Clinical predictive modelling of post-surgical recovery in individuals with cervical radiculopathy: a machine learning approach

Bernard X. W. Liew, Anneli Peolsson, David Rügamer, Johanna Wibault, Håkan Löfgren, Åsa Dedering, Peter Zsigmond, Deborah FallaScientific 2020 · DOI

Abstract

Prognostic models are vital for managing cervical radiculopathy (CR), yet the comparative performance of modern machine learning (ML) versus traditional statistical methods remains under-explored in this population. This study analyzed a prospective cohort of 201 individuals with CR using four modelling techniques: stepwise regression, LASSO, boosting, and multivariate adaptive regression splines (MuARS). Models were developed for four 12-month outcomes: disability (NDI), quality of life (EQ5D), neck pain, and arm pain. Results indicated that differences in mean predictive performance were clinically negligible across all methods. However, ML techniques like MuARS offered advantages in model parsimony and the ability to identify non-linear relationships, suggesting their utility in probing complex clinical recovery patterns.

Introduction: The Challenge of Predicting CR Recovery

Cervical radiculopathy (CR), characterized by nerve root compression in the neck causing pain and disability, imposes a significant health burden. While surgical intervention is often effective, individual outcomes vary considerably. Accurate prognostic models are essential for managing patient expectations and optimizing clinical resource allocation. Historically, "stepwise regression" has been the standard for building these models, despite criticisms regarding over-optimism and biased coefficients. Modern machine learning offers a potential alternative through automated variable selection and non-linear modeling, but its practical superiority in CR prognosis required rigorous validation.

Methodology

The researchers evaluated four distinct modelling techniques to predict recovery outcomes:

  1. Stepwise Regression: The traditional approach iteratively selecting variables based on p-values.
  2. LASSO (Least Absolute Shrinkage and Selection Operator): A penalized regression method that performs automatic variable selection by shrinking less relevant coefficients.
  3. Model-based Boosting: An iterative technique that builds models by sequentially correcting errors from previous iterations.
  4. MuARS (Multivariate Adaptive Regression Splines): A non-linear approach using "hinge functions" to capture complex interactions between predictors.

The study utilized data from 201 individuals across four Swedish spinal surgery centers. Predictors included 85 variables spanning demographics, neurological tests, physical function, and psychological assessments. Outcomes were measured at 12 months using the Neck Disability Index (NDI), EuroQol 5D (EQ5D), and Visual Analogue Scales (VAS) for pain. Robustness was ensured via 10-fold cross-validation repeated 50 times.

Key Results and the "Accuracy Plateau"

A primary finding was the "accuracy plateau": modern ML did not significantly outperform traditional regression in raw predictive accuracy. As shown in Figure 1, the differences in Root Mean Square Error (RMSE) between the four models were clinically minor across all outcomes—for instance, less than 1 point on the 50-point NDI scale.

Comparison of Root Mean Square Error (RMSE) across four modelling techniques for disability (a), quality of life (b), neck pain (c), and arm pain (d).

Despite similar accuracy, the models differed in complexity. ML methods, particularly MuARS, often produced more parsimonious models, achieving comparable results with fewer predictors. Furthermore, MuARS identified critical non-linear thresholds; for example, baseline disability only predicted future disability when the initial NDI score exceeded a specific threshold (>9).

Consistent predictors emerged across models: baseline NDI was a robust indicator for future disability and neck pain, while somatic awareness (MSPQ) predicted quality of life. A simple neurological exam—the C6 light touch test—was a key predictor for arm pain intensity.

Discussion and Future Directions

The study's findings suggest that while ML may not provide a "silver bullet" for accuracy in smaller clinical datasets (N=201), its ability to handle non-linearity and produce simpler models is highly valuable. The authors note that the "data-hungry" nature of ML might mean its full potential is only realized with larger, multi-center cohorts.

Future research should move beyond linear assumptions. The "hinge" relationships identified by MuARS indicate that the physiological and psychological processes of recovery are complex. Applying these advanced methods to larger datasets will be crucial for refining prognostic tools and improving personalized care for patients with cervical radiculopathy.

Summary updated

Abstract

Prognostic models play an important role in the clinical management of cervical radiculopathy (CR). No study has compared the performance of modern machine learning techniques, against more traditional stepwise regression techniques, when developing prognostic models in individuals with CR. We analysed a prospective cohort dataset of 201 individuals with CR. Four modelling techniques (stepwise regression, least absolute shrinkage and selection operator [LASSO], boosting, and multivariate adaptive regression splines [MuARS]) were each used to form a prognostic model for each of four outcomes obtained at a 12 month follow-up (disability-neck disability index [NDI]), quality of life (EQ5D), present neck pain intensity, and present arm pain intensity). For all four outcomes, the differences in mean performance between all four models were small (difference of NDI < 1 point; EQ5D < 0.1 point; neck and arm pain < 2 points). Given that the predictive accuracy of all four modelling methods were clinically similar, the optimal modelling method may be selected based on the parsimony of predictors. Some of the most parsimonious models were achieved using MuARS, a non-linear technique. Modern machine learning methods may be used to probe relationships along different regions of the predictor space.
BibTeX
@inproceedings{liew2020clinical,
  author = {Bernard X. W. Liew and Anneli Peolsson and David Rügamer and Johanna Wibault and Håkan Löfgren and Åsa Dedering and Peter Zsigmond and Deborah Falla},
  title = {{Clinical predictive modelling of post-surgical recovery in individuals with cervical radiculopathy: a machine learning approach}},
  booktitle = {Scientific},
  year = {2020},
  doi = {10.1038/s41598-020-73740-7},
}
Figure