Random Forest Model for Predicting Post-Lockdown Antenatal Depression Risk: A Cross-Sectional Study of Pregnant Women in China

This cross-sectional study in China found that post-lockdown antenatal depression is prevalent and strongly associated with sleep disorders, family support, and COVID-19 symptom severity, with a Random Forest machine learning model demonstrating superior predictive performance compared to other algorithms.

The Gist

The Big Picture: A "Mental Health Weather Report" for Pregnant Moms

Imagine pregnancy as a long, important journey. Usually, this journey has its own bumps and turns, but the recent pandemic added a massive storm cloud that followed everyone.

This study is like a post-storm damage report. The researchers wanted to know: Now that the strict "lockdown" rules in China have been lifted and people can move around again, how are pregnant women feeling? Are they still carrying the heavy emotional baggage of the pandemic?

They didn't just ask "Are you sad?"; they built a high-tech prediction machine (using Artificial Intelligence) to figure out exactly why some moms are struggling more than others, so doctors could spot the trouble early.

The Cast of Characters

• The Subjects: 459 pregnant women in China (from the busy city of Beijing and the diverse, mountainous region of Guizhou).
• The Timeframe: The first half of 2023, right after the strict lockdowns ended.
• The Tool: They used a standard "depression thermometer" called the EPDS. If the score was high enough, it meant the woman was likely experiencing Antenatal Depression (AD).

The Detective Work: What Did They Find?

The researchers found that about 1 in 4 (25.7%) of the pregnant women tested positive for depression. This is higher than usual, suggesting the pandemic's shadow is still long.

To understand why, they acted like detectives looking for clues. They compared the women who were struggling with those who were doing okay. They found three main "villains" that were making the depression risk much higher:

1. The Sleep Thief (Sleep Disorders): This was the biggest culprit. If a mom couldn't sleep well, her risk of depression skyrocketed. Think of sleep as the "recharge battery" for the brain; without it, the system crashes.
2. The Support Net (Family Support): It wasn't just about having a family; it was about the quality of that support. Women who felt their family was only "okay" at helping them (a middle level) were at much higher risk than those who felt fully supported. It's like having a safety net that has holes in it—it's better than nothing, but not enough to catch you if you fall.
3. The Virus Shadow (Symptom Severity): Women who got sick with COVID-19 and had moderate to severe symptoms were at higher risk. It wasn't just the worry about the virus; the actual physical sickness seemed to take a toll on their mental state.

The "Magic Crystal Ball": The Random Forest Model

The researchers didn't stop at just listing the problems. They wanted to build a crystal ball that could predict who was at risk before the depression got too bad.

They tried out six different types of "prediction engines" (mathematical models), including:

• Logistic Regression: The old-school, reliable calculator.
• SVM & KNN: The sharp-eyed detectives.
• XGBoost & GBDT: The heavy-duty power tools.
• Random Forest (RF): The wise, experienced forest ranger.

The Winner: The Random Forest model won the race.

• Why "Forest"? Imagine a single tree trying to guess the weather; it might get it wrong. But if you have a whole forest of trees, each looking at the data from a slightly different angle, and they all vote on the answer, the result is incredibly accurate.
• The Score: This "forest" was very good at spotting the difference between women who were depressed and those who weren't, beating all the other models in the test.

Making the "Black Box" Transparent

One problem with AI is that it's often a "black box"—it gives an answer, but you don't know how it got there. To fix this, the researchers used a tool called SHAP.

Think of SHAP as a spotlight. When the model makes a prediction, the spotlight shines on the specific clues it used.

• The spotlight showed that Sleep Disorders were the brightest, loudest clue.
• Family Support and COVID Symptoms were the next brightest clues.

This confirmed that the AI wasn't just guessing; it was correctly identifying the real-world factors that matter most.

The Bottom Line

This study tells us that even after the lockdowns ended, the pandemic left a mark on pregnant women's mental health.

• The Main Takeaway: If a pregnant woman is having trouble sleeping, feels her family isn't fully supporting her, or recently suffered from a bad case of COVID, she is at a much higher risk of depression.
• The Solution: The "Random Forest" model acts like a smart screening tool. By checking these three specific factors, doctors could potentially identify women who need help before their depression becomes severe.

The researchers are essentially saying: "We found the three biggest warning lights on the dashboard. If we watch those lights, we can help keep the journey safe."

