Deep learning with wearable based heart rate variability for prediction of mental and general health

Abstract

The ubiquity and commoditisation of wearable biosensors (fitness bands) has led to a deluge of personal healthcare data, but with limited analytics typically fed back to the user. The feasibility of feeding back more complex, seemingly unrelated measures to users was investigated, by assessing whether increased levels of stress, anxiety and depression (factors known to affect cardiac function) and general health measures could be accurately predicted using heart rate variability (HRV) data from wrist wearables alone. Levels of stress, anxiety, depression and general health were evaluated from subjective questionnaires completed on a weekly or twice-weekly basis by 652 participants. These scores were then converted into binary levels (either above or below a set threshold) for each health measure and used as tags to train Deep Neural Networks (LSTMs) to classify each health measure using HRV data alone. Three data input types were investigated: time domain, frequency domain and typical HRV measures. For mental health measures, classification accuracies of up to 83% and 73% were achieved, with five and two minute HRV data streams respectively, showing improved predictive capability and potential future wearable use for tracking stress and well-being.