The findings of the included studies are summarized across 4 domains of perinatal mental health: perinatal depression, prenatal stress, emotional states during pregnancy, and maternal loneliness. These findings are presented alongside the analytical approaches used, which ranged from traditional inferential statistical models to supervised machine learning and meta-learning frameworks. The analytical approaches, evaluation metrics, and key findings of the included studies are summarized in Table 3. To support the interpretation of the reported performance metrics, this section also summarizes the model validation methods used in the end.

Faherty et al [33] used generalized linear mixed effects models to examine associations between movement patterns and both self-reported mood and depression or anxiety scores. They found that the radius of travel was significantly associated with depression severity. Women with more severe depression had a median daily radius of 1.9 miles versus 2.7 miles in milder cases (P=.04). A worsening of daily mood from the prior day was associated with a 5% contraction in radius (P=.03), and each PHQ-9 point corresponded to a 64% smaller radius around assessment days (P=.003).

Hurwitz et al [34] used individualized multinomial machine learning classification models to classify 4 perinatal periods: prepregnancy, pregnancy, postpartum without depression, and PPD. The individualized random forest models using Fitbit-derived features distinguished these 4 periods with a mean multiclass area under the curve (AUC; mAUC) of 0.85 and Cohen κ of 0.80, outperforming generalized linear models (mAUC=0.82), support vector machines (mAUC=0.75), and k-nearest neighbors (mAUC=0.74). For the PPD class specifically, the model achieved a sensitivity of 0.79, a specificity of 0.95, a precision of 0.84, and an F₁-score of 0.81.

Li et al [24] reported a weighted risk scoring framework combining self-report scales with app-derived phenotypic scores estimated using multivariable logistic regression to classify PPD risk as low, moderate, or high. At 3 weeks postpartum, the multimodal model achieved 90.0% sensitivity, 84.1% specificity, and an AUC of 0.871. The study also showed that the mobile health intervention, a smartphone app integrating continuous self-report with passively collected behavioral data to generate automated PPD risk alerts, shortened the mean time to PPD identification from 26.9 days under standard care (ie, scheduled screenings) to 11.8 days in the intervention group (P<.001).

Tang et al [22] used feature selection followed by univariable linear mixed models to predict EPDS scores and identify behavioral features associated with PPD symptoms. They identified 8 behavioral features significantly associated with weekly EPDS scores after false discovery rate correction. Baby crying count on weekdays (β=+.21; q<0.05) and maximum nighttime GPS stay duration (β=+.30; q<0.05) were positively associated with depressive symptom burden, while afternoon active bouts on weekdays (β=−0.24; q<0.05) and afternoon stationary bout count (β=−0.23; q<0.05) were negatively associated. The study further identified a potentially stress-inducing greater travel distance on weekday morning pattern (β=+.17; q<0.10), contrasting with the typically inverse relationship between mobility and depression observed in nonpostpartum populations [50,51].

Jose et al [23] used a meta-learning framework combining long short-term memory networks, convolutional neural networks, and gradient boosting models to classify PPD risk levels, reporting a predictive accuracy of 98.7%, the highest figure reported in the included literature. However, this result should be interpreted cautiously, as key methodological details, including the wearable devices used, the mental health outcome measures, and the validation procedures, were insufficiently reported in this conference paper.

Singh Solorzano et al [38] used Pearson correlation and hierarchical linear regression to examine whether prepartum RMSSD was associated with and predicted postpartum depressive symptoms. They found that lower prepartum RMSSD independently predicted depressive symptoms 1 month postpartum (β=−0.22; P=.01), after adjustment for prepartum depressive symptoms and potential confounders. These findings suggest that pregnancy may represent a potentially important window for early risk stratification and preventive monitoring.

King et al [36] developed the microstress EMA framework to identify a micro-EMA item aligned with physiological stress for labeling passively sensed data in pregnant women. Using laboratory data from 18 nonpregnant women, they trained stress classifiers on wearable ECG or interbeat interval–derived features before applying the model to in-the-wild recordings from pregnant women. “WorriedStress” showed the most promise as a short-window micro-EMA label. However, as the model was developed in a nonpregnant laboratory sample and evaluated in a separate pregnant field sample, the findings are best interpreted as proof-of-concept transfer across populations and contexts rather than as externally validated perinatal stress prediction.

Ng et al [35] used machine learning classification models to predict next-day physiological and perceived stress in a cohort of 16 pregnant individuals. Random forest classifiers achieved an F₁-score of 0.84 for physiological stress and 0.74 for perceived stress. A notable observation was that physiological and perceived stress appeared to rely on different predictive signals: when EMA features were excluded, performance remained comparable for physiological stress but was substantially reduced for perceived stress.

