BACKGROUND
Without timely diagnosis and treatment, tachycardia, also named as tachyarrhythmia, could cause serious complications such as heart failure, cardiac arrest, and even death.
OBJECTIVE
The predictive performance of the conventional clinical diagnostic procedures needs improvement in order to assist physicians for the early risk detection.
METHODS
We propose a Deep Personalized Tachycardia Onset Prediction (DeePTOP) model based on a deep learning model (i.e. BiLSTM) and k-means clustering algorithm for the early individualized tachycardia diagnosis. DeePTOP leverages the vital signs including heart rate (HR), respiratory rate (RR), and blood oxygen saturation (SpO2) acquired continuously by wearable embedded systems and the electronic healthcare records (EHR) containing age, gender, admission type, first care unit, and the history of cardiovascular diseases. The model was trained by a large dataset from intensive care unit and was then transferred to a real-world scenario in general ward. In this study, three experiments including the individualized prediction characteristic, the temporal memory function, and different types of features combination were conducted and six metrics (area under the receiver operating characteristic curve (AU-ROC), precision-recall curve (AU-PR), sensitivity, specificity, accuracy and F1 score) were used for the evaluation of predictive performance.
RESULTS
Our DeePTOP outperformed the baseline models (AU-ROC: 0.806; AU-PR: 0.725, sensitivity: 0.743; specificity: 0.745; accuracy: 0.749; F1 score: 0.681) when predicting tachycardia onset (TO) six hours in advance on the large dataset. When predicting TO two hours in advance by the data acquired from our hospital using the transferred DeePTOP, the six metrics were 0.904, 0.843, 0.894, 0.898, 0.795, and 0.766, respectively. Among them, the best performance was achieved with comprehensive statistical information of vital signs (HR, RR, and SpO2).
CONCLUSIONS
DeePTOP is a personalized tachycardia prediction model using eight types of data recorded by wearable sensors and EHR. When validated by real clinical scenarios, the model achieved an outperforming prediction performance 0 - 6 hours before TO. Owing to our implementation and easily accessible data by wearable sensors, we suggest that the model can assist physicians to early discover the potential risks of patients in general wards and houses.
CLINICALTRIAL
No. S2018-095-01