020 - Robust individualized conformal predictive inference of patient's event times of Alzheimer's using real-world data

Alzheimer's disease (AD) and AD-related dementias (ADRD) are neurodegenerative diseases characterized by progressive loss of cognition along with other neurobehavioral symptoms. The heterogeneity in the progression pathways from normal cognition to various intermediate stages of the disease makes it difficult to accurately diagnosis and model mathematically. This paper aims to address some of these difficulties by proposing a reliable, accurate, effective, and patient-specific prediction of risks to AD/ADRD through mining the real-world OneFlorida database collected from healthcare organizations in Florida. The development is based on a machine learning method known as conformal prediction. With this method, no model assumption is needed to obtain a valid prediction, and it is also individualized. In addition to the standard conformal prediction setting in which the prediction is verified assuming the new individual is a random sample from the whole at-risk population, we use an individualized fusion learning method to develop the second type of prediction with conditional coverage: the validity is verified among patients with the same medical condition as the target new patients. Furthermore, we provide a novel data-adaptive imputation method to handle the dependent censoring case. All the desired properties of our development are demonstrated both theoretically and in simulation studies. We also apply our method to provide reliable and individualized risk assessments for at-risk patients and supply a time frame with confidence for each individual before which the individual is unlikely to develop AD/ADRD. This assessment is clinically and economically important for health system, patients, and stakeholders.

Alzheimer's disease and AD-related dementia

individualized risk prediction

conformal prediction

iFusion learning

censoring

imputation 

Mid-Level

Knowledge

International Conference on Health Policy Statistics 2023