Introduction and objectives: Advanced heart failure patients usually require a donor heart for destination therapy. However, since donor hearts are scarce, left ventricular assistant devices (LVADs) have become a popular alternative. Unfortunately, patients often experience complications with this treatment. Telemonitoring LVAD parameters such as power and flow may improve outcomes by detecting early signs of deterioration. Currently, LVADs use a simple monitoring approach with a fixed threshold for all patients, resulting in delayed detection of adverse events and false alarms. To address this, we developed a personalized algorithm that can detect unscheduled admissions caused by common complications like cardiac arrhythmia and major bleeding. This tailored algorithm allows for better detection and management of complications in individual patients. Methods: The algorithm uses patient-tailored thresholds to identify abnormal power and flow observations. It employs a linear mixed-effects (LME) model that considers pump parameters of a group of stable patients without any admission and the longitudinal data of each individual patient. This results in a personalized mean pump value that is flexible and reflects the patient's stable historical baseline. The patient-specific mean is then subtracted from real-time measurements to obtain residuals, which are smoothed with an exponentially weighted moving average (EWMA) statistical process control chart, and compared to upper and lower control limits determined by the EWMA control chart. If the smoothed residuals exceed these control limits, the algorithm triggers an alarm. Our findings indicate that PRECISION-LVAD was capable of detecting 59% and 79% of cases related to cardiac arrhythmia and major bleeding, with a low false alarm rate (FAR) of 2%. However, the FAR varied between patients. Furthermore, the median number of days between the first alarm and admission due to CA or MB was 6.5 and 7.0 days, respectively. Conclusion: Although PRECISION-LVAD shows promise as a powerful tool for detecting cardiac arrhythmia and major bleeding, some events were still not detected by the algorithm. Therefore, continuous refinement of the algorithm using data streams is necessary. One possible approach is to use latent variable models such as Hidden-Markov models to monitor patients based on their switching hidden states. This would allow for more accurate and timely detection of cardiac arrhythmia and major bleeding. Keywords: Personalized monitoring, linear-mixed effect model, Statistical process control chart.

We advise the users to import their datasets. If you do not have access to actual LVAD log files, you can use the attached synthetic datasets.