Research Paper Volume 13, Issue 6 pp 7900—7913

Deep longitudinal phenotyping of wearable sensor data reveals independent markers of longevity, stress, and resilience

Architecture of the neural network predicting biological age acceleration (BAA). GeroSense model predicts BAA once per day based on step counts recorded by wearable or mobile device sensors using each individual’s week-long physical activity tracks. The network components responsible for the feature extraction and BAA output are shown in green. BAA can be predicted for any sample of arbitrary length exceeding one week. For example, BAA on day 10 is predicted using the step counts data coming from day 4 through day 10, and so forth. Shown in red are the network components used only during the training procedure. One is the discriminator responsible for domain adaptation between e.g. smartphones and smartwatches. The other is the class predictor based on the log-odds ratio trained to predict morbidity binary status for UK Biobank and NHANES.

Figure 1. Architecture of the neural network predicting biological age acceleration (BAA). GeroSense model predicts BAA once per day based on step counts recorded by wearable or mobile device sensors using each individual’s week-long physical activity tracks. The network components responsible for the feature extraction and BAA output are shown in green. BAA can be predicted for any sample of arbitrary length exceeding one week. For example, BAA on day 10 is predicted using the step counts data coming from day 4 through day 10, and so forth. Shown in red are the network components used only during the training procedure. One is the discriminator responsible for domain adaptation between e.g. smartphones and smartwatches. The other is the class predictor based on the log-odds ratio trained to predict morbidity binary status for UK Biobank and NHANES.