Evaluation of Pacemaker Battery Level Estimation using Large Language Model and Remote Monitoring Historical Data

In recent years, remote monitoring (RM) where stored pacemaker's information is sent from home to hospitals, has become widespread. This leads to reduced burden of outpatient visits, and it is also possible to obtain useful information by analysing the large amounts of data stored with RM. However, two issues exists; the data format for RM is provided in different types of document data, which is time consuming to extract and aggregate the data manually by the clinical staff. In addition, RM reports are done once a month, resulting in only 12 reports a year, so as they are scarce, we need to unify reports for various types of patients. However, the data aggregation would require excessive workload for the healthcare professionals, therefore it is difficult to unify different types of patient data with the current RM.To solve these critical problems for the pacemaker, this study developed a pacemaker battery level estimation using a large language model (LLM) and remote monitoring historical data. With the aid of LLM, no coding is required as a relational database could be run fully automated, which should be a strong merit for non-engineers, namely, doctors and clinical staff. Based on the interactions through the LLM-based system, we can find similar patient data from the relational database easier, without the need for any technical skills. For evaluating the developed system, RM files were created for 30 people, simulating the discharge characteristics of pacemakers, and a high-order approximation curve of the characteristics was calculated from the RM history information using the OpenAI's generative pre-trained transformers (GPTs). The optimal order approximation curve was then selected using the Akaike information criterion (AIC), and a pacemaker battery remaining capacity prediction system was created. As a result, gpt-4o-latest showed the lowest root mean square error (RMSE), with a median value of 0.0124mV, which demonstrated that data analysis integrating RM and LLMs could become feasible in future pacemaker condition monitoring.