Prediction of stress levels with LSTM and passive mobile sensors

Abstract

Stress levels among people rise through the years, and passive sensing data from mobile phones or other ubiquitous devices have started to found its place in applications of mental health observation. With the ultimate goal of creating an automatic human mental health assistant that helps people to have a better mental condition, a step is taken by creating a stress recognition model. In previous works, the researchers have found correlations between sensor data and mental health conditions. They attempted to predict the stress level by different data sources. Due to there is no direct link between any sensor data with mental health, Machine Learning algorithms are employed to uncover relations with multiple sensors and mental well-being. The utilized machine learning algorithms work with non-sequence data hence the researchers need to extract features that represent historical sensor data as one value features. However, extracted features cannot completely represent time-varying sequence data. Within the scope of this study, we demonstrate that LSTM, CNN and CNN-LSTM algorithms, which accept sequence data as input can also work in passive mobile phone sensor data to predict human mental stress. The performance of the model on StudentLife dataset that includes passive mobile sensing data and stress feedbacks of college students has 62.83% accuracy on 460 test instances by training with 800 instances with LSTM model. The diversity of the data is limited and the size is tiny for the data-hungry LSTM model. Consequently, the model could not generalize on adapted features with the small sample size. Although we did not adapt complex features, the results encourage us to improve data size and continue to research on this topic.

Cite this work

No public copy is stored for this record, so only the metadata is cited.

Stored only when both papers are in the corpus — 37 citations exist upstream.

Papers that cite the same references — the relational fallback while embeddings are still backfilling.