Using Machine Learning to Predict Fetal Health: A Comprehensive Guide

Introduction

Did you know that timely diagnosis of fetal health conditions can significantly reduce complications during pregnancy? With technological advancements, machine learning (ML) has emerged as a powerful tool to assist healthcare professionals in making accurate predictions and decisions.

In this blog, I’ll walk you through my project where I built an ML pipeline to classify fetal health based on Cardiotocogram (CTG) data. The goal was to predict the health status of a fetus as:

1. Normal

2. Suspect

3. Pathological

By leveraging a dataset of fetal health readings, I developed and evaluated models to improve prediction accuracy, gaining insights into the most critical factors influencing fetal health.

Dataset Overview

The dataset for this project was sourced from the Fetal Health Classification Dataset on Kaggle. It contains:

• 2,126 records with 22 features extracted from CTG readings.

• A target variable fetal_health with three classes:

  • 1: Normal

  • 2: Suspect

  • 3: Pathological

Key Features

Some important features include:

• Baseline value: Fetal heart rate (FHR) baseline (beats per minute).

• Accelerations: Number of FHR accelerations per second.

• Fetal movement: Number of fetal movements per second.

• Uterine contractions: Number of uterine contractions per second.

• Histogram properties: Statistical metrics from FHR histograms.

I identified key patterns and trends that informed my machine-learning approach by exploring and visualising this dataset.

Project Workflow

This project followed a structured pipeline:

1. Data Loading and Exploration

   • Load the dataset and explore its structure.

   • Visualize class distributions and feature relationships.

2. Data Preprocessing

   • Handle missing or infinite values.

   • Scale numerical features using StandardScaler.

3. Model Training and Evaluation

   • Train a baseline Random Forest model.

   • Evaluate its performance using accuracy, confusion matrix, and ROC-AUC metrics.

4. Hyperparameter Tuning

   • Optimize the Random Forest model using GridSearchCV to improve performance.

5. Optimized Model Evaluation

   • Train the model with the best hyperparameters and evaluate its performance.

6. Feature Importance Analysis

   • Analyze the importance of features contributing to the predictions.

Exploratory Data Analysis (EDA)

Class Distribution

To start, I visualized the distribution of the target variable (fetal_health) and found that the dataset is slightly imbalanced, with most records belonging to the Normal class.

Feature Correlation

I generated a correlation heatmap to understand feature relationships. It revealed:

• Strong correlations between histogram-related features.

• Features like baseline value and accelerations showed significant potential as predictors.

Model Training and Results

Baseline Model

I began with a baseline Random Forest Classifier, which achieved:

• Accuracy: 92.4%

• ROC-AUC Score: 0.95

However, there was room for improvement in classifying the minority classes (Suspect and Pathological).

Hyperparameter Tuning

Using GridSearchCV, I optimized the following parameters:

• n_estimators: Number of trees in the forest.

• max_depth: Maximum depth of the trees.

The best parameters were:

{'n_estimators': 200, 'max_depth': None}

Optimized Model

The optimized Random Forest model achieved:

• Accuracy: 95.3%

• ROC-AUC Score: 0.97

This improvement highlighted the importance of hyperparameter tuning for better generalization.

Feature Importance

Analyzing feature importance provided key insights:

• Baseline value, accelerations, and histogram mean were the most significant predictors of fetal health.

• Visualizing these helped interpret the model’s behaviour and ensured that the predictions aligned with domain knowledge.

Challenges and Solutions

1. Class Imbalance

   • Challenge: Fewer records in the Suspect and Pathological classes.

   • Solution: Experimented with class weights in the model to address this imbalance.

2. Outliers

   • Challenge: Outliers in features like severe decelerations.

   • Solution: Applied boxplots and carefully handled outliers to reduce noise.

Future Improvements

While this project achieved strong results, there’s always room for improvement. Future steps could include:

• Exploring advanced models like XGBoost or LightGBM.

• Implementing cross-validation for more robust performance metrics.

• Deploying the model using Flask, FastAPI, or Streamlit for real-world applications.

Conclusion

This project demonstrates the potential of machine learning in healthcare, particularly for assisting in early diagnosis and decision-making. By leveraging CTG data, we can build accurate, interpretable models to predict fetal health and support healthcare professionals.

Interested in exploring the project? Check out:

• Kaggle Notebook

• GitHub Repository

Feel free to share your thoughts or suggestions in the comments!