Human Mental State Monitoring in the Wild: Are We Better Off with DeeperNeural Networks or Improved Input Features?

Abstract

Advances in wearable devices have allowed for the collection of multimodal biomedical data from hundreds of subjects in everyday environments (i.e., in the wild). This has enabled the development of real-time monitoring of various human mental states, such as stress and anxiety, in highly ecological settings. Within a hospital setting, for example, this allows for prediction of burnout within medical staff, as well as anxiety within the patient population, thus improving their quality-of-life. Long-term monitoring via wearables has allowed for large amounts of data to be collected – so-called big data– and thus has opened doors for new applications relying on data-heavy deep learning algorithms. One question that remains unanswered, however, concerns the benefits of blindly applying deep learning algorithms with the collected data versus spend-ing some time and resources on feature engineering prior to machine learning. Feature engineering relies on domain knowledge to extract relevant parameters from the collected signals. In this paper, we aim to answer this question. In particular, we use a dataset collected from 200 hospital workers over a period of 10weeks during their work shifts. We compare the advantages of using data directly from the wearable devices and applying them to deep learning algorithms versus carefully-crafted features ap-plied to conventional machine learning algorithms. Experimental results are reported for stress and anxiety measurement from heart and breathing rate signals.