Algorithm for Assessing the Quality of Medical Synthetic Data Based on The Multi-Criteria and Pac-Bayesian Model

The effectiveness of medical artificial intelligence systems is directly related to the quality and level of representativeness of the training sample. However, real clinical data are characterized by problems such as confidentiality limitations, class mismatch, and heterogeneity. In solving these problems, synthetic data is considered a promising solution, however, a comprehensive assessment of their quality remains a pressing issue. In this work, an algorithm for assessing the quality of synthetic data based on a multi-criteria approach and PAC-Bayesian theory is proposed. The proposed model combines statistical proximity (KL-divergence), inter-character relationships (mutual information), predictive efficiency, and subject-specific limitations within a single integrated functional. Experimental results showed that the synthetic data satisfactorily reflect the main features of the real distribution (KL = 0.6035), and inter-characteristic relationships are maintained with high accuracy (MI = 0.0072). The model, trained on the basis of synthetic data, achieved an accuracy of 80.4%, which confirms the possibility of its practical application. The domain matching criterion showed the maximum value (1.0). The results show that the proposed approach ensures a balance between the quality of synthetic data and model reliability.

Hali tarjima qilinmagan