PROJECT TITLE :

An Explainable Transformer-Based Deep Learning Model for the Prediction of Incident Heart Failure

ABSTRACT:

It can be difficult to make accurate predictions about the occurrence of complex chronic conditions such as heart failure. Deep learning models that are applied to rich electronic health records might improve prediction, but these models' inability to explain their results prevents them from being used more widely in clinical settings. Our goal was to develop a deep learning framework capable of accurate and, at the same time, explainable prediction of heart failure occurring within six months (HF). We applied a novel Transformer-based risk model to 100,071 patients from longitudinally linked electronic health records across the United Kingdom. The model took into account all community and hospital diagnoses as well as medications, and it contextualized these factors within the patient's age and the calendar year for each patient's clinical encounter. An ablation analysis was used to investigate the importance of features by comparing the performance of the model after removing features in a variety of different orders, as well as by contrasting the variations in temporal representations. For the purpose of carrying out feature contribution analyses, a post-hoc perturbation method was applied in order to propagate the changes that were made to the input to the outcome. Our model outperformed other deep learning models by achieving a score of 0.93 on the area under the receiver operator curve and a score of 0.69 on the area under the precision-recall curve during an internal 5-fold cross validation. Analysis of ablation showed that medication plays an important role in predicting HF risk, and that the year on the calendar plays a more significant role than age as measured in years. This was further supported by analysis of temporal variability. The contribution analyses uncovered a number of risk factors that have a close association with HF. Although many of them were in line with what was already known from clinical and epidemiological research, several new associations were found that had not been taken into account in expert-driven risk prediction models. In conclusion, the findings shed light on the fact that our deep learning model, in addition to having a high predictive performance, can contribute to the identification of data-driven risk factors.