Monitoring and diagnosis of cardiac anomalies in hospitalized patients using IoT and LSTM neural networks

Resumo

Continuous monitoring of patients in hospital settings is essential for preventing cardiac disease, representing a crucial measure to ensure the safety of hospitalized patients. However, current monitoring methods have significant limitations, such as patients only being monitored for short periods, which complicates the identification of changes in health status. This article presents a proof of concept for successfully implementing a continuous vital signs monitoring system using the Internet of Things (IoT), coupled with a Long Short-Term Memory (LSTM) neural network for detecting cardiac anomaly events.The development of the neural network hinged on a meticulously labeled database. This database, composed of 12-lead electrocardiogram (ECG) signals from 45,152 patients, was a labor of expertise. Collected at a sampling rate of 500 Hz, these data were labeled by experts, ensuring their high quality and reliability. This resource is fundamental for training machine learning models with high diagnostic precision, enabling accurate identification of cardiac rhythm patterns that indicate potential adverse conditions or those that deviate from a normal or stable electrocardiogram pattern. To make practical use of this extensive data, a robust system architecture was necessary to handle and analyze the data effectively.

Microcontrollers played a crucial role in our system, facilitating the transmission of data to a web server. This was achieved using the Message Queuing Telemetry Transport (MQTT) protocol, a highly efficient and fast method for message transmission. The MQTT server, acting as a central hub, received the information and distributed it to two destinations. First, it sent the data to the real-time monitoring site, ensuring immediate access to patient vital signs. Second, it forwarded the data to the system hosting the neural network model. In this environment, the model processed the information and promptly identified any irregularities detected in the data.The results demonstrate that the model achieved high precision in identifying these conditions, enabling the issuance of alerts and the implementation of preventive measures with a high efficacy rate. Such actions reduce patients' health risks, allowing real-time detection and remote communication of vital signs to healthcare professionals.