Newborn Jaundice Screening With Phone Color Calibration

ISEF Category: Systems Software

Subcategory: Mobile Apps  ·  Difficulty: Advanced  ·  Setup: Home Setup  ·  Time: 1 to 2 Months

The Hook

Newborn jaundice is common, but phone cameras do not all see yellow the same way. That means one app can look accurate on one phone and fail on another. Your project asks a real software question, can calibration make skin color measurements reliable across devices?

What Is It?

This project uses a phone camera to estimate how yellow a newborn’s skin looks. Jaundice raises bilirubin in the blood, and high bilirubin can make skin and the whites of the eyes look yellow. A mobile app can read color from a photo, then compare that color to a printed reference card so the result is less affected by the camera, screen, or lighting.

Think of it like a bathroom scale that needs to be zeroed before you step on it. Without calibration, two phones can give different readings for the same image. With calibration, you try to convert raw RGB values, red, green, and blue, into a more consistent score that can be compared across devices and photo conditions. The software challenge is to make those scores stable enough to help flag possible jaundice risk.

Why This Is a Good Topic

This is a strong science fair topic because you can test a clear input and output, then measure how much calibration improves prediction. You are not just making an app. You are testing whether a design choice, the printed reference card and color correction, changes accuracy, error, and device-to-device consistency. That connects to a real health problem in places where access to trained clinicians or bilirubin meters may be limited.

Research Questions

• How does cross-device color calibration affect agreement between phone models when measuring skin yellowing from photos?
• What is the effect of different lighting conditions on jaundice classification accuracy before and after calibration?
• Does adding a printed reference card reduce error in RGB-based jaundice scores across phone cameras?
• To what extent can a simple color score separate jaundice-positive and jaundice-negative images in published datasets?
• Which color channel, or color ratio, best predicts jaundice labels after normalization?
• How does image cropping strategy affect model performance when the face, chest, or eye region is analyzed?
• What is the effect of training on one dataset and testing on another on the app’s classification accuracy?

Basic Materials

• Smartphone with a camera.
• Second smartphone or tablet for cross-device testing.
• Printed color reference card with known color patches.
• Free image analysis app or coding environment.
• Computer with spreadsheet software.
• Published jaundice photo dataset from a journal article or public research repository.
• Notebook for recording lighting, device model, and image conditions.
• Plain background for taking controlled test photos.

Advanced Materials

• Color calibration target or professionally printed reference card.
• Colorimeter or spectrophotometer, if available.
• Mac or PC for model training and evaluation.
• Dataset with neonatal jaundice images and labels from published research.
• Python environment with computer vision libraries.
• Android Studio or Xcode for prototype app development.
• Statistical analysis package.
• Ethics and consent documents if collecting any human images in a school-approved setting.

Software & Tools

• Python: Cleans image data, computes color features, and runs classifier tests.
• ImageJ: Measures color values in selected skin regions and helps compare images.
• Google Colab: Lets you run Python code without installing software on your computer.
• R: Supports statistical tests, ROC curves, and device comparison plots.
• Android Studio: Builds a prototype mobile app if you want to test the workflow on Android.

Experiment Steps

1. Define the exact output your app will predict, such as a jaundice risk score or a binary positive or negative label.
2. Choose one calibration method and one baseline method so you can compare them fairly.
3. Decide which skin region, image crop, or feature set the app will analyze first.
4. Build a test plan that compares multiple phones under the same photo conditions and records calibration error.
5. Select evaluation metrics that match the goal, such as accuracy, sensitivity, specificity, and cross-device agreement.
6. Plan how you will separate training data, validation data, and final test data so the results stay honest.

Common Pitfalls

• Testing only one phone model, which hides how badly the app changes across different cameras.
• Using mixed lighting or auto-flash photos, which makes RGB values shift for reasons unrelated to jaundice.
• Picking a calibration card that is not printed consistently, which breaks the color reference from image to image.
• Training and testing on the same dataset split, which makes accuracy look better than it really is.
• Treating every yellow skin image as jaundice without checking labels, which turns the project into a color test instead of a clinical classifier.

What Makes This Competitive

A stronger project goes past a simple app demo and asks how well the method travels across phones, lighting, and datasets. You can compare raw RGB, calibrated RGB, and a more advanced normalization method, then test which one holds up best. A competitive entry also uses clean metrics, confidence intervals, and a true holdout dataset. That turns your project from a build exercise into a real evaluation of reliability.

Project Variations

• Test whether the same calibration method works on eye-sclera images instead of skin images.
• Compare a rule-based color score with a small machine learning model trained on calibrated photos.
• Check whether the app performs differently on images from low-cost Android phones versus higher-end iPhones.

Learn More

• PubMed: Search for review articles on neonatal jaundice screening, bilirubin measurement, and smartphone-based color analysis.
• NIH: Look for pediatric jaundice background pages and clinical guidance on bilirubin risk in newborns.
• NOAA Color Science resources: Find free background on illumination, color spaces, and why lighting changes camera measurements.
• NASA Earth Observatory color tutorials: Use the image analysis sections to understand RGB, calibration, and reflectance basics.
• MIT OpenCourseWare, computer vision and image processing courses: Search for free lecture notes on image features, color correction, and classification.
• Pediatrics and BMC Pediatrics: Search for peer-reviewed studies on jaundice detection, smartphone screening, and neonatal image analysis.