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

Bernard X. W. Liew, Francisco M. Kovacs, David Rügamer, Ana RoyuelaEuropean 2022 · DOI

Abstract

The improvement in prediction performance between ML and logistic regression methods in the present study could be due to the potential greater nonlinearity between baseline predictors and clinical outcome. The benefit of machine learning in prognostic modelling may be dependent on factors like sample size, variable type, and disease investigated.

Introduction

Non-specific neck pain is a leading cause of global disability, yet predicting patient recovery remains a significant clinical challenge. Traditionally, prognostic models have relied on logistic regression (LR). While LR is valued for its interpretability, it often struggles to capture the complex, non-linear relationships inherent in clinical data. As machine learning (ML) techniques become more prevalent in healthcare, this study evaluates whether the added complexity of ML algorithms translates into superior predictive performance for individuals with neck pain.

Methodology

The study utilized a large-scale clinical registry from the Spanish Back Pain Research Network (REIDE), comprising 3,001 participants. Researchers aimed to predict three primary outcomes at a 3-month follow-up: improvement in neck pain, arm pain, and disability.

The models incorporated 26 baseline predictors, ranging from demographic data like age and employment status to clinical findings such as pain intensity and imaging results.

Baseline variables and their characteristics used in the prognostic models.

A traditional stepwise logistic regression was compared against several ML algorithms, including LASSO, K-nearest neighbors (KNN), Random Forest (RF), Support Vector Machines (SVM), Artificial Neural Networks (ANN), and Gradient Boosting (Xgboost). To ensure robustness, the team employed nested cross-validation and extensive hyperparameter optimization.

Results

The analysis demonstrated that ML algorithms generally outperformed traditional logistic regression across all outcomes. The primary metric for evaluation was the Area Under the Receiver Operating Characteristic curve (AUC).

Performance comparison across metrics and models for arm pain, neck pain, and disability.

Key findings include:

  • Top Performer: Xgboost consistently emerged as the most effective model for all three clinical outcomes.
  • Predictive Gains: Xgboost achieved an AUC of 0.765 for arm pain (an improvement of 0.081 over LR), 0.726 for neck pain (improvement of 0.103), and 0.703 for disability (improvement of 0.077).
  • Metric Nuances: While ML dominated in discrimination (AUC), stepwise LR showed high accuracy in specific instances, such as predicting arm pain. Random Forest exhibited exceptional specificity, reaching 0.958 for neck pain improvement.

Discussion

The superiority of ML in this context likely stems from its ability to model non-linear interactions between predictors that traditional regression might miss. This represents a shift from statistical inference (understanding data generation) toward predictive accuracy, which is often more critical for clinical decision-making.

However, several limitations persist. The models currently lack calibration—meaning they rank patients by risk well but do not yet provide precise individual probabilities of recovery. Furthermore, the reliance on discrete, single-point measurements suggests that incorporating functional, high-dimensional data (e.g., from wearable sensors) could further enhance the predictive gap between ML and traditional methods.

Conclusion

Machine learning, particularly ensemble methods like Xgboost, offers a significant advantage over traditional logistic regression for prognosticating neck pain recovery. While further work is needed to ensure clinical meaningfulness and calibration, these findings suggest that sophisticated modeling can better navigate the complexities of musculoskeletal pain prognosis.

Summary updated

Abstract

The improvement in prediction performance between ML and logistic regression methods in the present study, could be due to the potential greater nonlinearity between baseline predictors and clinical outcome. The benefit of machine learning in prognostic modelling may be dependent on factors like sample size, variable type, and disease investigated.
BibTeX
@inproceedings{liew2022machine,
  author = {Bernard X. W. Liew and Francisco M. Kovacs and David Rügamer and Ana Royuela},
  title = {{Machine learning versus logistic regression for prognostic modelling in individuals with non-specific neck pain}},
  booktitle = {European},
  year = {2022},
  doi = {10.1007/s00586-022-07188-w},
}
Figure