A sequential tree-based classifier for personalized biomarker testing of Alzheimer's disease risk

Using baseline biomarkers to predict the conversion of mild cognitive impairment (MCI) to Alzheimer's disease (AD) has considerable clinical interest in recent years. The existing studies have several limitations, including unsatisfactory accuracy due to MCI heterogeneity, use of conventional classification models that require biomarkers to be measured all at once instead of sequentially and as needed, and use of raw numerical measurement of the biomarkers instead of discretized levels that are more robust to measurement errors and provide convenience for clinical utilization. To tackle these limitations, we propose a novel sequence tree-based classifier (STC) for predicting the conversion of MCI to AD. Different from conventional classification models, STC achieves a sequential, as-needed use of biomarkers and a three-category classification (high-risk converter, low-risk converter, and inconclusive diagnosis) by finding an optimal sequence of biomarkers and two-sided cutoffs of each biomarker that satisfy pre-specified accuracy requirements while minimizing the proportion of inconclusive diagnosis. STC is also a personalized approach, as it allows patient characteristic variables to be included to help identify patient-specific cutoffs for each biomarker. We apply STC to two important clinical applications using the data from the worldwide Alzheimer's Disease Neuroimaging Initiative project: prediction of MCI conversion and patient selection for AD-related clinical trials.