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.

Figure 1

Accuracy of age-predicting models. The biological age estimation according to deep convolutional neural network CNN_Age (A), the multivariate regression REG_Age (B), and the unsupervised PCA_Age (C) models. The solid lines and the transparent ± standard deviation bands are color coded as green, blue, and red, representing the whole population, the patients diagnosed with diabetes by a doctor, and individuals with self-reported high blood pressure, respectively. All the calculations were produced using the NHANES 2003–2006 cohort wearable accelerometers data, comprising one-week long activity tracks (1 min⁻¹ sampling rate).

Figure 2

Performance of the biological age models to distinguish between the longer- and the shorter-living individuals illustrated with Kaplan-Meier survival curves. Each participant was classified into the “high-” and the “low-” risks groups according to the deep convolutional neural network CNN_Age (A), the multivariate regression REG_Age (B), and the unsupervised PCA_Age (C) models. The p-values characterize the survival curves separation significance.

Figure 3

Architecture of convolutional neural network. The same network architecture was used to train both models for predicting age and for predicting mortality rates.

Figure 4

Learning curves for the CNN age and hazard predictor models. Learning curves for age-predicting model CNN_Age are shown for each of 5 folds for the training (A) and test (B) subsets. Cross-validation learning curves (B) show a minimum of MSE at epochs 500–800, while further training results in model overfitting. Panels (C,D) show the learning curves for the mortality risk-predicting CNN model. Here some of the cross-validation learning curves show a minimum of MSE at epoch 3000. Learning curves are color-coded according to the fold number in the 5-fold training/cross-validation setup.