Wearable device non-wear detection methodology using dynamical features and machine learning techniques to improve sleep measurement (QuanDev Brownbag)

Google map

Purpose:

Actigraphy data from wrist motion sensors are often used in sleep research. Without an on-wrist sensor, actigraphy data are often clouded with non-wear data that compromise the quality of the data and confound sleep/wake classification. We hypothesized that a machine learning algorithm supplemented by dynamic features could discern wear/non-wear episodes.

Methods/Approach:

Data was collected from employed adults from two different companies. Wrist actigraphy measurements(Spectrum device, Philips-Respironics) including movement and light levels in 30-s epochs were collected for 1 week. Non-wear periods were identified with the built-in on-wrist capacitive sensor and verified by two trained coders. Lower-cost devices with similar or identical measurement characteristics do not have this on-wrist sensor. Extreme gradient boosting (XGBoost), a tree-based classifier algorithm, was used to classify the data as wear/non-wear. Algorithm accuracy was further increased by introducing dynamic features based on activity levels, such as entropy, maximum level shift, and spikes of the activity levels, which were calculated over various time windows. To better encapsulate the circadian rhythm in the data, we converted the timestamp data into sinusoidal curves. Accuracy was further improved by using k-fold validation to tune hyperparameters.

Results/Findings:

The XGBoost classifier yielded substantial improvements in accuracy, sensitivity, specificity, and related metrics (e.g., balanced accuracy, the area under the receiver operating characteristic curve; AUC) under the inclusion of dynamic features, and in comparison, to default actiwatch classification algorithms. Evaluation of the SHapley Additive exPlanations (SHAP) values suggested that the first derivative of the activity levels, the maximum level shift of activity levels, and the Hurst exponent were among the dynamic features that led to notable differences in non-wear classification. These dynamic features complemented other important features such as the presence of ambient light, time of the day, and activity levels.

Conclusion/Practical Implications:

Our findings highlight the potential of XGBoost using dynamic features of varying activity levels across the time series to provide valuable insights on wear/non-wear classification using a large dataset. The methodology provides an alternative to laborious manual benchmarking of the data for similar devices that do not have a non-wear sensor. Adding dynamical features increases computational and memory allocation but substantially benefits accuracy and specificity. Since there are few non-wear epochs at the individual participant level, the data is inherently imbalanced leading to lower specificity. Hence for future research, incorporating more dynamical features may lead to better differentiation between wear and non-wear epochs.