Technical Summary

Technical Summary: Random Forest Model for Predicting Post-Lockdown Antenatal Depression

Problem Statement
Following the easing of COVID-19 lockdown measures in China, pregnant women faced an elevated risk of infection and associated psychological stress, potentially exacerbating antenatal depression (AD). While previous research focused heavily on the early stages of the pandemic or utilized singular statistical methods (e.g., logistic regression) primarily in Western populations, there was a lack of comparative studies employing multiple machine learning (ML) algorithms to predict AD risk among Chinese pregnant women in the post-lockdown period. Furthermore, existing models often lacked interpretability, limiting their clinical utility. This study aimed to investigate the prevalence of AD in this specific post-lockdown context, identify independent risk factors, and develop and compare predictive models using six different ML algorithms to determine the most effective approach for risk stratification.

Methodology

• Study Design and Population: A cross-sectional study was conducted between January and August 2023 in Chaoyang District, Beijing, and Qiandongnan Prefecture, Guizhou Province. The study included 459 pregnant women (effective response rate of 91.07%) who provided informed consent.
• Data Collection: Participants completed a self-designed questionnaire covering sociodemographic characteristics, pregnancy-related factors (including sleep quality assessed via the Pittsburgh Sleep Quality Index [PSQI] and family support levels), and COVID-19-related factors (infection severity, concerns, and economic impact).
• Outcome Measure: Antenatal depression was screened using the Edinburgh Postnatal Depression Scale (EPDS). A score of ≥10 was defined as a positive marker for AD.
• Statistical Analysis:
  • Variable Selection: Univariate and multivariate logistic regression analyses were performed to identify independent predictors of AD.
  • Model Development: The dataset was split into training (60%) and testing (40%) sets. Six ML algorithms were trained and compared: Logistic Regression (LR), Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Random Forest (RF), eXtreme Gradient Boosting (XGBoost), and Gradient Boosting Decision Tree (GBDT).
  • Evaluation Metrics: Models were evaluated using Area Under the Curve (AUC), precision, recall, accuracy, F1 score, sensitivity, and specificity.
  • Interpretability and Validation: The best-performing model underwent further examination using SHapley Additive exPlanations (SHAP) for feature importance, Calibration Curves (CC) for discrimination, and Decision Curve Analysis (DCA) for clinical utility.

Key Results

• Prevalence: The prevalence of AD in the study population was 25.7% (118/459).
• Independent Predictors: Multivariate analysis identified three independent risk factors for AD:
  1. Sleep disorders (OR = 4.19).
  2. Middle-level family support (compared to high-level support; OR = 2.84).
  3. Moderate to severe COVID-19 symptom severity (compared to non-infected/asymptomatic; OR = 3.63).
• Model Performance: The Random Forest (RF) model demonstrated superior predictive performance compared to the other five algorithms.
  • Training Set AUC: RF (0.842) > GBDT (0.834) > XGBoost (0.801) > LR (0.790) > SVM (0.776) > KNN (0.749).
  • Testing Set AUC: RF (0.724) > XGBoost (0.713) > GBDT (0.694) > SVM (0.680) > LR (0.679) > KNN (0.599).
• Model Interpretation: SHAP analysis confirmed that "sleep disorders" had the greatest impact on AD risk, followed by "family support level" and "COVID-19 symptom severity."
• Validation: The RF model showed good calibration (Hosmer-Lemeshow P > 0.05) and demonstrated significant clinical utility across threshold probabilities of 8%-95% (training) and 10%-58% (testing) in the Decision Curve Analysis.

Key Contributions

1. Post-Lockdown Focus: The study addresses a gap in literature by focusing on the mental health of pregnant women specifically after the easing of pandemic lockdowns, a period where infection risks increased.
2. Algorithmic Comparison: Unlike previous studies that often relied on a single statistical method, this research systematically compared six distinct ML algorithms, identifying RF as the most robust for this specific demographic and context.
3. Interpretability: By integrating SHAP values, the study overcomes the "black box" limitation of the RF model, providing clinicians with clear insights into the relative importance of specific risk factors (sleep, family support, infection severity).
4. Diverse Demographics: The study included participants from both a developed urban area (Beijing) and a region with a high concentration of ethnic minorities (Guizhou), offering broader applicability for policy formulation in diverse Chinese settings.

Significance and Claims
The authors claim that the EPDS-based Random Forest model effectively predicts AD risk among pregnant women in the post-lockdown period. They posit that the integration of ML into AD risk assessment offers a targeted tool for early screening, which could aid in policy adjustments during public health emergencies and help alleviate the medical burden associated with untreated depression. The study emphasizes that addressing sleep disorders, strengthening family support, and managing the psychological impact of infection severity are critical for intervention. The authors conclude that while the RF model shows promise, its application in clinical practice requires caution due to the cross-sectional nature of the data and the inherent complexity of ensemble learning models.