MEASURING BIOLOGICAL AGE AS A TASK OF INDIVIDUALIZATION: A FRAMEWORK FOR PREDICTOR SELECTION AND MODEL VALIDATION

Abstract

Myriads of methods have been proposed to measure biological age with the goal of enabling a personalized prognosis of age-related health issues and tailored treatments. However, it is highly unclear which method suits the goal of measuring biological age best in the context of a given data-set. Here, we present a new statistical definitional framework for quantifying biological age, generalizing a recently published work on individualization. We explicate the statistical prerequisites for the meaningful differentiation of two individuals of the same chronological age regarding their true biological age by supervised methods, namely that each contributing variable must be stochastically independent of chronological age given the true biological age. If this specific condition is not fulfilled, we reveal that unsupervised methods will often outperform supervised methods. In the consequence, we prove that the most commonly applied predictor selection methods based on maximizing the correlation to chronological age lead to a heavy loss of biological and clinical information in a wide range of scenarios. Finally, by using data from the large prospective cohort Study of Health in Pomerania (n=4308) with extensive follow-up and baseline phenotyping, we show that published metrics of brain ageing and metabolic ageing can be seriously improved by operationalizing our theoretical results. In conclusion, by reframing the measurement of biological age as a problem of individualization, we deliver an applicable and theoretically sound statistical framework which allows a qualified decision on the method of choice for constructing a metric for biological age in a certain data-situation.