Li et al [39] used machine learning classification models to predict 4 discrete emotional states (happy, anxious, sad, and frustrated) among 53 pregnant women across gestational weeks 23 to 32. Among the 10 machine learning classifiers evaluated, random forest achieved the highest area under the receiver operating characteristic curve of 0.70 for multiclass emotion prediction.

Sarhaddi et al [37] used machine learning classification models to classify maternal social loneliness status during pregnancy and the postpartum period using objective sensor data. Gradient boosting achieved a weighted F₁-score of 0.897, and decision tree achieved a weighted F₁-score of 0.872 based on smartwatch-derived HRV, sleep, and physical activity data from 39 data samples contributed by 31 participants. Adding sleep features to physical activity features slightly reduced classification performance among participants classified as lonely.

Reported model performance should be interpreted in the context of the validation strategies used, as high AUC, F₁-score, or accuracy values in small datasets may reflect overfitting, nonindependent observations, or optimistic validation rather than generalizable performance.

Internal resampling was the main validation approach among studies that developed predictive classifiers. Hurwitz et al [34] used 3 repeats of 10-fold cross-validation within individualized models; each model was developed and tested using data from a single participant. Ng et al [35] used 5-fold cross-validation with folds randomly drawn from all participant-days combined, and Li et al [39] similarly randomly split pooled observations into training and testing sets under 5-fold cross-validation; in both studies, repeated observations from the same individual could therefore fall into both partitions. In contrast, Sarhaddi et al [37] used leave-one-participant-out cross-validation, keeping observations from the same individual entirely within either the training or the testing set. Jose et al [23] used k-fold cross-validation but did not report whether folds were separated at the participant level.

One study (King et al [36]) used separate-sample validation, but the design involved 2 simultaneous shifts rather than conventional external validation. They trained the model on laboratory-induced stress data from nonpregnant women and then applied it to in-the-wild recordings from a separate group of pregnant women. This should therefore be interpreted as a proof-of-concept transfer across both populations and contexts, rather than as external validation in an independent perinatal cohort or as participant-level internal validation.

No study performed external validation in an independent cohort. The only related analysis was conducted by Hurwitz et al [34], who trained individualized models in participants with PPD and used a non-PPD control group from the same dataset as a specificity check. Finally, no study assessed calibration; model performance was reported using discrimination and classification metrics only. Taken together, the high discrimination metrics reported above should be interpreted cautiously.

Three studies integrated post hoc explainable AI (XAI) routines alongside their predictive pipelines. XAI refers to AI systems that, given a target audience, provide details or reasons that make their functioning clearer or easier to understand [52]. In applied predictive modeling, commonly used XAI techniques often aim to improve the interpretability of complex or “black box” machine learning models by estimating how input features contribute to model outputs. Jose et al [23] embedded both Shapley Additive Explanations (SHAP) [53] and Local Interpretable Model-Agnostic Explanations [54] within their Context-Aware Adaptive Meta-Learning framework for PPD risk estimation. Ng et al [35] applied SHAP to their random forest models to provide both global feature importance rankings and local, instance-level explanations for stress prediction. Hurwitz et al [34] combined SHAP with a permutation-based variable importance method to rank candidate features associated with PPD across individualized random forest models, using SHAP dependence plots to visualize the direction and magnitude of each feature’s relationship with the PPD class.

Other studies used more limited model-intrinsic or statistical interpretability approaches. Sarhaddi et al [37] examined decision tree structures and Gini-based feature importance from gradient boosting models to identify smartwatch-derived predictors of maternal social loneliness. Li et al [39] ranked 9 HRV indicators using random forest feature importance for predicting 4 emotional states, reporting that performance plateaued after the top 5 features were included. Tang et al [22] applied regularization-based feature selection followed by linear mixed effects models to identify passive sensing features associated with EPDS scores. In 2 further studies, interpretability was based on classical statistical modeling: Faherty et al [33] applied generalized linear mixed effects regression to examine associations between mood and mobility features, and Li et al [24] used multivariable logistic regression to identify independent features of PPD within a mobile health digital phenotyping model.

Across the included studies, a range of features were reported as useful for detecting or predicting perinatal mental health outcomes. These features mainly fell into 2 categories: physiological features, derived predominantly from cardiac signals, such as HRV features; and behavioral and contextual features, such as physical activity, sleep, and device usage patterns. As the included studies differed substantially in devices, feature-engineering pipelines, outcomes, time windows, and modeling approaches, feature importance was not directly comparable across studies. We therefore summarized these 2 categories of features reported as useful within each study, together with the criteria used, in Table 4.

