With the continuous improvement of the accuracy of predictive models, the problems of algorithmic bias and the lack of fairness have become increasingly prominent. This not only may lead to significant differences in the accessibility of medical services and the effectiveness of diagnosis and treatment among different groups, but also may trigger medical risks such as misdiagnosis and missed diagnosis, treatment delays, as well as ethical disputes and legal issues under algorithmic discrimination, becoming the key reasons restricting the implementation of technology. At present, research on fairness has shifted from ethical and conceptual discussions to quantitative calculations, exploration of bias layer mechanisms, and exploration of technical mitigation strategies. This article, through a comprehensive analysis and discussion of the fairness issue of clinical prediction models, systematically sorts out the sources of model unfairness, fairness assessment and mitigation methods, as well as commonly used toolkits, and at the same time explores the future development trends. This study aims to sort out the current status of fairness studies on predictive models and promote the responsible application of artificial intelligence in the medical field.
ObjectiveTo introduce and validate the censoring unbiased deep learning (CUDL) algorithm suitable for survival data and to explore its utility in survival analysis. MethodsThe principles of the CUDL algorithm were detailed, followed by an introduction to its software implementation steps and parameter settings. Using the construction of a long-term survival prognosis model for patients with myocardial infarction (MI) as a case study, the algorithms were compared with the conventional Cox proportional hazards model and random survival forest model. ResultsIn the task of predicting the 5-year survival probability of patients, the mean Brier score (BS) of both the Buckley-James censoring unbiased deep learning (BJDL) model and the regularized BJDL were superior to that of the Cox proportional hazards model (mean BS: 0.1722 and 0.1715 vs. 0.1765, P<0.001). The regularized BJDL also outperformed the random survival forest model (RSF500) (mean BS: 0.1715 vs. 0.1750, P<0.001). ConclusionThe BJDL can effectively leverage censored information within survival data to achieve accurate predictions without the need to satisfy the proportional hazards assumption. It provides a novel methodological alternative for the analysis of medical survival data.