Prediction of Heart-Failure with Machine Learning

Overview

In this monocentric observational study the research question is to what extent data collected via Apple Watch can predict the heart failure status of decompensated HF patients. For this purpose, physiological data from the Apple Watch (such as single-lead electrocardiogram, SpO2, respiratory rate, step count, nighttime temperature, etc.) will be extracted and used as predictor variables to forecast outcomes like risk of decompensation and rehospitalization within the follow-up period. Since this is a data-driven study, additional data collected as part of guideline-compliant treatment will also be included.

Wearable devices for measuring vital functions, known as "wearables" from the consumer sector, such as the Apple Watch, have gained significant popularity. Increasingly, they are also being used for cardiovascular assessments. For example, a previous study at the Department of Cardiology and Pulmonology demonstrated that the Apple Watch is well-accepted by HF patients and that the average daily step count correlates significantly with the 6-minute walk test. Research has since shifted from simple correlation analyses to more complex tasks, such as predicting clinical laboratory measurements or events, like decompensation in HF patients. Machine learning methods have proven to be suitable for various predictions in the field of heart failure. For instance, it has been shown that ECG data can be used to predict HF risk surrogates or comorbidities, such as NT-proBNP levels, age or gender, anemia, or renal insufficiency. Beyond ECG, multimodal approaches that combine multiple measurements have demonstrated the feasibility of data-driven HF risk assessment. Examples include combining cardiac MRI with clinical information to predict time to hospitalization or HF incidence rates in atrial fibrillation. Since ECG alone does not provide sufficient prognostic value for heart failure (HF) assessment, this study aims to advance the state of the art by incorporating additional sensor data. The Apple Watch will be utilized as the device of choice. Extracted parameters include respiratory rate, oxygen saturation, nighttime temperature, acceleration data, and automatically provided derived parameters (e.g., step count, sleep times). Specific Objectives: * Collect data from HF patients using the Apple Watch. * Extract and integrate the data into a unified format. * Perform correlation analysis with clinical parameters. * Develop an algorithm/model to predict clinical parameters. * Statistically evaluate the predictive power of the developed model. Clinical Parameters of Interest: * Changes in the Kansas City Cardiomyopathy Questionnaire (KCCQ). * Changes in N-terminal pro-B-type natriuretic peptide (NT-proBNP) levels. * Changes in hemoglobin concentration. * Changes in kidney function (creatinine, GFR). * Changes in standard clinical examination parameters: electrolytes, liver function, blood count, CRP, thyroid function. * Changes/adjustments in medication by treating physicians based on clinical symptoms. * Changes in volume status (edema, body weight, chest X-ray). The KCCQ and NT-proBNP levels will be measured upon admission, and changes from these baseline values will be assessed at two time points: on the day of discharge and 90 days post-discharge. Since this is a data-driven study, additional data collected as part of guideline-compliant treatment will also be included and analyzed for correlations. This includes the DZHK core dataset consisting of 42 items, laboratory parameters determined from patient blood samples (e.g., blood count, serum analysis, coagulation tests), echocardiography performed twice during hospitalization (e.g., LV-EF, LAVI, severity of valvular disease), body weight/edema status trends, and medications administered during the hospital stay. Procedural Description On the day of admission, eligible HF patients will be identified and recruited based on the above criteria. Each participant will receive an Apple Watch and an introduction to its use. On the day of discharge, the Apple Watch will be collected and handed over to the Institute for Medical Informatics for data extraction. The data analysis will be performed offline. Procedure for Informed Consent Informed consent will be obtained personally and through the distribution of written materials designed for clarity and readability. These materials are included in the appendix of this application. Follow-Up After 3 Months A follow-up appointment will be scheduled 90 days (±10 days) after discharge for laboratory and echocardiographic examinations. During this follow-up, laboratory parameters including blood count, serum analysis, and coagulation tests will be reassessed, and an echocardiography will be performed. Interim hospital admissions and associated clinical information (reason for admission, treatments administered) will also be recorded. Evaluation Measures To avoid biases and errors in conclusions, specific measures will be implemented during all evaluations: * The dataset will be divided into training, validation, and test subsets. * The test dataset will be extracted at the beginning and reserved exclusively for the final evaluation of the model. * Training and validation data will be split using cross-validation. * To prevent data leakage, all preprocessing steps will be performed solely on the training dataset. * Results will be reported based on the guidelines of the "Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD)." Study Population The study will include HF patients with acute decompensated heart failure of the HFrEF type (reduced left ventricular ejection fraction). Observation Period: April 1, 2024 - May 30, 2025 (including a 90-day follow-up). Expected Number of Participants: 32 patients. Expected Risks HF patients will be treated according to current guidelines. In addition, they will wear Apple Watches to collect physiological data. The Apple Watch has a CE marking, ensuring compliance with all EU safety and health protection requirements. No interactions between the Apple Watch and guideline-compliant treatment are expected. Therefore, participation in the study does not pose any medical disadvantages or risks to the participants. Methodology and Analysis To evaluate the developed models, standard statistical metrics will be applied: * For regression tasks (e.g., predicting changes in the KCCQ score), metrics such as Root Mean Square Error (RMSE) will be used. * For classification problems (e.g., predicting anemia), metrics such as the Area Under the ROC Curve (AUC) and the F1-score will be applied.