Actionable Suggestions in Support of Rehospitalization Risk Predicted by Artificial Intelligence

Abstract

Discharging a patient from hospital is a crucial decision, as wrong timing may lead to rehospitalization. In this scenario, Artificial Intelligence (AI) can support physicians in the decision-making process. However, AI provides a probability that an event will happen, without information regarding how to reduce such probability. EXplainable Artificial Intelligence (XAI) methods, allow to better understand how the score is computed but still lack in providing specific suggestions. We propose a method that creates a set of actionable suggestions on how to change the feature values to reduce a risk predicted by AI, and support the physicians in deciding when to discharge a patient. We utilize historical data distribution of features to understand how values for the current case should be modified. After retrieving a set of relevant historical data by considering the similarity with the current point, we use statistical analysis to highlight differences between the feature values of current and historical cases. A suggestion is made whenever a significative difference on a controllable feature is found. For the evaluation, we used an electronic health record (MIMIC-III) to produce suggestions to reduce 30 days rehospitalization probability. A ground truth set of suggestions was created for 54 patients, by considering how changes in feature values affected the risk, as well as if feature values fell within literature ranges. The proposed approach reached a precision of 80%. The adoption of this method will allow to decrease costs related to rehospitalizations and positively impact patients’ health.

Ähnliche Arbeiten