Postpartum depression (PPD) can occur in women after childbirth for up to one year, affecting around 15% of mothers, and is the most common complication of childbirth (1). Neonatal intensive care unit (NICU) admission is a stressful experience for parents and together with prematurity are well known risk factors for PPD. Multiple studies have determined that the incidence of PPD in parents whose infants are admitted to the NICU is approximately 40–45%, which is considerably higher than the 15% risk among general birthing population (2, 3). Therefore, early detection of PPD at critical times during admission and discharge through screening programs can play a significant role in preventing the negative consequences for the family and child (4). Given the importance of early diagnosis of depression symptoms, multiple NICUs have developed and implemented screening programs for PPD in the NICU (4–6). Early identification of depression symptoms in parents is crucial to mitigate adverse effects such as infant neurodevelopmental delays (7). However, the current screening process is both expensive and time-consuming, requiring a tracking system that could span multiple healthcare settings.

Having a predictive model to identify parents at risk of developing postpartum depression can assist in prioritizing those in need of screening. Prior research has focused on training machine learning (ML) models to predict postpartum depression (8–19). A review of these studies revealed several significant predictors, including age, education, marital status, income, ethnicity, lifetime depression, depression during pregnancy, anxiety, smoking, mode of delivery, gestational age, APGAR score (appearance, pulse, grimace, activity, and respiration), BMI (body mass index), and history of antidepressant use (10). Although ML models have been used to predict postpartum depression, no study has applied ML to predict postpartum depression of NICU parents. There is only one study that has utilized logistic regression to investigate the risk factors associated with parental depression symptoms at NICU (20) and found that higher levels of parental stress, older gestational age, and lower levels of social support contribute to parental depression symptoms at NICU (20). However, it is worth noting that this study has yet to present performance metrics for the model, which are essential for facilitating a comprehensive comparison with other predictive models. Notably, the study also did not employ segmentation of the data into training and testing sets, a practice pivotal for evaluating the model’s performance using unseen testing data. The absence of such data partitioning is a common feature of preliminary investigations which may raise concerns about the model’s ability to generalize beyond its training data.

This study contributes to the existing literature in three distinct ways. Firstly, it pioneers the application of machine learning (ML) approaches to NICU data to comprehensively investigate factors predicting postpartum depression among NICU parents. A central objective is to discern whether ML methodologies surpass the predictive capabilities of the traditional Logistic Regression (LR) model. Secondly, the study employs rigorous methodology by dividing the dataset into distinct training and testing sets. Multiple performance measures are reported to systematically compare and assess the efficacy of eight ML models on previously unseen data (testing dataset). The study also undertakes data imputation and parameter optimization, ensuring robustness and reliability of the findings. Thirdly, this research enhances the existing logistic regression model by incorporating two pivotal variables, namely anxiety level and self-efficacy. Moreover, improvements in data preprocessing steps contribute to a more nuanced understanding of the intricate relationship between these variables and parental depression in the NICU context.

In tackling the complexities inherent in forecasting the risk of parental depression upon NICU discharge, our study takes a comprehensive approach, aiming to identify and prioritize factors associated with this crucial outcome. We sought to establish the most effective predictive model by systematically comparing results obtained from various machine learning (ML) techniques and logistic regression (LR). While previous research in the domain of predicting parental mental health outcomes has delved into the application of ML models (8–19), the specific context of parental depression in the NICU remains underexplored, with only logistic regression studies to date (20). In the absence of conclusive evidence supporting ML’s superiority in predicting parental depression within the NICU, our study fills a critical gap by offering a rigorous comparison between ML techniques and logistic regression. This investigation emerges from a motivation to challenge the assumption that ML universally outperforms traditional methods, especially in the nuanced domain of parental mental health within the NICU. By providing empirical evidence and insights into the predictive efficacy of different methodologies, our research contributes to advancing the understanding of optimal prediction strategies in this unique healthcare context.

Building upon this motivation, our study endeavors to elevate the field by advancing beyond the limitations of the existing logistic regression study on parental depression in the NICU. We explore this uncharted territory by employing eight distinct machine learning (ML) models, each meticulously assessed and compared through comprehensive performance evaluations on previously unseen test data. This departure from conventional methodologies is facilitated by the implementation of a cross-validation technique, dividing the data into two subsets for model evaluation, ensuring robustness and applicability to real-world scenarios. Figure 3 presents a visual representation of our findings, encapsulating crucial performance measures such as accuracy, AUC, precision, sensitivity, and F-score. This not only facilitates an in-depth comparison of the models but also ensures the reproducibility of our results across different frameworks. Furthermore, our study enhances the existing paradigm by fine-tuning a previous model (20), incorporating additional independent variables such as state anxiety scale (STAI-Y1), trait anxiety scale (STAI-Y2), and perceived maternal parenting self-efficacy. This augmentation, coupled with refined preprocessing procedures, contributes to the evolution of predictive models in the NICU setting.

Our findings reveal that a higher level of perceived stress (PSS-10), lower perceived social support (MSPSS), and older gestational age (GA) significantly contribute to depression symptoms among parents of NICU infants. Importantly, our results align statistically with those reported in Soghier et al. (20). Additionally, we observed that parents with a female infant in the NICU face a higher risk of depression symptoms compared to parents of male infants. While the precise reasons for this gender difference remain elusive, analogous results have been noted in other studies, where the reported odds of depression are higher among mothers of female infants (26–28). These studies attribute this outcome to a potential preference for a male infant, suggesting societal influences. Limited evidence also suggests biological differences; mothers carrying a female fetus exhibit elevated levels of β-human chorionic gonadotropin. This indicates that hormonal changes, along with similar alterations, may provide a biological explanation for the impact of the child’s gender on postnatal depression (29, 30).

