Preterm birth (PTB) is the leading cause of child mortality under five years of age. While predictive models for PTB exist, effective, inexpensive interventions remain limited. Physical activity and sleep offer potential avenues for intervention, especially in low- and middle-income countries (LMICs). However, objective measurement of these factors is challenging, with self-reported data lacking accuracy and resolution. This study aimed to leverage wearable device data to develop a more accurate and interpretable model for pregnancy progression and identify associations between physical activity and sleep patterns with PTB risk. The hypothesis was that predictive models built from wearable data could offer a low-risk, cost-effective means of reducing PTB. The challenge lies in analyzing the continuous, long-length, and low-dimensionality wearable data to identify at-risk behaviors. Existing methods often rely on black-box toolboxes or limited non-parametric techniques. Disruptions to circadian rhythm are also known to impact birth timing and fetal/maternal health, but the role of physical activity and sleep changes during pregnancy remains largely unknown. This study proposed a new deep learning and inference pipeline, 'series2signal', to analyze wearable data, aiming to improve on existing approaches by incorporating data augmentation techniques, a novel deep learning architecture (inspired by ResNet), and an extensive post-hoc inferential and analysis pipeline for model interpretation. The study applied 'series2signal' to a cohort of pregnant women wearing wearable devices to discover relationships between deviations from typical sleep and activity patterns and PTB risk. The 'series2signal' pipeline was designed to combine prediction and explainable AI to identify activity and sleep behaviors associated with adverse pregnancy outcomes. The ultimate aim is to show the potential for using wearable data and computation to personalize risk assessment for PTB and enable cost-effective interventions.

Existing literature highlights the high global burden of preterm birth and the need for effective, low-cost interventions. Several studies have attempted to create predictive models using various data sources, but few have focused on readily available and modifiable factors such as physical activity and sleep. While wearable devices offer potential for objective measurement, their integration into clinical workflows remains limited due to analytical challenges. Previous studies have demonstrated the impact of circadian rhythm disruption on pregnancy outcomes, but extensive monitoring of sleep and activity patterns during pregnancy using wearable data has been lacking. The existing analytical methods for accelerometer data are often limited, using black-box commercial toolboxes, narrow applications, or non-parametric techniques with limited generalizability. This study aims to address these limitations by developing a comprehensive analytical pipeline that combines cutting-edge deep learning techniques with robust interpretability methods.

This study utilized a dataset from a cohort of N = 1083 pregnant individuals who wore wearable actigraphy devices throughout their pregnancy. The devices captured physical activity and motion data (acceleration across three dimensions) and light intensity at 1-minute intervals. Data were collected for one week following each gestational age (GA) measurement, resulting in N = 2305 data points. Electronic health record (EHR) data were also collected and included variables such as BMI, prior PTB history, comorbidities, and delivery outcomes. The researchers first analyzed the data using standard actigraphy methods, calculating metrics such as interdaily stability (IS), intradaily variability (IV), and relative amplitude (RA). They found no significant differences in these metrics between pregnancies that resulted in PTB and those that did not. Standard activity metrics such as the Kaiser Permanente Activity Scale (KPAS) and Pittsburgh Sleep Quality Index (PSQI) were also explored, yielding limited results. The core of the methodology involved the development and application of 'series2signal', a novel machine learning and explainable AI pipeline. This pipeline comprises four main components: (1) a machine learning pipeline that involves data pre-processing, a custom deep learning architecture (based on ResNet but modified for time-series analysis), and a novel data augmentation scheme; (2) model error analysis that identifies subgroups based on error modes and clinical variables; (3) a feature attribution module to interpret model predictions; and (4) a module for predicting subgroup membership and phenotyping based on the model's learned representations. The researchers' deep learning model was compared to seven other machine learning methods (kNN, RandomForest, Gradient Boosting, TimeSeriesForest, GRU, VGG-1D, Inception Time), demonstrating that 'series2signal' significantly outperformed the others in predicting GA from actigraphy data. To analyze model error, the researchers identified three error modes: higher-than-actual GA, lower-than-actual GA, and a low-error group. They used permutation testing to assess the association between these error groups and clinical variables, revealing significant differences in the prevalence of PTB across error modes. The model interpretability module, using gradient-based feature attribution, investigated the model's reliance on sleep and activity patterns in making predictions. Finally, they evaluated the utility of the model's learned representations for predicting other metadata variables and for phenotyping patients using time-series clustering.

