Towards Optimizing Left Ventricular Assist Device (LVAD) Therapy for Patients with Advanced Heart Failure: Exploring Machine Learning Applications in Pre- and Post-LVAD Therapy

Heart failure is a significant global public health concern affecting millions of individuals. Implantable Left Ventricular Assist Devices (LVADs) have proven beneficial for patients with advanced heart failure who are unresponsive to conventional treatments. However, despite the improved survival rates associated with LVADs, this therapy carries a high risk of recurrent and severe adverse events (AEs) that lead to increased morbidity and mortality. Many clinical studies have examined the AE profiles of LVAD patients, but these studies often focus on statistical analysis and treat each AE as a separate event, overlooking potential relationships and interactions among AEs. This thesis aims to overcome these limitations by exploring three machine learning applications integrated with a National registry dataset during both pre- and post-implantation time frames. The first module involves post-LVAD sequential AE pattern mining and clustering, providing a comprehensive view of the AE landscape in the LVAD population. By identifying critical time points and subgroups with distinct AE patterns that significantly impact therapy outcomes, this module may serve to inform personalized care strategies, resource allocation, and follow-up scheduling. The second module developed a pre-implant predictive risk model that assesses AE risk based on patients' preoperative clinical profiles. This model may serve to enhance patient selection and management of high-risk patients prior to the procedure. The third module created a post-LVAD risk model predicting future AEs based on pre- and post-LVAD clinical profiles supplemented with the patients' AE history in first critical weeks. This risk may help clinicians to anticipate subsequent AEs and employ preventive measures. These three modules collectively represent a crucial step towards optimizing LVAD therapy and supporting human decision-making by integrating comprehensive clinical information, treatment history, and temporal events. Future translation of these modules into clinical application software will contribute to a clinical decision support system through the integration of various data mining tasks linked with temporal clinical milestones