Trustworthy AI for Predictive Analytics under Missing Data

Reliable predictive analytics in healthcare and other data-driven domains requires machine learning models that remain reliable in the presence of incomplete data. Missing values in electronic health records can introduce bias and reduce the robustness of predictive models. This paper proposes an adaptive framework for predictive analytics under missing data that discovers latent missing-data regimes using unsupervised learning. Statistical features describing missingness structures are extracted and clustered to identify distinct regimes, which are then used to guide model selection. Experiments on the UCI Heart Disease dataset under controlled missingness scenarios demonstrate that the proposed approach maintains reliable predictive performance under varying levels of data incompleteness. The framework contributes to the development of trustworthy AI systems capable of robust decision-making when missing data are present.