Using explainable artificial intelligence to predict sleep interruptions from indoor environmental conditions: An empirical study

Research has proven that the ideal thermal comfort parameters for sleep differ from those for awake conditions. However, predicting thermal comfort for sleep is challenging, especially studies that permit subjects to perform normal adaptive behaviour such as choosing their own sleepwear and adding or removing blankets. Therefore, this study uses empirical data to predict subjects' sleep interruptions from indoor environmental quality (IEQ) conditions. We monitored 15 human subjects in their own homes over 378 total person-nights, under their preferred sleeping conditions, recording asleep, awake, and restless periods, via wristband fitness trackers. We simultaneously monitored indoor environmental variables including dry-bulb temperature, relative humidity, sound pressure levels, and carbon dioxide concentrations. By using explainable Artificial Intelligence (XAI), specifically, the XGBoost model, the study revealed that CO2 levels and heat index demonstrate the most significant association with sleep classification. Within the observed conditions of 16–25°C (with most observations falling within 21–23°C), an increase of 1.4°C in the average temperature and a 2–6% fluctuation in relative humidity tended to increase restlessness in the subjects. When temperature fluctuations exceeded 60% relative to the mean temperature, these fluctuations were correlated with a significant 50% reduction in sleep efficiency.