Predicting readmission risk for diabetic patients: Make artificial intelligence work in real life with interpretable machine learning
Background: The expenditures of healthcare services associated with unplanned readmissions are enormous. Recognising the reasons that contribute to readmission and identifying at-risk patients are the essential steps to reduce such readmissions. Artificial intelligence is changing the practice of healthcare. It has enabled medical practitioners to provide high-quality treatments to reduce readmissions. While it is essential to employ such solutions, making them transparent to medical experts is more critical.
Aims: Apparently, healthcare stakeholders do not have a strong data science background. Doctors consider the cause of a prognosis rather than the binary outcome from the machine learning methods. Most of the previous work presented predictions, but they did not explain the models. Treating these models as “black boxes” diminishes confidence in their predictions. This study aims to report high-performing predictive models for readmission along with transparent interpretations.
Methods: Six machine learning applications are employed in predicting 30-day readmission after hospitalisation for diabetic patients. This study also utilised the white-box machine learning framework by exploring Shapley Additive Explanations and Local Interpretable Model-agnostic Explanations.
Results: Results show that the CatBoost had the best performance with a higher area under the receiver operating characteristic curve than other models. Prior readmission, discharged at home, the number of emergencies and age were strong predictors. To demonstrated explainability at the individual level, we interpreted the relative variable influence of patient observations.
Conclusions: The findings could be helpful in medical practice and provide valuable recommendations to stakeholders for minimising readmission and reducing public healthcare costs in the future.