Across studies examining physiological features, HRV-related features were examined most frequently, reflecting ANS function and its association with depression risk and severity [18,55]. Li et al [39] reported that the CVRR, RMSSD, SDNN, HF, and LF power were among the most important HRV features in their random forest models for detecting maternal emotions. Singh Solorzano et al [38] found that lower prepartum RMSSD independently predicted higher postpartum depressive symptoms after adjustment for baseline depression and covariates. In models of maternal social loneliness, Sarhaddi et al [37] highlighted resting HR, resting SDNN, LF, ratio of LF-to-HF power (LF/HF), and average NN intervals as key contributors alongside activity-related features. In individualized PPD prediction models, Hurwitz et al [34] identified average HR, first-quartile HR, and minimum HR as key contributors.

Behavioral and contextual features derived from smartphones and wearables were also frequently reported as useful, particularly those capturing activity, mobility, sleep, and daily routines. Sarhaddi et al [37] identified physical activity intensity and the kurtosis of hourly step counts as the most important features in their models of maternal social loneliness, with additional contributions from distributional statistics of step count, activity duration, and sedentary time. Faherty et al [33] found that the radius of travel was most sensitive to mood, with worsening mood and higher PHQ-9 scores associated with reduced mobility range. Tang et al [22] reported that infant-related sleep disruptions, increased morning crying events, longer nighttime stay duration, reduced afternoon activity, and altered mobility patterns were associated with higher EPDS scores.

Perinatal mental health conditions affect approximately 20% of pregnant and postpartum individuals and are associated with a broad spectrum of adverse outcomes, including significant functional impairment, increased risk of chronic psychiatric disorders, and increased suicidality. These conditions can also disrupt mother-infant bonding and are linked to adverse developmental outcomes in offspring. The scale and severity of these impacts, together with persistent barriers to timely care during the perinatal period, underscore the potential value of continuous, real-world monitoring using mobile sensing technologies.

This systematic review examined the use of smartphones and wearable devices to detect and predict perinatal mental health outcomes. Overall, the findings suggest that mobile sensing shows preliminary potential for perinatal mental health assessment, but the current evidence base remains limited by narrow outcome coverage, restricted technological approaches, and substantial methodological weaknesses.

This section first synthesizes 3 core observations emerging from the included studies and then situates the current evidence base within the broader perinatal mental health and digital phenotyping literature to characterize the scope of mental health outcomes examined and the technological approaches represented to date.

The included studies converge on 3 key observations. First, ANS-related physiological features, particularly HR and HRV features, were frequently reported as important model contributors within individual studies. Li et al [39] reported that CVRR, RMSSD, SDNN, HF, and LF contributed to the detection of maternal emotions. Sarhaddi et al [37] highlighted resting HR, resting SDNN, LF, LF/HF, and average NN intervals as key contributors to maternal loneliness prediction. Hurwitz et al [34] identified average HR, first-quartile HR, and minimum HR as key contributors in individualized PPD prediction models. The relevance between these autonomic features and mental health is consistent with the neurovisceral integration model, which links vagally mediated HRV to the functional integrity of prefrontal-subcortical circuits involved in emotion regulation [56].

Notably, Singh Solorzano et al [38] found that lower prepartum RMSSD independently predicted higher depressive symptoms at 1 month postpartum after adjustment for baseline depression and covariates. This finding is supported by evidence from nonperinatal populations showing that reduced vagal tone is associated with depression, anxiety, and impaired stress reactivity [18,55]. This finding further suggests that ANS activity during pregnancy may provide information about subsequent postpartum mood outcomes, rather than merely reflecting concurrent emotional states. One possible explanation is that pregnancy is characterized by substantial autonomic adaptation to support maternal-fetal hemodynamic demands [57] and that disrupted autonomic adaptation or reduced parasympathetic reserve may be expressed both as altered HRV during pregnancy and as greater vulnerability to postpartum mood symptoms.

Second, behavioral and contextual features were also reported as useful for detecting and predicting perinatal mental health outcomes, and their interpretation appears to require context-specific consideration within perinatal populations rather than direct extrapolation from general population findings. Sarhaddi et al [37] identified physical activity intensity and hourly step-count kurtosis as key features for maternal social loneliness. Faherty et al [33] found reduced radius of travel to be associated with worsening mood and higher PHQ-9 scores. Tang et al [22] reported that infant-related sleep disruptions, increased morning crying events, longer nighttime stay duration, reduced afternoon activity, and altered mobility patterns were associated with higher EPDS scores. These findings further indicate that behavioral features may provide meaningful information about mental health [7].

