Predictive Medicine Using Interpretable Recurrent Neural Networks Andr´ e Crist´ ov˜ ao Neves Ferreira andre.c.n.ferreira@tecnico.ulisboa.pt Instituto Superior T´ ecnico, Lisboa, Portugal October 2020 Abstract Deep learning has been revolutionizing multiple aspects of our daily lives, thanks to its state-of- the-art results. However, the complexity of its models and its associated difficulty to interpret its results has prevented it from being widely adopted in healthcare systems. This represents a missed opportunity, specially considering the growing volumes of Electronic Health Record (EHR) data, as hospitals and clinics increasingly collect information in digital databases. While there are studies addressing artificial neural networks applied to this type of data, the interpretability component tends to be approached lightly or even disregarded. Here we demonstrate the superior capability of recurrent neural network based models, outperforming multiple baselines with an average of 0.94 test AUC, when predicting the use of non-invasive ventilation by Amyotrophic Lateral Sclerosis (ALS) patients, while also presenting a comprehensive explainability solution. In order to interpret these complex, recurrent algorithms, the robust SHAP package was adapted, as well as a new instance importance score was defined, to highlight the effect of feature values and time series samples in the output, respectively. These concepts were then combined in a dashboard, which serves as a proof of concept in terms of a AI-enhanced detailed analysis tool for medical staff. Keywords: Deep learning, interpretability, recurrent neural network, electronic health records, disease progression, data visualization 1. Introduction Through deep learning models, academia and in- dustry alike have disrupted a wide variety of areas. However, compared to previous machine learning models, these high-performing yet more complex deep learning models are less intuitive, in terms of interpreting their outputs. This observation started a performance and interpretability tradeoff, as while in some cases one might desire accuracy above all, in other, more critical scenarios, it is also very important to validate and understand how the model gets to each result. One case where interpretability matters particu- lary is in healthcare. When a decision can define re- covery or deteorating health, life or death, any error can result in serious consequences. So, each decision must be carefully thought of, neatly planned out and made with thorough understanding of the situ- ation. Medics cannot afford to just blindly trust an algorithm, no matter how good it claims to be. This difficult interpretability of deep learning models is likely why AI has not yet been massively integrated into healthcare systems worldwide. Traditional ma- chine learning models, with relatively low accuracy, do not represent enough value, while deep learning cannot be trusted. Towards reaching an AI approach that can be considered for integration in a healthcare system, several steps need to be taken. This work hopes to push in that direction, by contributing in key areas and showcasing the potential of it. To do so, the main objectives have been defined as: • Training machine learning models that excel in the prediction of disease progression, based on Electronic Health Records (EHR) data. • Find or develop an adequate interpretability technique to allow for model validation and output explainability. • Create a prototype of a platform that can allow for intuitive interaction with the trained mod- els, gathering insights from it and interpreting its outcomes. 2. Related Work Over the last few years, there has been an increas- ing number of papers published on deep learning applied to EHR data [4, 3, 7, 13, 16, 1, 14]. These 1