Machine Learning for Assessing Vital Signs in Humans in Smart Cities Based on a Multi-Agent System

Healthcare professionals face numerous challenges when analyzing data and providing treatment, including determining which parameters to measure, the frequency of measurement, i.e., how frequently to measure them, and the responsibility for monitoring patient health with new medical devices. Machine learning (ML) techniques are efficient predictive models used to improve early prediction of patient care and reduce the cost of implementing healthcare systems. This study proposes a new model (data prediction and labeling using a negative feature based on a multi-agent system (PLPF-MAS)) that provides a smart city-based healthcare system for the continuous monitoring of patients’ vital signs, such as heart rate, blood pressure, respiratory rate, and blood oxygen saturation. It also predicts future states and provides suitable recommendations based on clinical events. The MIMIC-II database of the MIT physio bank archive is used, which contains 1023 patient records. Additionally, the EHR dataset is used, which contains 10,000 patient records. The models were trained and evaluated for six bio-signals. The PLPF-MAS model is distinguished from traditional methods in its advanced system, which combines the activities of several agents and the intelligent distribution of responsibilities among them. The LR agent measures the model’s reliability in parallel with the AE-HMM agent to predict the Prisk; it then sends the data to a coordinator and a supervisory agent to monitor and manage the model. Our model is characterized by strong flexibility and reliability, the ability to deal with large datasets, and a short response time. It provides recommendations and warnings about risks, and it can predict clinical states with high accuracy. The new model achieved an accuracy of 98.4%, a precision of 95.3%, a sensitivity of 99.2%, a specificity of 99.1%, an F1-Score of 97.1%, and an R2 of 98%, when the MIMIC-II dataset was used. Conversely, it achieved an accuracy of 93%, a precision of 92%, a recall of 94%, an F1-Score of 93%, an AUC-ROC of 94%, and an AUC-PR of 89% when the EHR dataset was used.