Postpartum Depression
Risk Assessment

Project Information

This project proposes a trajectory-based framework for postpartum depression risk assessment using longitudinal depressive symptom data collected across pregnancy and postpartum.

Structure of Dataset

• Timeline: Longitudinal data collected across pregnancy and after birth
  • Early pregnancy (Baseline – F)
  • Second trimester (S)
  • Third trimester (T)
  • Postpartum (FP)
• Entities: Pregnant and postpartum individuals receiving mental health screening
• Variables: 61 variables per participant including demographic, health, psychosocial, and mental health assessment data

Analysis Findings Overview

Descriptive Analysis

• High-risk prevalence drops through pregnancy from 4.0% at first trimester to 0.9% at third trimester, but rebounds to 3.2% postpartum
• The distribution of residence type: ~87% urban, ~13% rural (Women in rural areas are underrepresented)
• The household income among participants: ~75% in the middle bracket (50k–200k CNY). Very few extremely low income (<50k)
• Unplanned pregnancy: ~21% unplanned, ~79% planned
• Number of previous children: ~65% first-time mothers, ~34% have one previous child, very few have two
• FEPDS mean 5.72 vs FPEPDS mean 4.38; symptoms generally improve from first trimester to postpartum on average
• FPEPDS max of 29: someone scored nearly the maximum possible, showing severe cases exist despite the low mean

Diagnostic Analysis

• Most participants are first-time mothers, which reflects China's two-child policy era
• Weak correlation between first trimester EPDS and postpartum EPDS (r=0.31)
• Both ACE score (r=0.22) and SLE score (r=0.26) show modest correlations with postpartum EPDS
• A meaningful subgroup of women experience worsening depressive symptoms that the average conceals
• Symptom trajectories are not uniform and that postpartum deterioration is a distinct and clinically important pattern

Predictive Analysis

• We tried different models to predict Postpartum Depression (measured by the FPEPDS score) using data collected at different stages of pregnancy:
  • Model F only: Uses only baseline and First Trimester data (FEPDS, FGAD7) — ~10% explained variance
  • Model F+S: Adds Second Trimester data — more than doubles explained variance
  • Model F+S+T: Incorporates all trimester data — best overall predictive performance ✓ (deployed)
• Handling class imbalance: RandomOverSampler duplicates high-risk cases in training so the model learns minority class patterns better
• Two pipelines: classification (high-risk yes/no using Youden threshold) and regression (predicted EPDS severity score)

Domain Answers

• Early screening alone is inadequate — a longitudinal trajectory-based approach integrating psychosocial risk factors is necessary for early identification
• ACE and SLE scores show only modest correlations, meaning mental health risk is influenced by multiple factors, not just past trauma
• Care shouldn't stop after delivery; new mothers still need mental health support
• Not all mothers follow the same risk pattern — some improve over time, others worsen after giving birth
• The closer to delivery, the better we can identify who is at risk