Towards Trustworthy Explanations in Clinical AI: A Framework of Causal Screening and Clinical Constraints

The importance of artificial intelligence continues to increase in disease diagnosis and risk prediction. However, the clinically used prediction models based on AI nowadays are often established upon non- causal features, limiting their interpretability and trustworthiness among doctors. To address this issue, the Causal-Clinical Explainability (CCX) framework is put forward in this paper. In addition to the use of clinical prior knowledge for the purpose of guiding the selection of features, the framework also carries out causal discovery via the PC method for the elimination of wrongful associations. Through the double strategy mentioned above, the quality of the causal-clinical feature subsets for the establishment of any following prediction models can be ensured. Experiment results demonstrate that the CCX framework outperforms baseline models both on prediction performance and robustness. The paper offers an effective approach for the development of clinical decision support systems and provides a feasible solution for the promotion of the usage of AI for clinical work under practical situations.