Building on these significant findings, our study introduces an additional layer of insight by delving into feature importance through SHAP analysis. By not exclusively relying on black box ML-based models, we were able to extract nuanced information about the contributors to parental depression symptoms in the NICU. Particularly noteworthy are the state anxiety scale (STAI-Y1) and trait anxiety scale (STAI-Y2), identified as crucial predictors. This finding emphasizes the importance of not only screening for depression but also for anxiety and social support as both naturally predict the onset of depression. While the connection between anxiety, social support, and depression is well-established, it’s crucial to highlight that many NICUs primarily screen for postpartum depression (PPD), often assuming that certain questions indirectly address anxiety. Our study challenges this assumption, emphasizing the distinct and significant impact of both anxiety and depression on parental mental health.

Having uncovered nuanced insights into the contributors of parental depression symptoms through SHAP analysis, we turn our attention to the performance aspect. Remarkably, the logistic regression model, a key focus of our study, exhibits comparable effectiveness when benchmarked against commonly used ML models. This finding aligns with broader research on depression, where logistic regression has consistently demonstrated either superior or comparable performance compared to alternative ML models (31–33). Our observation prompts consideration of two pivotal factors that contribute to this alignment. First, the richness of our dataset, encompassing a broad spectrum of variables and free from biases, ensures that optimized models consistently exhibit stable performance across diverse algorithms. All eight models achieved an area under the curve (AUC) above 0.8, suggesting that the logistic regression-based model’s performance is comparable to other common ML models. Second, the common ML models typically outperform logistic models in larger datasets. However, the comparable performance observed in our study, possibly attributed to the dataset’s size (three hundred observations), underscores the value of an easily interpretable logit model for predicting postpartum depression among NICU parents, boasting a high accuracy of 0.77.

Based on our findings, while Logistic Regression offers its own advantages and remains one of the top-performing models, it is essential to consider the broader performance metrics displayed in Figure 3. Notably, algorithms such as Naïve Bayes, XGBoost, and Artificial Neural Network demonstrate a remarkable balance between precision and sensitivity as evidenced by their notably high F-score values. This underscores their ability in effectively identifying positive cases while simultaneously minimizing both false positives and false negatives. Naïve Bayes stands out as a rapid algorithm with minimal training time, making it ideal for clinical decision support systems where speed is a crucial constraint (12). XGBoost exhibits the robust ability to mitigate overfitting issues commonly encountered in datasets (34). On the other hand, Neural Network emerges as an excellent choice when dealing with substantial amounts of data sourced from diverse healthcare organizations (35). By highlighting the strengths and distinct qualities of each ML technique in relation to predicting PPD in the NICU, our study expands the potential for accurate predictions, enhances the understanding of PPD risk factors, and provides valuable insights for developing targeted interventions in NICU settings.

It is essential to acknowledge the limitations inherent in this study. Notably, the dataset under consideration exhibited a relatively small number of patients. While the size of our sample is limited, it is imperative to underscore the high quality of the data therein. This dataset originates from a meticulously conducted clinical trial, ensuring a high standard of data integrity. It is crucial to emphasize that the sample is devoid of biases, and further enhances its reliability by maintaining a balanced representation across various racial groups which instill confidence in the validity of our study outcomes. To address potential challenges associated with small sample sizes, rigorous monitoring of learning curves for all prediction models was undertaken throughout the training process. Employing a strategic combination of techniques, including cross-validation, regularization, and hyperparameter tuning, we actively mitigated the risk of overfitting, thereby reinforcing the integrity of our study’s analytical approach. Another limitation of this study is that the dataset utilized was exclusively sourced from the Children’s National Hospital in Washington, DC. Therefore, the generalizability of our study’s results to other healthcare systems monitoring parental depression symptoms in the NICU may be limited. Future studies should aim to include larger and more diverse datasets from multiple institutions to enhance the external validity and generalizability of the predictive models developed in this research.

In conclusion, the findings of this study contribute to the ongoing efforts of improving parental depression screening in the NICU context. The implementation of more accurate and targeted screening systems can ease the burden on both patients and healthcare systems by reducing unnecessary interventions and optimizing resource allocation. Our findings emphasize the importance of evaluating perceived stress, perceived social support, and state anxiety scale as essential factors to be screened in NICU parents. Moreover, our results show that the performance of the logistic regression as an interpretable and easy to use model is comparable with other commonly used ML-based models. This finding facilitates informed decision-making for healthcare providers, empowering them to select the most appropriate model for their specific contexts. These advancements aim to enhance the overall well-being of parents and their infants in the NICU by effectively identifying and addressing parental depression.

The data analyzed in this study is subject to the following licenses/restrictions: this is the analysis of an existing dataset. We obtained the de-identified data from authors of this paper. Requests to access these datasets should be directed to LS, LSoghier@childrensnational.org.

This study was approved by the Children’s National Institutional Review Board and it was registered with https://clinicaltrials.gov (NCT02643472). The studies were conducted in accordance with the local legislation and institutional requirements. Written informed consent for participation was not required from the participants or the participants’ legal guardians/next of kin in accordance with the national legislation and institutional requirements.

FS: Conceptualization, Formal analysis, Methodology, Visualization, Writing – original draft, Writing – review & editing. NH: Conceptualization, Funding acquisition, Supervision, Writing – review & editing. VA: Methodology, Writing – review & editing. LS: Data curation, Funding acquisition, Writing – review & editing.