Fig 1.
A) MOVES data set showing different exercise activities over time, with each color representing a different activity. B) Weight over time from the Withings data set. C) An example of human behavioral data, the use of SeeClickFix, over time. D) Various neighborhoods in New Haven and their crime data over time. E) Spatial map of New Haven illustrating the use of SeeClickFix and the associated geographical data. Open source map shapes can be found at https://rdrr.io/github/CT-Data-Haven/cwi/man/neighborhood_shapes.html. F) Various measurements taken from an indoor air quality sensor averaged over a day across a month.
Fig 2.
Schematic and illustration of the Bayesian structural time series.
Latent parameters and hyperparameters are shown in blue, observations are shown in green, covariates in yellow (with a property of posterior inclusion probability), coefficients in orange, and predictions in purple. Predictions have an associated credible interval shown in light purple. An illustrative example shows weight over time, with various covariates, being modeled. Covariate posterior inclusion probability is given by the size of the circle. Assuming an intervention of increased diet, the model detects a strong impact on the post-intervention weight.
Fig 3.
Performance of the Bayesian structural time series model in model experiments with known interventions.
A) Using an indoor air quality sensor, CO2 is measured with a variety of other covariates. The intervention (arrival of office members) causes an increase in CO2 and is determined to be impactful using the model. B) Accelerometer measurements of a person spinning in a chair holding sensor near body and then extended to arm’s length at the intervention; C) with simulated “noise” produced by a hop during the intervention period and D) with a paired covariate (second sensor) that is not affected by the intervention but experiences the “noise” hop. Sensitivity analysis shows comparison of mhealthCI to ARIMA, with the vertical axis as the fraction of the intervention that correctly identified a non-zero impact.
Fig 4.
A) Continuous glucose, estimated insulin on board (IOB), and Apple watch data, including heart rate variability (HRV), daily steps taken, and energy expended, over 12 weeks for the participant of interest. B) Analysis of a 10-week exercise regimen’s causal impact on the percentage of daily glucose readings in the target range (percent-in-target) of individual 1. C) Analysis of a 10-week exercise regimen’s causal impact on the percentage of daily glucose readings above the target range (percent-above-target) for individual 1. HRV, heart rate variability. IOB, insulin on board. D) Percent-in-target range analysis for individual 2. E) Percent-below-target range analysis for individual 2.
Fig 5.
A) Impact of intervention (use of SeeClickFix) on all of New Haven. B) Impact of intervention on only one neighborhood, Wooster Square. C) Spatial data of New Haven showing variety of SeeClickFix usage throughout various neighborhoods. D) Impact of intervention on Wooster Square crime using West River as a paired covariate. E) Illustration showing the observed data, Wooster Square, versus the paired covariate, West River. All open source map shapes can be found at https://rdrr.io/github/CT-Data-Haven/cwi/man/neighborhood_shapes.html.