Methods/Statistics
What goes wrong in prediction models if you ignore mortality? Stephanie Armbruster* Stephanie Armbruster Sebastien Haneuse Harrison Reeder Daniel Kramer
Prediction is a central pillar to epidemiology; it supports informed healthcare decisions on a patient and population level and is essential to research in the design and analysis stage. In many settings of epidemiological interest, the force of mortality is strong. Mortality – as a competing risk – prevents health-related outcomes from happening. It partly determines a prediction; it truncates and implies non-existence of outcomes post-death. Yet, standard prediction models, such as regression-based methods, frequently ignore mortality and models imply a ‘pretend reality’ in which patients are considered immortal. Ignoring mortality may result in bias. If used in clinical practice, it can compromise patient-specific health-care decision making, mislead risk stratification in targeted trial design or misinform the implementation of public health interventions. The agnosticism to mortality also misaligns with a patient’s reality; health-related decisions naturally involve a complex trade-off between survival and quality-of-life.
We will illustrate what goes wrong in prediction when ignoring mortality by presenting real-world scenarios in which bias arises based on data from the PIPER-ICD study. Sudden Death (SD) is the primary cause of death in the US. Patients at increased risk of suffering a SD are primarily treated with an implantable cardioverter defibrillator (ICD). ICD-related predictions among older patients are biased if they do not account for higher risk of all-cause mortality at the end-of-life. In this vein, PIPER-ICD observed geriatric conditions, frailty metrics, and quality-of-life endpoints among patients over the age of 65 with ICD devices to analyze the end-of-life experience with an ICD.