Evaluating the Accuracy of AI Models in Clinical Environments: A Study of EHR-Based Patient Readmission Predictions

Abstract

The work is a personal study that assesses the performance of AI systems in predicting patient readmission rates in healthcare settings, with Electronic Health Records (EHR). With healthcare practice systems transferring to AI-based clinical decision-making technologies, it is clear that the evaluation of system performance and shortcomings cannot be neglected. This paper has two objectives, the first one is to ascertain the performance of machine learning models in predicting patient readmission, and the second is to understand some circumstances that influence real-world model performance in medical practice. Complexity of sample Cosmically large sample of EHR data was exploited to run many types of AI algorithms, e.g. Random Forests, Support Vector Machines, and Neural Networks. The model performance was evaluated with the help of such measurement as accuracy, precision, recall, F1-score and compared to the existing benchmarks. The results indicate that AI models hold the potential to increase the accuracy of the prediction but issues associated with poor data quality, explainability, and generalization of models are still there. The overall results outlined in the above support the requirement of integrating domain expertise along with machine learning tools towards enhancing clinical AI models. The study ends by outlining implications of the adoption of AI in healthcare, suggesting avenues in terms of future research on the same in relation to enhancing the models with regards to transparency and robustness.

Key-words: machine learning, artificial  intelligence, patient readmission, Electronic Health Records, the accuracy of the models, healthcare, predictive analyses