Frequency-Aware Attention based LSTM Networks for Cardiovascular Disease

There are various medical features associated with cardiovascular disease in the EMR data, but the frequency of each medical feature is different. Less frequent feature may be considered as non-critical feature, although cardiovascular disease is closely associated in the cardiovascular disease risk prediction model. We propose a frequency-aware based Attention-based LSTM (FA-Attn-LSTM) that weighs on important medical features using an attention mechanism that considers the frequency of each medical feature. Our model predicts the risk for cardiovascular disease using the ejection fraction as a prediction target and shows RMSE = 3.65 and MAE = 2.49.