Explainability-Enhanced Neural Network for Thoracic Diagnosis Improvement

Thoracic problems are medical conditions that affect the area behind the sternum and include the heart, lungs, trachea, bronchi, esophagus and other structures of the respiratory and cardiovascular system. These problems can be caused by a variety of conditions, such as respiratory infections, lung conditions, heart conditions, autoimmune diseases, or anxiety disorders, and can vary in symptoms and severity. In this paper, we introduce a supervised neural network model that is trained to predict these problems and to further more increase the level of accuracy by using explainability methods. We chose to use the attention mechanism to be able to get a higher weight after training the data set. The accuracy of the trained model reached the value of more than 80%. To be able to analyze and explain each feature, we use Local Interpretable Model-Agnostic Explanations, which is a post-hoc model agnostic technique. Our experiments showed that by using explainability results as feedback signal, we were able to increase the accuracy of the base model with more than 20% on a small medical dataset.