Machine learning versus logistic regression for prognostic modelling in individuals with non-specific neck pain

Area of Science:

• Clinical epidemiology
• Biostatistics
• Health informatics

Background:

• Prognostic models are crucial for managing neck pain disorders.
• Previous research has not compared modern machine learning (ML) with traditional regression for neck pain prognostication.

Purpose of the Study:

• To compare the performance of ML techniques against traditional regression in developing prognostic models for individuals with neck pain.

Main Methods:

• Utilized a clinical registry of 3001 neck pain patients.
• Assessed three dichotomous outcomes at 3-month follow-up: neck pain, arm pain, and disability improvement.
• Compared seven modeling techniques: logistic regression, LASSO, Xgboost, KNN, SVM, RF, and ANN, using Area Under the Curve (AUC) as the primary performance metric.

Main Results:

• Gradient boosting (Xgboost) demonstrated the highest AUC for predicting arm pain (0.765), neck pain (0.726), and disability (0.703).
• ML algorithms improved classification AUC by 0.081 to 0.103 over stepwise logistic regression across the outcome measures.

Conclusions:

• ML methods, especially Xgboost, offer superior predictive performance for neck pain prognosis compared to logistic regression.
• The enhanced prediction by ML may stem from modeling greater nonlinearity between predictors and outcomes.
• The benefits of ML in prognostic modeling are influenced by sample size, variable type, and the specific disease studied.