Developing and applying machine learning algorithms to improve prediction in patients with heart failure
The goal of the project is to develop novel statistical methodology to a) improve self-management of patients with heart failure, and b) improve prediction of complications in patients with a heart pump. The project is a close collaboration between the Department of Methodology and Statistics (Faculty of Social Sciences, Utrecht University) and the Department of Cardiology (University Medical Center Utrecht).
The PhD project is composed of two related, methodological projects. The first project concerns the improvement of self-management of patients with heart failure. In UMC Utrecht, heart failure (HF) patients can safely monitor and manage their disease using a telemedicine platform called EMPOWER. Using this, patients are enabled to safely act first on changes in their health status: when changes in biometric measurements exceed a clinical threshold, signaling risk, an alert is triggered. Patients can act on these alerts by changing diet, fluid intake, and selected pharmacological therapy themselves. Currently, these thresholds are uniformly prespecified (and thus equal for all patients), and can be manually changed by the patient’s physician or specialist HF nurse. However, thresholds that are set either too wide or too tightly could lead to missing episodes of acute deterioration or too many unhelpful alert triggers. This problem warrants an approach tailored to the individual patient. The goal of this project is to achieve this by extending statistical process control techniques.
The second project concerns the prediction of complications in patients with a heart pump (also known as Left Ventricular Assist Device; LVAD). Major adverse events in patients with LVAD therapy are common and often occur suddenly and unpredictably. However, it is possible to retrieve a wealth of information from the LVAD itself, as well as from patients on LVAD support. Currently, appropriate analysis models to link adverse events to these data are lacking.
Both the EMPOWER platform and the LVAD provide a wealth of intense longitudinal information on biometric measures of the patient. These could be exploited to predict a) automated, personalized and more accurate thresholds for EMPOWER patients, and b) (personalized) clinical outcomes for patients on LVAD
The project comprises three studies that will: develop patient-specific predictions of thresholds/clinical outcomes for separate biometric measurements using proven time series modeling (i.e. statistical process control(Benneyan, Lloyd et al. 2003, Suman and Prajapati 2018) techniques);
develop a statistical model in which separate patient-level biometric measurements are integrated into a single model using machine learning techniques (i.e. Markov related models and/or dynamic factor models(Stanculescu, Williams et al. 2013, Huang, Cohen et al. 2018)). The statistical ensemble model developed for both patient groups could be very similar, or it could be that the data of the EMPOWER and LVAD patients require two different approaches;
validate the developed algorithms with feedback loop using a) routinely collected individual outcome data of patients using the EMPOWER platform, or b) alternative sample datasets of patients on LVAD support.
Dr. Emmeke Aarts
Dr. Daniel Oberski
Prof. Dr. Folkert Asselberg
Dr. Linda van Laake
1 December 2020 – 1 December 2024