Machine Learning-Based Non-Invasive Diagnosis of Anemia in Children Using Palm Image Analysis

This study presents a machine learning framework that utilizes palmar image analysis to achieve highly accurate, non-invasive anemia classification and hemoglobin level prediction in children, demonstrating robust performance across diverse skin tones and strong potential for point-of-care screening.

The Gist

Imagine your body is a bustling city, and hemoglobin is the fleet of delivery trucks responsible for carrying oxygen to every neighborhood. When there aren't enough trucks (anemia), the city starts to slow down, especially in children whose growth and learning depend on a steady supply of energy.

Usually, to count these trucks, doctors have to stop the city traffic, draw blood, and run lab tests. But this new study suggests a simpler way: just take a picture of a child's palm.

Here is how the researchers turned a simple photo into a medical detective tool, explained in everyday terms.

The Problem: Finding the "Ghost" in the Machine

Anemia is like a silent thief; it steals oxygen without making a loud noise. In many parts of the world, checking for it requires needles and expensive machines. The researchers wanted to know: Can we spot the "thief" just by looking at how pale a child's hand looks?

The Solution: The Digital Detective

The team built a computer system that acts like a super-observant detective. Here is their step-by-step process:

1. The "Magic Crop" (Segmentation)
First, the computer had to find the hand in the photo. It used a smart tool called U-Net (think of it as a digital pair of scissors) to cut out just the palm, ignoring the background, the wrist, or the fingers. It did this with incredible precision, getting it right 96% of the time.

2. Translating Colors (The Color Spaces)
Human eyes see colors, but computers see numbers. The researchers didn't just look at the photo in its normal "RGB" (Red, Green, Blue) mode. They translated the image into two other "languages" that are better at describing human skin:

• CIELab: A language where one axis measures "Redness" vs. "Greenness." Since anemia makes skin look less red (paler), this was a crucial clue.
• HSV: A language that separates color from brightness, helping the computer ignore shadows or bright lights.

3. The "Redness" Clues
The computer didn't just guess; it measured specific things, like:

• The Red Fraction: How much of the pixel is actually red?
• The Erythema Index: A fancy math way of saying "how red is this?"
• The 'a' value: In the CIELab language, a high 'a' value means "very red," and a low 'a' value means "pale."

The Training: Teaching the Computer

The team fed the computer photos of 174 children from a hospital in Tehran. They knew the exact hemoglobin levels of these kids from blood tests (the "ground truth").

• The Class: 150 kids were healthy; 24 were anemic.
• The Lesson: The computer learned that when the "Redness" numbers dropped below a certain line, the child was likely anemic.

To make sure the computer didn't just memorize the answers, they used a trick called Data Augmentation. They took the existing photos, rotated them, flipped them, and zoomed in/out (like playing with a photo album) to create more practice examples. This helped the computer learn the concept of anemia, not just the specific faces of the kids in the study.

The Results: How Good Was the Detective?

The computer tried to solve two puzzles:

1. "Is this child anemic?" (Classification)

   • The best models (Logistic Regression and a custom CNN) got it right about 95% of the time.
   • It was like a test where the computer only made a few mistakes out of 100 tries.

2. "What is the exact hemoglobin number?" (Regression)

   • The computer tried to guess the exact number (e.g., 10.5 g/dL) based on the palm color.
   • Using a Random Forest model (which acts like a committee of decision-makers), it predicted the numbers with amazing accuracy, especially for children with darker skin tones.
   • In fact, for darker-skinned children (who made up most of the group), the prediction was almost perfect (99% accuracy).

The "Why" Behind the Magic

The researchers used a tool called SHAP to ask the computer, "Why did you make that guess?"
The computer answered clearly: "I looked at the redness."
It confirmed that the most important clues were the red-related features. If the palm looked less red, the computer correctly flagged it as a low-hemoglobin case.

The Bottom Line

This study shows that you don't always need a needle to check for anemia. By taking a photo of a child's palm and letting a smart computer analyze the shades of red, you can get a very reliable estimate of their health.

• The Tool: A smartphone camera + AI software.
• The Target: Children under six.
• The Promise: A fast, painless, and non-invasive way to screen for anemia, particularly effective for the population studied (Iranian children with diverse skin tones).

The paper concludes that this "palm-reading" method, powered by artificial intelligence, is a strong, reliable tool that could help doctors and parents catch anemia early, without the fear of a blood draw.

