Development and validation of an XGBoost model with SHAP-based interpretability and a web-based calculator for predicting extrauterine growth restriction in preterm infants

This study developed and validated an interpretable XGBoost-based machine learning model with a web-based calculator that accurately predicts extrauterine growth restriction in preterm infants using early clinical variables, leveraging SHAP analysis to identify key risk factors and support individualized clinical management.

The Gist

Imagine a premature baby is like a tiny, fragile seedling that has been planted in a garden (the mother's womb) but had to be moved to a greenhouse (the hospital) before it was fully ready. The goal is to help this seedling grow strong enough to survive and thrive outside.

However, sometimes these seedlings struggle to catch up in size and strength after leaving the womb. Doctors call this Extrauterine Growth Restriction (EUGR). It's a big problem because if a baby doesn't grow well early on, it can lead to health issues later in life, like diabetes or learning difficulties.

For a long time, doctors tried to predict which babies would struggle using simple math formulas (like a basic recipe). But human bodies are messy and complex; the ingredients don't always mix in a straight line. Sometimes, two factors together create a bigger problem than just adding them up.

Here is what this new study did, explained simply:

1. The "Super-Brain" Detective (XGBoost)

Instead of using a simple recipe, the researchers built a Super-Brain Detective called XGBoost. Think of this as a highly advanced AI that can look at a baby's early history and spot hidden patterns that a human or a simple calculator would miss.

They fed this AI data from 1,431 premature babies (about 860 for "training" and 570 for "testing"). The AI looked at things like:

• How small the baby was at birth.
• If the baby was born too early.
• If the mother had high blood pressure or an infection.
• If the baby had trouble breathing or a heart issue.
• Whether the baby was breastfed.

2. The "Why" Behind the "What" (SHAP)

Usually, AI is a "black box." You put data in, and it gives an answer, but it doesn't tell you why. That's scary for doctors.

This study added a special feature called SHAP. Imagine SHAP is a transparent flashlight that shines on the AI's decision. It shows exactly which "clues" made the AI decide a baby was at high risk.

• The Flashlight revealed: Low birth weight and infections were like heavy anchors dragging the baby down.
• The Good News: Breastfeeding was like a life raft, helping to keep the baby safe.
• The Twist: The AI found that some problems work together. For example, a baby with a heart issue (PDA) who wasn't breastfed was in much more danger than if they had been breastfed. It's like a storm being much worse if you don't have an umbrella.

3. The "Magic Calculator" (Web Tool)

The researchers didn't just keep this Super-Brain in a lab. They built a free, easy-to-use website (a web-based calculator).

• How it works: A doctor or nurse goes to the website.
• The Input: They slide a few bars or click a few boxes with the baby's early info (like "Birth Weight: 1.5kg" or "Breastfed: Yes").
• The Output: In seconds, the site gives a risk percentage (e.g., "This baby has a 75% chance of growth restriction").
• The Bonus: It also draws a little chart showing why that baby is at risk, so the doctor knows exactly what to watch out for or how to help (like prioritizing breastfeeding).

Why This Matters

Think of the old way of predicting risk as using a paper map in a foggy forest. You might get lost because the map is too simple.

This new tool is like a GPS with a live camera. It sees the terrain in real-time, knows exactly where the potholes are (the risk factors), and tells the driver (the doctor) the best route to take.

The Bottom Line:
This study created a smart, easy-to-use tool that helps doctors spot premature babies who are likely to struggle with growth very early on. By knowing this sooner, doctors can give these babies the special care they need immediately, helping them grow stronger and healthier for the rest of their lives.

Technical Summary

1. Problem Statement

Extrauterine Growth Restriction (EUGR) is a critical complication in preterm infants associated with adverse short-term outcomes (e.g., neurodevelopmental impairment) and long-term metabolic risks (e.g., obesity, diabetes).

• Current Limitations: Existing prediction models often rely on traditional logistic regression, which assumes linear relationships between variables and the log-odds of the outcome. This approach fails to capture complex nonlinear associations and interaction effects inherent in clinical data. Furthermore, traditional models (like nomograms) often lack individual-level interpretability, making it difficult for clinicians to understand why a specific patient is flagged as high-risk.
• Objective: To develop a high-performance, interpretable machine learning model for early EUGR prediction (within 48–72 hours of birth) and translate it into a user-friendly clinical tool.

2. Methodology

Study Design and Data

• Population: Retrospective analysis of 1,431 preterm infants (gestational age 28–36+6 weeks) admitted within 24 hours of birth.
• Dataset Split:
  • Training Set: 863 infants from the Yangpu campus (May 2020 – March 2025).
  • Validation Set: 568 infants from the Huangpu campus (May 2022 – March 2025).
• Outcome: Incidence of EUGR was 34.1% (488 cases).
• Data Preprocessing: Missing data were handled via multiple imputation using the mice package (random forest-based).

