JMIR Publications

ABSTRACT

Background:

Momentary affect intensity is central to many psychological processes and behaviors. Ubiquitous smartphone access and advances in artificial intelligence (AI) offer opportunities to continuously track affect intensity. Temporal Fusion Transformer (TFT) can (a) handle short (proximal effects) and long (distal effects) time dependencies between interdependent predictors and between predictors and outcome, testing multiple time lags simultaneously; and (b) model each individual’s data while borrowing group information.

Objective:

We examined how well TFTs predicted unseen momentary affect intensity without active inputs and used TFTs to reveal relations between affect and smartphone-tracked behaviors.

Methods:

Community adults (N = 102) reported emotional experiences five times a day for 14 days, with continuous collection of smartphone sensor data. Testing data (n = 408 surveys) was from the last four surveys for each individual, with the rest for training (n = 5264 surveys).

Results:

The best TFTs explained 40.7% variance in unseen negative affect (NA) and 38.0% variance in unseen positive affect (PA). Greater NA was associated with initiating more calls, greater N of people contacted, and weekdays. Greater PA was associated with more scattered locations, traveling further distances, and weekends. For PA, time spent at workplaces was an important distal predictor, whereas time spent at leisure was an important proximal predictor.

Conclusions:

Results suggest that after briefly collecting self-reported affect, it is possible to track affect intensity with behavior signals alone. NA seems to link with phone calls, and PA seems to link with physical mobility. We also provided preliminary evidence on the varying lagged effects (proximal vs. distal) of smartphone-tracked behaviors on momentary affect.