Feature Attribution for Automatic Medical Coding

Assigning medical codes to clinical reports is an important task in the healthcare domain, primarily for documentation, billing, and research purposes. Multipage reports and a high-dimensional label space cause the manual mapping process to be time-consuming and error-prone. While deep neural networks have recently achieved promising performance in automatic medical coding, the opaque decisionmaking process of these models hampers their adoption in real-world scenarios. This points to the need for explanations revealing the reasons behind model predictions. In this master’s thesis, we address a specific type of explanations, namely feature attributions for text classification models. These explanations indicate which words in the input are relevant for the model to produce a particular output. Several existing models for automatic medical coding provide such feature attributions. However, they all lack an evaluation of faithfulness, i.e. of how accurately the attributions reflect the model’s reasoning process. In high-stakes decision scenarios such as medical coding, the faithfulness of attributions is crucial to account for the legal, financial, and ethical relevance of the code assignment. The contribution of this thesis is threefold: First, we analyze recent literature on feature attribution to examine the theoretical adequacy of the most commonly used methods with respect to our use case. Second, we investigate in experiments the applicability of a popular gradient-based attribution method, called Integrated Gradients, to two medical coding models. Third, we discuss the suitability of a recent faithfulness metric, called infidelity, from a theoretical perspective, and scrutinize its applicability to the medical coding task in experiments. Our attribution results indicate that Integrated Gradients cannot be readily applied to the current state-of-the art medical coding model, implying the need for further research in this area. Our discussion of infidelity suggests that this metric represents a reasonable notion of faithfulness. Our experiments appear to confirm its suitability for the medical coding task. Based on these results, we believe that this metric has the potential to become a standard tool for evaluating the faithfulness of feature attributions.