Feature Selection

• Algorithm: The Boruta algorithm (a wrapper method based on Random Forests) was used to screen 19 potential clinical variables.
• Selected Features: Nine variables were confirmed as important:
  1. Birth Weight
  2. Small for Gestational Age (SGA) status
  3. Gestational Age
  4. Breastfeeding status
  5. Multiple gestation
  6. Neonatal Respiratory Distress Syndrome (NRDS)
  7. Patent Ductus Arteriosus (PDA)
  8. Maternal Hypertension
  9. Maternal Group B Streptococcus (GBS) infection

Model Development and Comparison

Five models were trained and compared on the validation set:

1. Logistic Regression (Baseline)
2. XGBoost (Extreme Gradient Boosting)
3. Random Forest
4. Decision Tree
5. Support Vector Machine (SVM)

Interpretability and Analysis

• SHAP (SHapley Additive exPlanations): Used to interpret the optimal XGBoost model.
  • Global Interpretability: Feature importance ranking and beeswarm plots to visualize the direction (risk vs. protection) and magnitude of each feature.
  • Local Interpretability: Waterfall plots to explain individual predictions.
  • Nonlinearity & Interactions: SHAP dependence plots and interaction surface plots were generated to uncover complex relationships (e.g., between birth weight and gestational age).

Clinical Tool Implementation

• A web-based calculator was built using Python 3.11 and the Streamlit framework.
• It accepts the 9 selected inputs, computes real-time risk probability, and visualizes the specific contribution of each factor for the individual patient.

3. Key Results

Predictive Performance

The XGBoost model outperformed all other models on the validation set:

• AUC: 0.922 (highest among all models).
• Accuracy: 0.849.
• Sensitivity: 0.757.
• Specificity: 0.895.
• Brier Score: 0.108 (indicating high calibration).
• Comparison: XGBoost significantly outperformed Logistic Regression (AUC 0.899) and other tree-based models.

Feature Importance and Interpretability

• Top Risk Factors: Low birth weight, SGA status, Maternal GBS infection, and PDA.
• Protective Factors: Higher birth weight, higher gestational age, and breastfeeding.
• Nonlinear Effects:
  • Birth Weight: A sharp decrease in risk was observed as weight increased from ~800g to 2000g. Above 2000g, the risk reduction plateaued.
  • Gestational Age: Showed a weaker nonlinear effect compared to birth weight.
• Interaction Effects:
  • Birth Weight × Gestational Age: Low gestational age significantly amplified the risk associated with low birth weight.
  • Breastfeeding × PDA: The protective effect of breastfeeding was significantly stronger in infants with PDA compared to those without. Conversely, the absence of breastfeeding in infants with PDA drastically increased EUGR risk.

Clinical Utility

• Calibration: The model showed excellent agreement between predicted probabilities and observed frequencies.
• Decision Curve Analysis (DCA): The XGBoost model demonstrated a higher net benefit across a wide range of threshold probabilities (0.1 to 0.9) compared to "treat all" or "treat none" strategies.

4. Key Contributions

1. Superior Predictive Model: Demonstrated that XGBoost captures complex, nonlinear clinical patterns better than traditional logistic regression for EUGR prediction.
2. Deep Interpretability: Moved beyond "black box" AI by using SHAP to quantify individual feature contributions and visualize specific interaction effects (e.g., the synergistic risk of PDA + no breastfeeding).
3. Clinical Translation: Successfully deployed the model as a real-time, web-based calculator, bridging the gap between algorithmic research and bedside clinical decision-making.
4. Feature Discovery: Identified and validated the critical interaction between breastfeeding and PDA, suggesting targeted nutritional interventions for this specific high-risk subgroup.

5. Significance and Limitations

Significance:
This study provides a scientifically robust, interpretable, and accessible tool for neonatologists. By identifying high-risk infants early (within 48–72 hours) and explaining why they are at risk, clinicians can implement personalized nutritional and clinical management strategies to mitigate EUGR and its long-term sequelae.

Limitations:

• Retrospective Nature: Reliance on historical medical records introduces potential information bias.
• Single-Center Data: The model was developed and validated within one hospital system (Fudan University), limiting generalizability to different populations or healthcare settings without further multi-center validation.
• Missing Variables: Factors such as socioeconomic status, genetic markers, and detailed postnatal care variables were not included.
• Prospective Validation: The tool has not yet undergone prospective clinical trials to verify its impact on actual patient outcomes.

Conclusion:
The authors conclude that the XGBoost-based model, enhanced by SHAP interpretability and deployed via a web calculator, is a promising tool for early EUGR risk stratification, offering a pathway toward more precise, data-driven neonatal care.