Personal comfort models based on a 6-month experiment using environmental parameters and wearables

Personal comfort models ?challenge the?one-size-fits-all?approach of thermal comfort models like the PMV and the adaptive. Instead of an average response from a group of people, a single model is trained and tested for each participant. Nevertheless, their aggregated outputs can still be used to predict the thermal preference of a large group of people sharing the same environment.

Personal comfort models are flexible in their needed input and can leverage data collected using a wide array of environmental and onboard sensors in wearable devices and smartphones. However, collecting feedback from participants is one of the major hurdles to developing new models. To solve the issue, we contributed to creating a seamless data collection method (ecological momentary assessment ), named Cozie.

Cozie is an open-source application that one can install on Fitbit (Versa 2 and Ionic) or Apple smartwatches. More information at?https://cozie.app ?and?https://cozie-apple.com .

Cozie allows people to complete a Right-Here-Right-Now survey via their smartwatches conveniently. Subjects' perceptions, preferences, and behaviors collected via Cozie can then be coupled with environmental data collected from wireless sensing devices and physiological data collected by the smartwatch.

Methodology

We conducted a longitudinal field study comprising 20 participants who answered Right-Here-Right-Now surveys using a smartwatch for 180 days. We collected more than 1080 field-based surveys per participant. We collected the data using the system depicted in the Figure below.

Surveys were matched with environmental and physiological measured variables collected indoors in their homes and offices. We then trained and tested seven machine learning models per participant to predict their thermal preferences. The methodology we used to clean the data, train and test the models is shown below.

Results

Participants indicated 58% of the time to want?no change?in their thermal environment despite completing 75% of these surveys at temperatures higher than 26.6°C.?

All but one personal comfort model had a median prediction accuracy of 0.78 (F1-score). Indoor air temperature, wrist near body temperature, heart rate, wrist skin temperature, and humidity ratio indoors had the highest average marginal contribution to the prediction accuracy.

We found that ≈250–300 data points per participant were needed for accurate prediction. We, however, identified strategies to reduce this number significantly. The Figure below shows the model's accuracy as a function of the number of training data points. Each line represents a participant.

Conclusions

We were able to determine that:

• Cozie, an open-source Fitbit and Apple application, is a reliable and robust solution to non-intrusively collect participants’ feedback in field studies.
• Personal comfort models can accurately predict (median F1-micro score 0.78) occupants’ thermal preferences. Personal comfort models outperform general models like the PMV but they require more data, in?another work ?we showed how the needed data could be substantially reduced.
• Indoor air temperature, wrist near body temperature, heart rate, wrist skin temperature, and humidity ratio indoors, listed in decreasing order of importance, had the highest contribution to prediction accuracy.
• The thermal personal comfort model prediction accuracy (F1-micro) plateaued at around 300 data points across all participants. Individual personal models are sensitive to dataset size to varying degrees. The amount of data required to characterize thermal comfort could be reduced with the development of?targeted sampling , which strategically requests feedback only when it is necessary.
• We made?available publicly ?the data we collected and open-sourced the Python code we used to analyze them to enable other researchers to test different hypotheses utilizing our data.

Reference

Tartarini F, Schiavon S, Quintana M, Miller C. Personal comfort models based on a 6-month experiment using environmental parameters and data from wearables . Indoor Air 2022; 32:e13160. doi:10.1111/ina.13160

GitHub rep

Acknowledgment

This research is funded by the Republic of Singapore’s National Research Foundation through a grant to the Berkeley Education Alliance for Research in Singapore (BEARS) for the Singapore-Berkeley Building Efficiency and Sustainability in the Tropics (SinBerBEST) Program. BEARS has been established by the University of California, Berkeley as a centre for intellectual excellence in research and education in Singapore.