The series2signal model significantly outperformed other machine learning algorithms in predicting gestational age (GA) from wearable actigraphy data (minimum average absolute error of 7.52 weeks, Spearman's ρ = 0.45, P < 0.001). The model's prediction error, when compared to actual GA, emerged as a strong predictor of preterm birth. Specifically: * When the model underestimated GA (lower-than-actual GA group), there were 0.52 fewer preterm births than expected (P = 1.01e−67, permutation test). * When the model overestimated GA (higher-than-actual GA group), there were 1.44 times more preterm births than expected (P = 2.82e−39, permutation test). * Model error was negatively correlated with interdaily stability (IS), indicating that less precise daily rhythms were associated with a higher predicted GA (P = 0.043, Spearman's). * The model attributed higher importance to sleep periods in predicting higher-than-actual GA compared to lower-than-actual GA (P = 1.01e−21, Mann–Whitney U). Model interpretability analysis revealed that the model relies more on sleep patterns than wake patterns when making predictions, particularly in the higher-than-actual GA error group. Feature importance scores were significantly higher during sleep periods in the higher-than-actual GA group compared to the lower-than-actual GA group (P < 0.001, Mann-Whitney U). The analysis also showed associations between model error and various clinical variables, such as hypertension, pregestational diabetes, prior PTB, and others. The model's learned representations demonstrated utility in predicting other clinically relevant variables, and unsupervised clustering of the embeddings revealed semantically meaningful clusters enriched for specific clinical characteristics. Analyses stratified by spontaneous vs. iatrogenic preterm birth demonstrated that 'series2signal' could be used to predict risk for both types of birth based on physical activity, sleep, and light exposure patterns.

This study demonstrates the potential of using wearable device data and deep learning to identify associations between physical activity and sleep patterns and preterm birth risk. The 'series2signal' model's ability to predict GA from wearable data and the strong association between model error and PTB risk provide compelling evidence for the clinical utility of this approach. The interpretability analysis shed light on the mechanisms driving these associations, highlighting the importance of sleep patterns in predicting deviations from expected GA. This suggests that interventions targeting sleep quality and possibly physical activity could potentially mitigate PTB risk. The model's performance on ancillary tasks and its ability to identify meaningful clusters from patient data further underscore its value for comprehensive pregnancy monitoring. The findings advocate for the development of clinical decision support systems that leverage passive monitoring of activity and sleep to identify at-risk individuals and promote timely interventions.

This study successfully developed and validated 'series2signal', a novel deep learning-based approach for analyzing wearable data to monitor pregnancy progression and identify risk factors for preterm birth. The model’s superior performance in predicting GA and its strong association with PTB risk highlight the potential for using wearable devices for personalized risk assessment and the development of effective interventions. Future research should focus on larger-scale validation studies and clinical trials to test the efficacy of interventions guided by the insights from 'series2signal'. Further exploration into disentangling the independent effects of sleep and physical activity disruptions is also warranted.

The study's limitations include the relatively small sample size (N = 1083 patients) and the potential for overfitting, despite implemented measures to mitigate this. The sleep detection method used was relatively simple, and more sophisticated approaches might provide further insights. The generalizability of the findings to different populations and clinical settings may be limited, and future work could explore this by including a more diverse population and using data from different wearable devices. The focus on one-week of data may be limiting, and exploration of shorter or longer periods could improve prediction and reduce the data needed for inference. The lack of direct intervention studies also restricts the conclusion to correlation rather than causation. Future work could involve randomized controlled trials to test the impact of interventions aimed at improving sleep and activity patterns on preterm birth rates.