In nonperinatal adults, reduced mobility, increased home dwell time, and disrupted circadian patterns are well-established correlates of depressive symptoms [50,51]. Faherty et al [33] replicated this pattern, with a reduced radius of travel tracking worsening mood in pregnant women. However, Tang et al [22] identified an inverse association during the postpartum period: greater weekday-morning travel distance was associated with higher EPDS scores. This divergence likely reflects the distinct behavioral context of early postpartum life, in which increased mobility may indicate infant care demands, medical visits, or disrupted routines rather than behavioral activation. Similarly, longer nighttime GPS stay duration, typically indicative of healthy rest in general populations, was associated with higher symptom burden [22], possibly reflecting prolonged sedentary awakenings related to infant care. These findings further emphasize the need to interpret passive sensing features within context, rather than assuming that feature-outcome relationships generalize across life stages [58].

Third, the limited available evidence suggests that multimodal and individualized modeling strategies may offer advantages over population-level, single-modality approaches. On the multimodal side, the 2 highest-performing PPD models in this review both integrated heterogeneous inputs: Jose et al [23] integrated passive sensing data (HR, HRV, sleep quality, and physical activity) with other app-derived data (social interaction logs; collection method not reported) and clinical data (prenatal, delivery, and postnatal examination records) within a meta-learning framework, and Li et al [24] combined passive sensing data (phone usage features), other app-derived data (voice features extracted from voice diaries), and active self-report data (weekly self-reported scales and daily diaries on mood, sleep, breastfeeding, and infant interaction). These patterns indicate that stronger reported performance may be related to the incorporation of data sources beyond passive sensing alone while also highlighting potential trade-offs between predictive performance and implementation demands across data source categories. Different data source categories have distinct implementation implications. Passive sensing can reduce deliberate input, but background collection can still involve technical burden, device-dependent or operating system–dependent missingness, and privacy concerns related to behavioral traces [59]. Active self-report inputs may improve predictive performance but reintroduce participant burden, particularly in the perinatal period, when sleep disruption and caregiving demands may limit engagement with self-monitoring [60]. Other app-derived data, such as social interaction logs and voice-derived features, may capture social or affective functioning but depend on app engagement and raise privacy concerns if they reveal identity, relationships, or sensitive affective information [61]. Clinical data may add contextual information, but their availability and completeness depend on documentation practices, access to care, and record linkage, which may introduce selection bias and limit scalability [62].

On the personalization side, Hurwitz et al [34] reported strong discrimination of PPD using individualized random forest models under internal validation, consistent with findings in nonperinatal populations that within-person deviations may be more informative than between-person comparisons [63]. Although only one included study implemented an individualized approach, the perinatal period, with its rapid within-person physiological and behavioral changes, is a context in which personalized modeling is likely to offer substantial advantages. Its limited use, therefore, represents a clear methodological gap.

Beyond these methodological observations, placing the current evidence base within the broader perinatal mental health and digital phenotyping literature highlights additional gaps in what has been studied. Two are particularly notable: the limited range of mental health outcomes examined and the restricted use of available sensing and analytical technologies.

The 10 included studies converged on a narrow set of outcomes. PPD dominated the evidence base (6 studies), followed by stress (2 studies), discrete emotions (1 study), and maternal social loneliness (1 study); anxiety appeared only as a co-tracked outcome rather than a primary target. This pattern reflects the longstanding clinical and research emphasis on PPD, while leaving substantial areas of perinatal psychopathology largely unexamined within the digital sensing literature.

Several clinically important conditions are notably underrepresented. Perinatal anxiety disorders affect 15% to 20% of pregnant and postpartum individuals [64,65], a prevalence comparable to depression, but were not examined as a primary outcome in any of the included studies. This represents a missed opportunity rather than a methodological barrier: anxiety is associated with physiological and behavioral features, such as elevated sympathetic tone, fragmented sleep, and nighttime phone use, that can be captured by current passive sensing pipelines and have already been studied in nonperinatal populations [12]. Fear of childbirth is another increasingly recognized perinatal concern, associated with obstetric interventions, such as cesarean section, poorer birth experience, and elevated risks of postpartum depression, anxiety, and posttraumatic stress disorder [66], yet none of the reviewed studies targeted this outcome.

