Daytime alertness quantification and modelling: results from a large observational study

1. OBJECTIVE

Subjective alertness variations throughout the day can be characterized using the two-process model (TPM) of sleep regulation, which combines sleep homeostasis and the circadian rhythm to derive a theoretical daytime alertness curve. The TPM has been adopted to model the effect of sleep deprivation on memory, circadian misalignment, temperature regulation, and brain function; however, despite its broad influence, evidence supporting the TPM-derived alertness comes largely from small-scale, controlled studies. Here, we show that a similar three-parameter alertness measure can scale to a large study sample under real-world conditions.

2. HIGH LEVEL STUDY SETUP

Subjective alertness was voluntarily rated on a scale from 1 to 10 by Sleep Number® smart bed users (N=22 499) through the SleepIQ® app. Three age groups (18–40, 41–65, and 66–90 years) were analyzed. A three-parameter version of the TPM-derived alertness curve was fit to the self-rated alertness responses using nonlinear least-squares fitting.

3. RESULTS

A total of 65,528 sleep sessions were gathered over 95 days and analyzed. Overall, subjective alertness followed a similar trend to that reported in published literature: mean hourly alertness increased in the morning, dipped slightly in the afternoon, increased during the evening, and dropped again during the night. In contrast to previous studies, mean alertness ratings only changed by approximately 1 unit from low to high, and a greater increase in alertness occurred from afternoon to evening. Age-group analyses found that youngest sleepers’ mean daily alertness was more stable throughout the day, and the amplitude of alertness variation decreased with age. These experimental results showed high agreement with model prediction (R2=0.96, P<0.001) and that the TPM-derived alertness can effectively predict daily alertness trends in a large sample under real-world conditions.

DOWNLOAD PRESENTATION

Unobtrusive Sensing of Skin Temperature During Sleep Using a Mattress Sensor

1. OBJECTIVE

Sleep is associated with temperature changes in the human body. At sleep onset, distal (hands and feet) and proximal (abdomen) temperatures increase by ~1°C and ~0.5°C, respectively, forming a distal-to-proximal gradient that increases throughout the first half of a night’s sleep. However, core temperature decreases during this period. Few devices can measure these temperatures unobtrusively. Our aim was to estimate distal skin temperature unobtrusively during sleep using a temperature sensor array on a mattress.    

2. HIGH LEVEL STUDY SETUP

Skin temperatures were measured using an array of five equally spaced thermistors distributed laterally across the mattress region aligning with the torso. Participants wore an Empatica smart watch to provide benchmark distal skin temperatures. Data from 249 sleep hours were used to build predictive models estimating distal skin temperature. The preprocessed data were grouped by participant and segmented into training and test sets (~60%/40%, respectively), with earlier sleep sessions selected for training and later sessions selected for testing. Using an automated machine learning (AutoML) software, a model was developed that optimized R2. This model was applied to the test set and its performance evaluated by Bland-Altman analysis, using predicted and benchmark distal temperatures. 

3. RESULTS

The AutoML selected an XGBoost decision-tree model to predict distal skin temperature for each minute. The mean difference in temperature between predicted and benchmark readings was 0.13°C (R2=0.35), with lower and upper limits of agreement (LOA) of -1.59 and 1.84, respectively. Next, all minute-level data were averaged by sleep session, and model performance was re-evaluated across all sleep sessions. This approach resulted in smaller LOAs (-0.58, 0.91), with a mean difference in temperature of 0.16°C (R2=0.73).   

DOWNLOAD PRESENTATION