Predicting anthropometric body composition variables using 3D optical imaging and machine learning.

Area of Science:

• Biomedical Engineering
• Machine Learning in Healthcare
• Medical Imaging Analysis

Background:

• Accurate anthropometric body composition prediction (Appendicular Lean Mass, Body Fat Percentage, Bone Mineral Density) is crucial for diagnosing chronic diseases.
• Current gold standard, Dual-Energy X-ray Absorptiometry (DXA), is expensive and time-consuming.
• Healthcare data extraction faces significant technical and legal hurdles, limiting supervised machine learning model development.

Purpose of the Study:

• To propose and evaluate a novel semi-supervised p-Laplacian model as an alternative to DXA for body composition analysis.
• To assess the model's performance using biomarkers from 3D optical images.
• To address data scarcity challenges in healthcare machine learning.

Main Methods:

• Applied statistical and machine learning models, including a semi-supervised p-Laplacian approach, on biomarkers from 3D optical images.
• Utilized a dataset of 847 patients from the Pennington Biomedical Research Center.
• Compared the p-Laplacian model's performance against standard supervised algorithms like Support Vector Regression (SVR).

Main Results:

• The p-Laplacian model achieved prediction errors of ~13% for ALM, ~10% for BMD, and ~20% for BFP with only 10% of training data.
• SVR demonstrated superior performance for ALM and BMD (~8% error), while Least Squares SVR excelled for BFP (~11% error) when trained on 80% of the data.
• The p-Laplacian model effectively leverages unlabeled data, showing promise in data-constrained healthcare environments.

Conclusions:

• The game-theoretic p-Laplacian model is a viable and promising tool for regression tasks in healthcare, especially with limited labeled data.
• This approach offers a cost-effective and efficient alternative to DXA for body composition assessment.
• Semi-supervised learning models can overcome data limitations inherent in clinical settings.