Technical Summary

Technical Summary: Machine Learning-Based Non-Invasive Diagnosis of Anemia in Children Using Palm Image Analysis

Problem Statement
Anemia, particularly iron-deficiency anemia, remains a critical global health issue affecting 42% of children under six years of age. Early detection is vital to prevent long-term developmental and cognitive impairments. However, standard diagnosis relies on Complete Blood Count (CBC) tests, which are invasive, require blood draws, and necessitate repeated monitoring. While non-invasive alternatives exist, they often rely on manual visual inspection of pallor in regions like the conjunctiva or nail beds, which can be subjective and inconsistent. There is a need for an objective, automated, and non-invasive screening tool capable of estimating hemoglobin (Hb) levels and classifying anemia status, specifically tailored for pediatric populations where frequent blood testing is burdensome.

Methodology
The study proposes a computer vision and machine learning framework utilizing palmar images of children to screen for anemia and predict hemoglobin levels.

• Data Collection: The dataset was sourced from the Children's Medical Center Hospital in Tehran, Iran, involving 174 pediatric subjects (under 7 years old). Images were captured under standardized lighting conditions using a Nikon D7500 camera. Following imaging, standard CBC tests were performed to establish ground truth. The final dataset comprised 150 non-anemic samples (Hb ≥ 11 g/dL) and 24 anemic samples (Hb < 11 g/dL due to iron deficiency).
• Preprocessing and Segmentation: A U-Net deep learning model was employed to automatically segment the palm region of interest (ROI), achieving a Dice coefficient of 0.96. Images were resized to 224 × 224 pixels, and white balance was corrected using calibration cards. To address the small dataset size and class imbalance, data augmentation techniques (rotation, flipping, zooming, and translation) were applied, expanding the training set.
• Feature Extraction: Images were processed across three color spaces: RGB, CIELab, and HSV.
  • RGB: Features included mean red intensity, red fraction, erythema index, and percentile-based distributions (5th, 25th, 50th, 75th, 95th) of pixel intensities.
  • CIELab: The 'a' channel (red-green axis) was prioritized. Features included the mean 'a' value, normalized 'a' value, and percentile distributions.
  • HSV: Mean and minimum values for Hue (H) and Saturation (S) channels were extracted.
• Modeling:
  • Classification: Multiple machine learning algorithms (Logistic Regression, Gradient Boosting, AdaBoost, Random Forest, k-NN, Naive Bayes, Decision Tree) and a custom Convolutional Neural Network (CNN) were evaluated for binary classification (anemic vs. non-anemic). The CNN architecture featured four convolutional blocks with batch normalization, global average pooling, and dropout regularization.
  • Regression: Linear Regression and Random Forest models were used to predict continuous hemoglobin values.
  • Interpretability: SHAP (SHapley Additive exPlanations) values were calculated to determine feature importance and the direction of influence on predictions.

Key Results

• Segmentation: The U-Net model demonstrated robust segmentation performance with an Intersection over Union (IoU) of ~0.935 and a Dice coefficient of ~0.96.
• Classification Performance:
  • In the RGB color space, Logistic Regression, Gradient Boosting, and AdaBoost achieved the highest test accuracy of 94.5% (AUC 0.96).
  • The custom CNN model achieved a test accuracy of 95.53% with a loss of 0.175, demonstrating strong generalization with minimal overfitting (training and validation accuracy remained within ±0.02).
  • In the CIELab color space, the mean 'a'-axis value was identified as the most predictive feature, with a threshold of 115 distinguishing anemic samples.
• Hemoglobin Prediction:
  • The Random Forest model in the CIELab color space achieved a coefficient of determination ($R^2$) of 0.95 and a Pearson correlation coefficient of 0.98 between actual and predicted Hb values.
  • Performance varied by skin tone. The model achieved an $R^2$ of 0.9926 for darker-skinned individuals (the majority of the Iranian cohort) compared to 0.8501 for lighter-skinned individuals.
  • SHAP analysis confirmed that red-related features (specifically average red intensity and erythema index) were the dominant predictors, with higher values correlating to higher hemoglobin estimates.

Significance and Claims
The paper claims that the proposed framework offers a reliable, non-invasive, and effective tool for pediatric anemia screening. By leveraging the CIELab color space and advanced machine learning, the study demonstrates that palmar pallor analysis can accurately classify anemia and estimate hemoglobin levels without blood extraction. The authors highlight that the method is particularly robust for darker skin tones, addressing a gap in many existing diagnostic tools. The integration of automated segmentation (U-Net) and explainable AI (SHAP) supports the potential for this technology to be deployed as a point-of-care application or a home-based monitoring tool, facilitating early detection and improving health outcomes for children in resource-constrained settings. The study emphasizes that while the results are promising, the approach is intended to serve as a complementary diagnostic tool for physicians rather than a standalone replacement for clinical blood tests.