Evaluation of Sequential and Temporally Embedded Deep Learning Models for Health Outcome Prediction

Deep learning sequential models are increasingly being used to predict patients’ health outcomes by analyzing their medical histories. In this paper, we investigate the design decisions and challenges of using deep learning sequential models for predictive health modeling. Our results show that the most successful deep learning health models to date, called transformers, lack a mechanism to analyze the temporal characteristics of health records. To address this gap, we propose and evaluate a new model called DTTHRE: Decoder Transformer for Temporally Embedded Health Records Encoding. DTTHRE analyzes patients’ medical histories, including the elapsed time between visits. We also evaluate the performance of DTTHRE on a real-world medical dataset for two health outcomes: (1) diagnostic and (2) readmission prediction. DTTHRE successfully predicted patients’ final diagnosis (78.54 ± 0.22%) and readmission risk (99.91 ± 0.02%) with improved performance compared to existing deep learning sequential models in the literature. DTTHRE predicts the health outcome for each medical visit, which increases the training examples available from limited medical datasets with no additional training time.