Browsing by Author "Choto Segovia, Alicia"

Now showing 1 - 1 of 1
  • Thumbnail Image
    Item
    Domain Adaptation, Stress, and Burnout Prediction in Shift Workers Using Wearable Data
    (2023-04-21) Choto Segovia, Alicia; Sano, Akane
    Healthcare workers are part of the shift worker population, and due to their occupational demands, they are at higher risk of developing mental health problems. With mental health disorders on the rise and healthcare workers’ shortages, it is imperative that workers have the tools to manage stress and prevent burnout. There has been a lot of progress in the applications of wearable sensors and machine learning to predict well-being, including stress, anxiety, and mood forecasting. However, wearable sensors collect time series data with high variability between domains with a small number of labels, making it challenging to maintain good predictive performance across domains. A domain is composed of an input space, an output space, and an associated probability distribution. The challenge of domain adaption is to train a model on a source domain and achieve a small error when tested on the target domain. The source and target domains share the same input and output space, but they have different distributions. Machine learning models need to adapt to new test users with data distribution different than the training data and leverage the unlabeled data to improve their predictions. In this work (1) we designed a personalization framework to adapt stress classifier models to new users. We found the closest points in the training set to the testing set and use them to personalize user-independent models. Our results showed that we can use the test users’ unlabeled data to tailor the training set to new users and improve performance from the user-independent model. (2) We proposed a pipeline to study the domain divergence between two datasets, and subsets within the datasets, to identify the best groups to perform domain adaptation. We measured domain divergence using Proxy-A-distance and used domain adversarial neural networks to extract domain invariant representations for stress prediction. We concluded that even if we extract invariant representations, it does not guarantee good performance on the target domain if the domain distributions are too different. (3) We implemented machine learning models to directly predict if a shift worker is at high risk of burnout by analyzing physiological and rhythm features. We analyzed the data of workers with low and high risks of burnout to understand the differences between their features and find markers indicative of burnout. The analysis showed that users with a low risk of burnout have better sleep regularity and lower anxiety. We also found that heart rate-related features and rhythm features are the most important for models to predict burnout. For all our models, we provided interpretability in relation to the input data that contributes to the prediction of well-being by analyzing the model’s feature importance and the correlation of the features with the activations of a convolutional neural network.