Development of digital measures for nighttime scratch and sleep using wrist-worn wearable devices

Pruritus (itch), a primary symptom in atopic dermatitis (AD), often leads to scratching, inflammation, and sleep disruption, significantly impacting quality of life. Traditional assessments rely on subjective clinical outcome assessments (COAs) and patient-reported outcome assessments (PROs), which have limitations in objectivity and capturing daily fluctuations. Wearable sensors offer a potential solution for objective, continuous monitoring of nighttime scratching and sleep. While previous studies have explored using wrist-worn accelerometers for sleep and scratch detection, many lack the capacity for passive, unsupervised monitoring during daily life, relying on simulated scratching or failing to segment data into sleep periods, potentially increasing false positives. This study aimed to develop a validated method combining both sleep and nighttime scratch detection for real-world application using a hierarchical approach, leveraging machine learning for complex tasks and heuristic rules for simpler ones.

Existing methods for measuring nocturnal scratching and sleep primarily rely on subjective patient reporting or laboratory-based polysomnography (PSG). COAs, such as assessing total body surface area (BSA) and lesion severity, provide limited insight into symptom fluctuations outside clinical settings. PROs, while offering patient perspective, are subjective and prone to bias. Previous attempts at objective measurement using accelerometers have shown promise, but limitations include reliance on simulated scratching movements, lack of continuous measurement capabilities, and absence of automatic sleep detection. Studies utilizing machine learning techniques like k-means clustering and logistic regression achieved high sensitivity in controlled settings, but free-living performance might suffer. Recurrent Neural Networks (RNNs) have been used for continuous measurement but lack interpretability. While some mobile applications leverage heuristic scratch detection, validation often involves small sample sizes or lacks automatic sleep detection. This research aimed to address these limitations by developing a robust and validated method suitable for continuous, unsupervised monitoring.

Forty-five AD patients participated in a study involving two in-clinic nights and two nights at home. In-clinic monitoring included infrared thermal videography for scratch annotation and, during the second night, limited polysomnography (PSG) for sleep assessment. Participants wore two GeneActiv Original wrist-worn devices, capturing triaxial acceleration, ambient light, and near-body temperature data. The analytical approach comprised a hierarchical paradigm: (1) context detection (wear detection, total sleep opportunity (TSO) detection, hand movement detection); and (2) symptom severity estimation (sleep quantity, nighttime scratch detection). The sleep module employed a heuristic approach for TSO identification, using arm angle changes and near-body temperature to exclude non-wear periods. A heuristic algorithm, modified to use an open-source activity index, then classified sleep/wake states in 1-minute epochs. The scratch module used a binary machine learning (ML) classifier trained on 3-second accelerometer data windows within the TSO. A heuristic hand movement detector pre-filtered data, and 36 time and frequency-domain features were extracted. Feature selection used recursive feature elimination with cross-validation. A random forest classifier (50 estimators) was trained using leave-one-subject-out validation. Epoch-level and summary endpoint performance were evaluated against reference data using metrics such as accuracy, sensitivity, specificity, F1 score, AUC, Pearson correlation, Bland-Altman plots, and SHAP analysis.

Of the 45 recruited patients, 33 were included in the analysis. Sleep-state detection achieved high sensitivity (0.95) and F1 scores (0.90) for both wrists compared to PSG, although specificity was lower (0.44). Scratch detection yielded sensitivity of 0.61 and specificity of 0.80. SHAP analysis revealed that signal periodicity and smoothness were important features for scratch detection. Total sleep opportunity (TSO) and total sleep time (TST) derived from the sensor data showed strong correlations with PSG data (TSO: r = 0.72, p < 0.001; TST: r = 0.76, p < 0.001). Log-transformed total scratching duration showed strong agreement with video annotations (r = 0.82, p < 0.001). The sleep module underestimated TSO by 29.7 minutes on average and overestimated TST by 24.2 minutes. A strong correlation (r=0.9, p<0.001) was observed between scratch events and wake after sleep onset (WASO), while there was no significant association between scratch and total sleep time (TST).

The study successfully developed and validated a method for continuously and objectively measuring nighttime scratching and sleep in AD patients using wrist-worn accelerometers. The hierarchical approach, combining heuristic rules and machine learning, proved effective in achieving high correlations with reference standards, supporting the use of this method for real-world clinical trials. The strong correlation between scratch duration and video annotations demonstrates the accuracy of the scratch detection algorithm. While the sleep module showed slightly less precise estimations of TSO, this is consistent with challenges in defining TSO boundaries, and the correlation remains significant. The strong association between scratch and WASO suggests that increased scratching leads to more disturbed sleep, rather than shorter total sleep duration. The interpretability of the ML model, unlike RNN-based approaches, offers advantages in regulated environments. Future work could explore other sensor modalities to improve low-intensity scratch detection and further investigate the relationship between objective scratching measures and subjective itch perception.

This study presents a novel, validated method for continuous and objective assessment of nighttime scratching and sleep using wrist-worn accelerometers. The hierarchical approach, integrating heuristic and machine learning algorithms, demonstrates strong agreement with reference measures. This method facilitates large-scale deployment in clinical studies, enabling more precise monitoring of AD symptoms and intervention efficacy. Future research should focus on incorporating additional sensing modalities to capture lower-intensity scratching and on investigating the longitudinal relationship between these digital biomarkers and disease progression.

The study's relatively small sample size might limit the generalizability of the findings. The reliance on near-body temperature for non-wear detection could limit applicability to devices without this capability. Although the algorithm achieved good accuracy, low-intensity scratching events might be missed. The study primarily focused on detecting scratching events rather than assessing the intensity or severity of scratching, which may have a different relationship with disease progression. The analysis is limited to the context of the specified study design and may not fully capture diverse scratching behaviors in all AD patient populations.