Hypotension prediction using vital signs

Background: Intraoperative hypotension is linked to adverse events. Forecasting incidence of hypotension is a taxing problem. To tackle this issue, machine learning and deep learning methods can be utilized.
Methods: Using feature engineering, 11 features were formed from blood pressure signals and together with patient age and sex were used to train five different machine learning models in combination with 10 different oversamplers to classify hypertension occurrence. Furthermore, raw wave-form data was used to train an RNN and an LSTM with the same goal of classifying hypertension occurrence.
Results: The best machine learning model was found to be random forest classifier with Lee oversampler and it yields balanced accuracy of 96.87%, average precision of 99.87% and 99.51% ROC-AUC on test data, whilst the best deep learning implementation was LSTM with balanced accuracy of 97.06%, average precision of 99.82% and 99.21% ROC-AUC on test data.
Conclusions: This paper demonstrates that prediction of hypotension occurrences using blood pressure signal is achievable. Among the tested algorithms, random forest and LSTM yielded the highest accuracy, whilst the gradient boosting algorithm was able to achieve high accuracy and offer high levels of interpretability.