Midlands and Lancashire CSU (MLCSU)

=LEFT(J64,6)

Validation and calibration of algorithms predicting the risk of outcomes in heart failure patients in NWL

Heart failure (HF) is a chronic, progressive condition in which the heart cannot pump enough blood to meet the body’s metabolic needs. Heart failure is a leading case of hospital admissions in older patients in the UK, with the average age at diagnosis being 77, it equates to 5% of all emergency admissions, adding pressure to the already busy health system. These emergency department attendances continue to rise year on year. Timely diagnosis and management of heart failure (HF) is critical, but identification of patients with suspected HF can be challenging, especially in primary care. The Division of Pharmacoepidemiology and Pharmacoeconomics at the Department of Medicine, Brigham and Women’s Hospital and Harvard Medical School have developed predictive models in HF that can be applied to a population so that interventions could be focused to improve clinical outcomes and/or reduce the cost of care for subgroups of patients with HF. Aim: Evaluate the performance of the Harvrad predictive model for predicting outcomes of interest in selected patients with heart-failure in NWL in terms of the following end-points: - 1-year HF-related hospital admissions - 1-year all-cause mortality The model may then be calibrated / adjusted to optimise their performance. Methods: Model Validation Harvard-developed predictive models for the key outcomes (heart failure hospitalisation and total costs of care) will be evaluated in the NWL population. The models’ performance will be access using receiver operating characteristics (ROC) curves (including the measures of sensitivity, specificity and positive predictive value (PPV)), calibration, and Brier score. Note: since only model performances will be evaluated, the study population (retrospective sample data) will be used as a test dataset, and the validation process will entail comparing the expected HF identified via the algorithms with actual HF identified in the observational data via clinical coding. Model Calibration - Adjusting the parameters/ coefficients for better fit to the validation dataset. - Considering the optimal cut off considering markers such as sensitivity, specificity, and the positive predictive value. - Identifying an alternative fit model using machine learning (artificial intelligence software). Target Population Assessment After the model calibration/adjustment, the population identified from the predictive model will be assessed as follows - three high-risk population cohorts will be identified based on the predicted probability from the model. 1. Having heart failure hospital admission 2. Having top 10th percentile of total healthcare costs

16/12/2021

TRE