continuous and categorical features.Source: A. Experimental Setup · p. 5
- Each dataset used an 80/20 train–test split, with synthetic data generated from training data and evaluation on the real test set.This setup assesses generalization and downstream utility.Source: A. Experimental Setup · p. 5
B. Results
Under TSTR evaluation, the FCM framework is competitive with TVAE and Gaussian Copula and outperforms CTGAN across utility measures, while balancing fidelity, privacy, and computational efficiency. The conclusion identifies temporal data and joint-distribution metrics as future work.
B. Results
- Evaluation followed TSTR, measuring classification utility, marginal and inter-feature fidelity, and privacy risk on real test data.The metrics include Accuracy, F1-score, AUROC, KS Complement, Correlation Similarity, Statistical Similarity, and DCR Baseline Protection.Source: B. Results · p. 5
- The FCM framework outperforms CTGAN and remains competitive with TVAE and Gaussian Copula on TSTR utility, while maintaining lower computational complexity.On Heart Disease, its AUROC is nearly identical to TVAE’s under CPU-only constraints; the reported comparison is summarized in Table I.Source: Table I · p. 5, B. Results · pp. 5–6
- Fidelity results show high KS Complement and correlation similarity, although Gaussian Copula has marginally higher fidelity in some cases alongside reduced predictive utility.The framework is reported to balance realism with downstream task performance.Source: B. Results · p. 6
- DCR Baseline Protection consistently exceeds TVAE’s, while sparse linguistic activations and FCM propagation support efficient inference.Higher DCR scores indicate greater distance from real records and lower re-identification risk; only a limited subset of concepts becomes active during reasoning.Source: B. Results · p. 6
- The conclusion describes fuzzy-set feature marginals and sparse FCM dependencies as an interpretable, computationally efficient alternative to conventional density estimation.Records transform random linguistic profiles into causally stabilized concept states and then numerical feature vectors.Source: B. Results · p. 6
- Across benchmark datasets, the conclusion reports competitive utility, fidelity, and privacy under CPU-only constraints and identifies temporal data and joint-feature metrics as future work.The paper also proposes automated structure refinement as a future direction.Source: B. Results · p. 6