Deep learning health space model for ordered responses.

Area of Science:

• Biomedical informatics
• Computational biology
• Personalized medicine

Background:

• Objective health status measurement is crucial for personalized medicine but challenged by increasing data dimensions.
• Existing statistical Health Space (HS) models, like logistic regression and proportional odds models, struggle with complex non-linear biological relationships.

Purpose of the Study:

• To develop and evaluate deep learning-based Health Space (HS) models capable of capturing complex non-linear biological relationships.
• To compare the performance of deep learning HS models against established statistical HS models.

Main Methods:

• Five deep learning HS models were formulated, including four binary deep neural networks (DNNs) and one deep ordinal neural network (DONN).
• Models were trained on 32,140 samples from the Korea National Health and Nutrition Examination Survey (KNHANES) and validated on external datasets (Ewha-Boramae cohort and KARE project).

Main Results:

• The deep learning HS models, particularly the deep ordinal neural network (DONN), demonstrated superior performance in discriminating health status.
• DONN outperformed the existing statistical HS model based on the proportional odds model (POM) in both training and external validation datasets.

Conclusions:

• Deep learning HS models effectively capture complex non-linear biological relationships for health status visualization.
• These models offer a promising approach for objective and meaningful visualization of individual health status in the era of personalized medicine.