Harnessing consumer wearable digital biomarkers for individualized recognition of postpartum depression using the All of Us Research Program dataset

Abstract

Postpartum depression (PPD), afflicting one in seven women, poses a major challenge in maternal health. Existing approaches to detect PPD heavily depend on in-person postpartum visits, leading to cases of the condition being overlooked and untreated. We explored the potential of consumer wearable-derived digital biomarkers for PPD recognition to address this gap. Our study demonstrated that intra-individual machine learning (ML) models developed using these digital biomarkers can discern between pre-pregnancy, pregnancy, postpartum without depression, and postpartum with depression time periods (i.e., PPD diagnosis). When evaluating variable importance, calories burned from the basal metabolic rate (calories BMR) emerged as the digital biomarker most predictive of PPD. To confirm the specificity of our method, we demonstrated that models developed in women without PPD could not accurately classify the PPD-equivalent phase. Prior depression history did not alter model efficacy for PPD recognition. Furthermore, the individualized models demonstrated superior performance compared to a conventional cohort-based model for the detection of PPD, underscoring the effectiveness of our individualized ML approach. This work establishes consumer wearables as a promising avenue for PPD identification. More importantly, it also emphasizes the utility of individualized ML model methodology, potentially transforming early disease detection strategies.

Figure 1:. An overview of the analysis workflow to evaluate the potential for digital biomarkers in postpartum depression (PPD) recognition.

*RF = random forest, GLM = generalized linear models, SVM = support vector machine, KNN = k-nearest neighbors

Figure 2:

A schematic of PPD computational phenotyping.

Figure 3:. Individualized RF models exhibited the best performance for multinomial time period classification.

A: The multiclass AUC (mAUC) across individual random forest (RF), generalized linear model (GLM), support vector machine (SVM), and K-nearest neighbor (KNN) models. B: The kappa value across individual RF, GLM, SVM, and KNN models. C: The sensitivity, specificity, precision, recall, and F1 score across individualized multinomial RF models for the PPD time period. The average mAUC and kappa values of individualized RF models were higher than that of GLM, SVM, and KNN models. Utilizing individualized RF models resulted in an average sensitivity, specificity, precision, recall, and F1 score with good performance for recognizing the PPD time period. Data are expressed as mean ± standard deviation (SD) in A-C.

Figure 4:. ML models did not accurately detect the PPD-equivalent time period in women without PPD.

The sensitivity, specificity, precision, recall, and F1 score of individualized ML models in women without PPD for predicting the pre-pregnancy (top left), pregnancy (top right), postpartum (bottom left), and PPD-equivalent (bottom right) time periods. The sensitivity, specificity, precision, recall, and F1 score were diminished for recognizing the PPD time period compared to the pre-pregnancy, pregnancy, and postpartum time periods. Data are expressed as mean ± SD.

Figure 5:. Individualized ML models for PPD recognition outperformed those in women without PPD detecting the PPD-equivalent time period.

The sensitivity, specificity, precision, recall, and F1 score across individual RF models for women in the non-PPD cohort for the pre-pregnancy (top left), pregnancy (top right), postpartum (bottom left), and PPD or PPD-equivalent time periods (bottom right). On average, individualized model performance was not significantly different for sensitivity, specificity, precision, recall, and F1 score for predicting the pre-pregnancy, pregnancy, or postpartum time periods between women in the PPD and non-PPD cohorts. Individualized model performance was reduced for sensitivity, precision, recall, and F1 score, while specificity did not differ between the PPD and non-PPD cohorts. Data are expressed as mean ± SD.

Figure 6:. The variable importance rankings demonstrated that calories burned during the basal metabolic rate (calories BMR) is the most predictive digital biomarker of the PPD class.

A: The percentage of women with the top five digital biomarkers ranked among the top five most predictive features of the PPD class (top left) and the most predictive feature of the PPD class (top right) based on SHAP. B: The percent of women with the overall top five digital biomarkers ranked among the top five most predictive features of the PPD class (bottom left) and the most predictive feature of the PPD class (bottom right) based on a permutation-based approach of variable importance. Calories BMR, average HR, Q1 HR, lightly active minutes, and minimum HR most frequently emerged as the top five digital biomarkers with the highest predictive value for the PPD time period within the PPD cohort. Calories BMR most often ranked as the number one digital biomarker predictive of PPD.

Figure 7:. The direction of digital biomarkers in ML models for PPD classification was heterogeneous.

A: The percent of women in the PPD cohort with a significant Pearson correlation (left) and the net relationship (right) for the top five overall ranked digital biomarkers for PPD classification. B: The percent of women in the non-PPD cohort with a significant Pearson correlation (left) and the net relationship (right) for the top five overall ranked digital biomarkers for PPD-equivalent classification. The proportion of women displaying a significant Pearson correlation coefficient between SHAP values and digital biomarkers varied in both the PPD and non-PPD cohorts. Among those demonstrating a significant relationship in SHAP dependence plots during the pre-pregnancy/PPD (and pre-pregnancy/PPD-equivalent) time periods, the correlation pattern for SHAP values and calories BMR differed: the majority of women exhibited a positive correlation in the PPD cohort, while there was no uniform pattern amongst women in the PPD-negative cohort. For women in the pregnancy/PPD and postpartum/PPD (and PPD-equivalent) time periods, the majority of women demonstrated a negative relationship between SHAP values and calories BMR in both the PPD-positive and negative cohorts.

Figure 8:. Individualized ML models outperformed a cohort-based model for PPD recognition.

The sensitivity, specificity, precision, recall, and F1 score of individualized ML models in women in the PPD cohort detecting the PPD time period were compared to a cohort-based model for PPD classification. Data are expressed as mean ± SD. On average, the sensitivity, specificity, precision, recall, and F1 score were elevated in individualized ML models using digital biomarkers among women in the PPD cohort, specifically for the PPD time period, compared to a conventional binomial model designed for PPD or non-PPD classification using digital biomarkers.