Beyond these, the evidence was also silent on trauma-related disorders, particularly childbirth-related posttraumatic stress disorder [67], as well as perinatal substance and alcohol use disorders [68,69], perinatal obsessive-compulsive disorder [70], bipolar disorder [71], and eating disorders [72]. Moreover, almost all studies modeled outcomes in isolation, rather than addressing the high comorbidity that characterizes perinatal psychopathology.

The technological choices made across the included studies reflect only a narrow slice of the methods now increasingly used in digital mental health and wearable biosignal processing. In terms of signal modality, the majority of studies relied on traditional physiological and behavioral features, such as HRV, mobility, sleep, and screen use, with Li et al [24] being the only study to incorporate voice-derived features. In terms of modeling approach, analyses were dominated by tree-based ensembles and classical regression. In terms of training paradigm, most models were trained from scratch on relatively small, study-specific datasets, without using pretrained representations from external biosignal corpora.

These choices leave several gaps relative to the methodological possibilities now available in the broader field. In terms of modality coverage, voice and language have become well-established streams in digital mental health phenotyping. Acoustic features, such as speaking rate, pitch variability, and pause frequency, as well as linguistic content from free text, are particularly informative for anxiety and trauma-related symptoms [73,74]. In terms of temporal modeling, the perinatal period is characterized by pronounced and sustained change, including autonomic adaptation, progressive sleep disruption, and evolving caregiving demands. However, models based on aggregated weekly or daily features and applied within cross-sectional or short-window classification frameworks can capture this temporal structure only to a limited extent [63,75]; the most clinically relevant signal may lie in the trajectory of change rather than in any single observation window. In terms of data scale and training paradigm, perinatal cohorts are inherently difficult to recruit and retain in sufficient numbers for end-to-end supervised learning [76], and the resulting small sample sizes raise well-documented risks of overfitting and unstable performance estimates [77,78]. This represents a structural mismatch rather than merely a resource limitation, particularly as the field moves beyond feasibility demonstration toward generalizable clinical utility.

The field remains at an early stage of development: only 10 studies met the inclusion criteria, and several methodological constraints limit the strength of the current evidence. First, sample sizes were consistently small relative to model complexity, with 5 of 10 studies enrolling fewer than 40 participants [22,33,35-37], raising risks of overfitting and optimistic performance estimates [77,78].

Second, validation designs were generally limited and susceptible to optimistic performance estimates. External validation was largely absent; no study validated a model in an independent perinatal cohort. Given that health AI models may perform worse when applied to populations that differ in cultural, caregiving, and health care contexts, this limits confidence that the reported models would generalize across heterogeneous perinatal populations, care settings, or stages of pregnancy and postpartum.

Third, observation windows were often short, with several studies collecting data for 4 weeks or less. As perinatal mental health follows trajectory-based onset rather than static states [79], such short windows limit inference about precisely the temporal dynamics digital sensing is best placed to capture.

Fourth, methodological reporting was often insufficient to evaluate model quality: handling of missing data, hyperparameter tuning, and validation strategies were frequently underspecified, and few studies followed standardized frameworks such as TRIPOD+AI (Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis With AI) [80]. This was reflected in the PROBAST+AI assessment, where 80% (12/15) of assessment units were rated at high overall quality concern or risk of bias, driven primarily by the analysis domain, which further limits confidence in the reported performance.

Fifth, although most studies used validated scales to measure self-report outcomes, 2 studies used nonvalidated outcome measures; their performance estimates should therefore be interpreted cautiously and not as evidence of prediction against established clinical or psychometric criteria.

Several limitations of the review process should also be acknowledged. First, substantial heterogeneity across devices, derived features, analytical pipelines, and outcome definitions precluded meta-analysis, necessitating a narrative synthesis. As a result, we could summarize patterns across studies but could not provide pooled estimates of model performance or formally quantify between-study differences.

Second, our search strategy may have introduced bias: searches in databases were restricted to title and abstract fields, which may have missed studies in which perinatal status, mental health outcomes, or sensing methods were described only in keywords or full text. We also did not include gray literature, preprints, or unpublished trials, and restricting the search to English-language publications may have excluded relevant studies and introduced language bias. Given that only 10 studies were included, even a small number of missed studies could have affected the synthesis and the strength of the conclusions.

Third, data extraction was checked by a second reviewer for a randomly selected 40% (4/10) of the included studies rather than for all studies. Although agreement was high and discrepancies were minor and nonsubstantive, full independent verification was not conducted for every study; therefore, a small risk of undetected extraction errors cannot be entirely excluded.