Artificial Intelligence algorithm development in healthcare

Abstract

This project aims to complement the artificial intelligence module of the BCLink application by integrating new algorithms for critical event prediction. BCLink collects patient data, analyses it and displays biomarkers in a graphical user interface. Specifically, the project covers the areas of ECG delineation and patient-machine asynchronisms identification in mechanical ventilation. In both cases, the data is analysed, the prevalence of multiple approaches in the literature is studied and the optimum solutions are built and implemented. On top of that, different procedures for the deployment of the algorithms into the application are evaluated and ultimately their integration is achieved by use of Pythonnet and ONNX. In the case of ECG delineation all the P, QRS and T wave onset, peak and offset are assessed automatically by our algorithms both for on-demand and real time purposes. To that end, modifications of the former with regard to memory management are applied to fit the real-time analysis’ necessities. From the information in terms of duration and amplitude of each wave the following measures are computed: RR interval duration and heart rate, PR interval duration, QT interval duration, QT interval duration corrected, QRS wave duration, P wave duration, QT dispersion, QT dispersion corrected, QRS axis, P axis and T axis. Additionally, the computation of the ST deviation in real-time is implemented by means of an algorithm capable of identifying the J point. Regarding mechanical ventilation, the specific patient-machine asynchronism studied is reverse triggering. For its identification, the approach chosen is the use of a classification model and, in consequence, a comparison of the state-of-the-art machine learning tools is conducted. All of them are compared using a specific type of repeated stratified K-Fold validation that ensures the metrics obtained are as robust as possible. These are selected carefully in order to obtain performance evaluations that are unaffected by imbalances in the datasets, as it is our case. Finally, the best three models – KNN, Support Vector Machines and Random Forest – are deployed in a simulated environment, where Random Forest shows to be the one with better performance with unseen data