Prediction of Disease Progress Using Medical Visit Records with Irregular Time Intervals

Time irregularity is common in medical visit record data. For disease progress prediction, irregular time intervals occur in two places. One is in the input historical data for prediction (denoted as input intervals), the other is the interval between the last input data and the future prediction point (denoted as prediction interval). Most previous work only considered the irregularity in the first case. The irregularity in prediction interval is usually ignored. In this paper, a new prediction model TA2D-LSTM is proposed by studying the impact of different kinds of irregular time interval as well as the different contribution of each visit for disease progress prediction. The proposed framework consists of two stages, the input sequence modeling stage and the prediction point modeling stage. In the first stage, the time aware LSTM structure is used to consider the irregular input intervals and mainly focuses on the contribution of recent visits. Then the attention network is built to learn the different contribution of all input visits. After that, the decay structure DECAY is designed to adjust the output of the first stage based on the irregular prediction interval. The adjusted results from the time aware LSTM and the attention network are then merged together to produce the final predictions results. The experimental results on a large-scale real-world hyperthyroidism disease data set verify the effectiveness of the proposed TA2D-LSTM for disease prediction task.