Onset acceleration: a new kind of biological aging metric
There is a lot of public interest in the idea of aging clocks, functions that output an organism or tissue’s estimated “biological age”. Biological age is often conceptualized as the chronological age of physiologically or functionally comparable subjects in a healthy reference population:
An individual’s age defined by the level of age-dependent biological changes, such as molecular and cellular damage accumulation. In practical use, this is often summarized as a number (in units of time) matching the chronological age where the average person in a reference population shares the individual’s level of age-dependent biological changes.
In a new paper, I propose an adaptation of the traditional Cox proportional hazard model used in popular health-risk clocks such as PhenoAge and GrimAge. In the study, we set the Cox temporal baseline to birth instead of trait measurement time to directly model of rate of aging (in the sense of rate of reduction of diagnosis-free time since birth). Under this setting, time-to-event corresponds to age at diagnosis, and the Cox hazard corresponds to onset acceleration risk. The risk model was subsequently used to uncover structure in how onset correlation of different disease conditions are related:
A summary of contributions is given below:
Using blood biochemistry, MRI imaging and electronic health records of 60,396 participants from the UK Biobank prospective cohort study, a neural network Cox model was trained to predict individual log-risk of first diagnosis of each age-associated condition. For the methodological contribution, our results validated that Cox models can be utilized to predict prognostically informative risk scores for time-to-event since birth. Evaluation of input biomarker combinations and model architectural choice demonstrated the importance of information rich predictor traits and the performance benefits of neural networks over linear models on in-distribution test data, whereas linear models outperformed on out-of-distribution test data. Correlation analysis of onset acceleration score between diseases showed highly structured associations between onset acceleration risk of different age-associate conditions. Disease clusters with highly correlated onset acceleration included cardiometabolic, vascular-neuropsychiatric, and digestive-neuropsychiatric groups, corroborating the notion that age-associated diseases do not co-occur randomly but in synchronous fashion (Prados-Torres et al. 2014, Ng et al. 2018). Saliency analysis on trained neural networks inferred pre-supported associations between biomarkers and disease in a precise manner, in particular suggesting the importance of body composition, grip strength, gamma glutamyltransferase, glycated hemoglobin, high-density lipoprotein cholesterol and urate for predicting onset acceleration risk in both sexes.
The paper is currently under review. A preprint can be downloaded here.