Early Diagnosis Prediction with Recurrent Neural Networks

Abstract

Predicting medical diagnoses early is critical as it can improve treatment outcomes and ultimately save patient lives. Machine learning can help doctors make early predictions by leveraging an abundance of electronic health data. However, medical data is difficult to feed into predictive models because it typically has a large number of missing values. In this work, we propose an early diagnosis prediction model. First, we design an LSTM (Long Short-Term Memory), a type of neural network effective for modeling long time series. Second, to address the issue of missing data values, we implement various data imputation techniques and evaluate their effectiveness when used with the LSTM. Finally, we develop a novel LSTM model named Multi-Label Early Detection (MED), which has the goal of predicting patient diagnoses early in their hospital stay. We compare MED to state-of-the-art baselines using a subset of time series from the MIMIC-III database. We verify that our model obtains comparable AUC scores to that of standard LSTMs while encouraging early predictions.