Metric selection and promotional language in health artificial intelligence

Background Recent advances in Artificial intelligence (AI) have the potential to substantially improve healthcare across clinical areas. However, there are concerns health AI research may overstate the utility of newly developed systems and that certain metrics for measuring AI system performance may lead to an overly optimistic interpretation of research results. The current study aims to evaluate the relationship between researcher choice of AI performance metric and promotional language use in published abstracts. Methods and findings This cross-sectional study evaluated the relationship between promotional language and use of composite performance metrics (AUC or F1). A total of 1200 randomly sampled health AI abstracts drawn from PubMed were evaluated for metric selection and promotional language rates. Promotional language evaluation was accomplished through the development of a customized machine learning system that identifies promotional claims in abstracts describing the results of health AI system development. The language classification system was trained with an annotated dataset of 922 sentences. Collected sentences were annotated by two raters for evidence of promotional language. The annotators achieved 94.5% agreement ({kappa} = 0.825). Several candidate models were evaluated and, the bagged classification and regression tree (CART) achieved the highest performance at Precision = 0.92 and Recall = 0.89. The final model was used to classify individual sentences in a sample of 1200 abstracts, and a quasi-Poisson framework was used to assess the relationship between metric selection and promotional language rates. The results indicate that use of AUC predicts a 12% increase (95% CI: 5-19%, p = 0.00104) in abstract promotional language rates and that use of F1 predicts a 16% increase (95% CI: 4% to 30%, p = 0. 00996). Conclusions Clinical trials evaluating spin, hype, or overstatement have found that the observed magnitude of increase is sufficient to induce misinterpretation of findings in researchers and clinicians. These results suggest that efforts to address hype in health AI need to attend to both underlying research methods and language choice.