Poster: Using Deep Attention Multiple Instance Learning To Improve Malpractices Detection In The Health Sector

In this study we developed a machine learning model to quantify the risk of malpractice based on historical medical claims data. The results are applied to improve the detection of malpractices in the context of the Catalan healthcare system. The system considers data about the medical facility, clinical staff, patient, law, and environmental conditions.
Our model is based on a supervised learning framework, where real claim data is available in tabular form and tagged with a binary label indicating risk or no risk. However, differently from traditional supervised learning tasks, in our setting each claim consists of a variable number of rows accounting for different instances of e.g. treatments or facilities involved.
To overcome that challenge, we propose a risk prediction method based on the state-of-the-art Deep Attention Multiple Instance Learning (DAMIL) model, specifically designed for healthcare applications, with a primary focus on predicting and preventing malpractice. The model is capable of assigning a risk probability to each claim, but also incorporates an explainability mechanism to indicate, when possible, to what degree each individual instance in the claim weighs in the calculation of the risk estimate. This enhancement is geared towards improving accuracy, interpretability, and overall applicability of AI algorithms in complex medical scenarios and opens interesting directions for the application of artificial intelligence in the interest of a safer and more efficient healthcare system.
Furthermore, we analize the impact of predicting risk in medical claims using our model compared to the traditional assessment from experts. We prove that our model is more efficient than expert classification of claims, suggesting that economic postive impact can be derived from automatically predicting the claims with medical malpractice.