Cyber-Physical System Framework for Cerebrovascular Accidents using Machine Learning Algorithm

This paper provides a proposed framework for a medical cyber-physical system for personalized, attentive care for patients at risk of cerebrovascular accidents (CVA). It compares machine learning techniques for their preliminary implementation. The proposed framework was developed by analyzing the clinical business process and modifiable/non-modifiable risk factors for CVA before being mapped onto accessible IoMT devices that can be used for the modified CPS framework. A publicly available dataset from the NY Open Data repository with class imbalance was treated using SMOTE then used with six ML techniques, namely: artificial neural network (ANN), logistic regression (LR), random forest (RF), ensemble voting, gradient boost (GB) and AdaBoost. They were then compared based on ROC AUC as the primary metric to determine individual classification ability. Results show that models that dealt with the imbalanced data could have high accuracy, like RF with 97.28% but still not perform well enough with an AUC of 0.71. However, ensemble techniques with over 95% accuracy obtained an AUC of 0.82. The results obtained from the conduct of this study will be used for further implementations of the MCPS framework and further improving the algorithms used as a component of the proposed CPS framework.