Extracting biological age from biomedical data via deep learning: too much of a good thing?

Abstract

Age-related physiological changes in humans are linearly associated with age. Naturally, linear combinations of physiological measures trained to estimate chronological age have recently emerged as a practical way to quantify aging in the form of biological age. In this work, we used one-week long physical activity records from a 2003–2006 National Health and Nutrition Examination Survey (NHANES) to compare three increasingly accurate biological age models: the unsupervised Principal Components Analysis (PCA) score, a multivariate linear regression, and a state-of-the-art deep convolutional neural network (CNN). We found that the supervised approaches produce better chronological age estimations at the expense of a loss of the association between the aging acceleration and all-cause mortality. Consequently, we turned to the NHANES death register directly and introduced a novel way to train parametric proportional hazards models suitable for out-of-the-box implementation with any modern machine learning software. As a demonstration, we produced a separate deep CNN for mortality risks prediction that outperformed any of the biological age or a simple linear proportional hazards model. Altogether, our findings demonstrate the emerging potential of combined wearable sensors and deep learning technologies for applications involving continuous health risk monitoring and real-time feedback to patients and care providers.

Biological age is a universal marker of aging, stress, and frailty

“…The authors also mentioned that its performance was superior to epigenetic biomarkers of aging (see, e.g., the value of p = 1.7E − 21 reported in for Hannum model (Hannum et al 2013)) trained to predict chronological age. This result also seems consistent with with our findings here with clinical blood markers and earlier in analysis of human physical activity-based bioage models (Pyrkov et al 2018b). The overall conclusion is that the most accurate chronological age predictors produced the poorest associations with all-cause mortality (see Table I) and thus should be avoided whenever possible in favor of the explicit mortality or morbidity models.…”

“…Advanced machine learning tools are naturally called to improve the biological age predictions, see, e.g., a model trained from clinical blood markers (Putin et al 2016) and facial photos (Bobrov et al 2018). The examples presented here show that the full power of the deep learning architectures could be harnessed for feature extraction and non-linear models fitting of risks, rather than chronological age models (Pyrkov et al 2018b). The risk models, however, require follow-up information involving the incidence of age-related diseases or death.…”

“…The first choice was the Cox proportional hazards model with blood markers as covariates only, with no explicit age or sex labels. The log-scaled hazards ratio is a linear combination of blood features and can be calibrated in years as outlined in (Liu et al 2018, Pyrkov et al 2018b and was used as an alternative biological age predictor, the "MORTAL-bioage". Although the correlation of the predicted age with the chronological age was lower (r = 0.35 with RM SE = 17.5 years), its association with mortality was notably improved (HR = 1.08, p = 4.5E−138) relative to that of the models based on the chronological age only, see Table I.…”

“…We, therefore, held the baseline model parameters fixed in Eq. 1 and optimized the deep network to include the physiological-state dependent variables through the network output variable x n (Pyrkov et al 2018b). As for all other models above, the log-hazard ratio of the deep mortality model was re-scaled to years and referred to as the "Deep HAZARDS-Blood-Survey" bioage.…”

“…This can be best seen from a Principal Component Analysis (PCA), which is a mathematical procedure aimed at the identification of the most correlated or collective features in multidimensional data (Ringnér 2008). The method is an almost always the first choice in biological data analysis and searches for aging signatures in a multitude of biological signals ranging from blood samples to locomotor activity patterns (Pyrkov et al 2018b), anthropological and functional measures (Park et al 2009), and microbiome composition (Odamaki et al 2016). In most cases only a few, most often the first, PCA scores were associated with age and hence could be naturally interpreted as biomarkers of age Miyao 2007, Nakamura et al 1988), see Jia et al (2016) for a recent review and references therein.…”