Magnitude and Impact of Hallucinations in Tabular Synthetic Health Data on Prognostic Machine Learning Models: Validation Study.

Abstract

Generative artificial intelligence (AI) for tabular synthetic data generation (SDG) has significant potential to accelerate health care research and innovation. A critical limitation of generative AI, however, is hallucinations. Although this has been commonly observed in text-generating models, it may also occur in tabular SDG. This study aims to investigate the magnitude of hallucinations in tabular synthetic data, whether their frequency increases with training data complexity, and the extent to which they impact the utility of synthetic data for downstream prognostic machine learning (ML) modeling tasks. On the basis of 12 large and high-dimensional real-world health care datasets, 6354 training datasets of different complexity were created by varying the subset of variables included in each dataset. Synthetic data were generated using 7 different SDG models. Hallucinations were defined as synthetic records that did not exist in the population, and the hallucination rate (HR) was the proportion of hallucinations in a synthetic dataset. Classification was the downstream prognostic modeling task, conducted via an ML approach (light gradient boosted machine) and an artificial neural network (multilayer perceptron). Mixed-effects models were fitted to examine the relationship between training data complexity and the HR and the HR and the predictive performance of AI and ML models when trained on the synthetic data. The HR ranged from 0.3% to 100% (median 99.1%, IQR 98.5%-100.0%) and increased with training data complexity. However, in most SDG models, the HR did not affect AI and ML prognostic model performance. In the SDG models in which a significant association was detected, the estimated effect was very small, with a maximum decrease in the area under the receiver operating characteristic curve of -0.0002 (95% CI -0.0003 to -0.0002, P<.001) in light gradient boosting machine and -0.0001 (95% CI -0.0002 to -0.0001, P=.002) in multilayer perceptron. These findings suggest that while hallucinations may be very common in synthetic tabular health data, they do not necessarily impair its utility for prognostic modeling.