Physical Features and Vital Signs Predict Serum Albumin and Globulin Concentrations Using Machine Learning

Area of Science:

• Biochemistry
• Medical Informatics
• Machine Learning

Background:

• Serum protein concentrations (albumin, globulin) are critical biomarkers for various diseases.
• Traditional blood tests for these proteins are invasive, costly, and inconvenient.
• Non-invasive prediction methods are needed to improve accessibility and early detection.

Purpose of the Study:

• To investigate the feasibility of predicting serum albumin, globulin, and albumin-globulin ratio using readily available physical characteristics and vital signs.
• To develop and evaluate machine learning models for this predictive task.

Main Methods:

• Utilized a dataset of 46,951 healthy adults from Hangzhou, China, including physical characteristics and vital signs.
• Trained computational models to predict serum protein concentrations from these predictors.
• Evaluated model accuracy on an independent dataset and determined feature importance.

Main Results:

• Achieved prediction accuracies (Pearson r) of 0.540 for albumin, 0.250 for globulin, and 0.373 for albumin-globulin ratio.
• Key predictors for albumin included age, gender, pulse, and BMI.
• Age, gender, BMI, and height were most important for predicting the albumin-globulin ratio.

Conclusions:

• Machine learning models can predict serum protein concentrations with notable accuracy using non-invasive data.
• This approach shows promise for augmenting existing screening tools and identifying individuals for further blood testing.
• Non-invasive prediction of serum proteins offers a potential advancement in disease risk assessment.