An interpretable risk prediction model for healthcare with pattern attention

Abstract

Background: The availability of massive amount of data enables the possibility of clinical predictive tasks. Deep learning methods have achieved promising performance on the tasks. However, most existing methods suffer from three limitations: (1) There are lots of missing value for real value events, many methods impute the missing value and then train their models based on the imputed values, which may introduce imputation bias. The models’ performance is highly dependent on the imputation accuracy. (2) Lots of existing studies just take Boolean value medical events (e.g. diagnosis code) as inputs, but ignore real value medical events (e.g., lab tests and vital signs), which are more important for acute disease (e.g., sepsis) and mortality prediction. (3) Existing interpretable models can illustrate which medical events are conducive to the output results, but are not able to give contributions of patterns among medical events. Methods: In this study, we propose a novel interpretable Pattern Attention model with Value Embedding (PAVE) to predict the risks of certain diseases. PAVE takes the embedding of various medical events, their values and the corresponding occurring time as inputs, leverage self-attention mechanism to attend to meaningful patterns among medical events for risk prediction tasks. Because only the observed values are embedded into vectors, we don’t need to impute the missing values and thus avoids the imputations bias. Moreover, the self-attention mechanism is helpful for the model interpretability, which means the proposed model can output which patterns cause high risks. Results: We conduct sepsis onset prediction and mortality prediction experiments on a publicly available dataset MIMIC-III and our proprietary EHR dataset. The experimental results show that PAVE outperforms existing models. Moreover, by analyzing the self-attention weights, our model outputs meaningful medical event patterns related to mortality. Conclusions: PAVE learns effective medical event representation by incorporating the values and occurring time, which can improve the risk prediction performance. Moreover, the presented self-attention mechanism can not only capture patients’ health state information, but also output the contributions of various medical event patterns, which pave the way for interpretable clinical risk predictions. Availability: The code for this paper is available at: https://github.com/yinchangchang/PAVE.