Applying binary mixed model to predict knee osteoarthritis pain.

Area of Science:

• Biomedical data science
• Rheumatology
• Machine learning in healthcare

Background:

• Knee osteoarthritis (KOA) data often lacks person-level context, focusing on individual knees.
• Ignoring correlations between a person's knees can lead to inaccurate KOA outcome predictions.
• Existing machine learning models typically analyze each knee independently, overlooking within-person correlations.

Purpose of the Study:

• To develop a flexible, data-driven framework for predicting knee osteoarthritis pain outcomes.
• To incorporate advantages of random forest (RF) and mixed effects models for correlated data.
• To compare a novel Binary Mixed Models (BiMM) algorithm against standard RF for KOA pain prediction.

Main Methods:

• Utilized baseline data from the Osteoarthritis Initiative (OAI) cohort.
• Applied the Binary Mixed Models (BiMM) algorithm to predict two binary KOA pain indicators.
• Compared BiMM performance against standard random forests (RF) that do not account for knee correlation.

Main Results:

• The BiMM algorithm demonstrated potential as a predictive tool for KOA pain.
• BiMM achieved comparable or slightly improved performance over traditional RF models.
• BiMM successfully accounted for within-person correlations among knees, unlike standard RF.

Conclusions:

• BiMM offers a nuanced, person-level approach to KOA outcome prediction, addressing limitations of existing models.
• This method advances the understanding of KOA pain by considering within-person joint correlations.
• The BiMM framework is applicable to other musculoskeletal conditions and situations with within-person correlations.