Machine learning-based prediction of disability in subacute low back pain: a primary care study on clinical and
José Alberto Benítez-Andrades1, Luis Laballos-González2, Lorena Pujante-Otalora2
1Advanced Learning for Biomedical Analytics (ALBA) Laboratory, Escuela de Ingenierías Industrial, Informática y Aeroespacial, Universidad de León, Campus de Vegazana s/n, León, 24071, Spain. jbena@unileon.es.
Background:
Subacute low back pain (LBP) is a highly prevalent condition and a major contributor to disability and health care burden. Early identification of individuals at risk of poor functional recovery is essential to support decision-making in primary care. Although prior research has identified relevant clinical and psychosocial predictors, the application of machine learning techniques for modeling disability and pain outcomes in this population remains limited.
Methods:
We conducted a prospective observational study involving 92 adult patients with subacute LBP attending primary care physiotherapy services. At baseline, participants completed standardized assessments of disability (percentage of perceived limitation in daily activities), pain intensity (numeric rating scale), self-efficacy (confidence to perform activities despite pain), fear of movement, and psychosocial risk classification. Predictive models were developed for categorical outcomes of functional disability and pain intensity at discharge and at three-month follow-up using Gaussian Naive Bayes, Complement Naive Bayes, k-Nearest Neighbors, and Decision Tree classifiers. Models were trained under three feature set configurations (intrinsic, extrinsic, combined) and evaluated with and without oversampling techniques (RandomOverSampler, Synthetic Minority Oversampling Technique). Performance metrics included accuracy and F1-score. Feature importance analysis was performed using SelectKBest and ExtraTreesClassifier.
Results:
The best results were obtained with the Decision Tree classifier using the combined feature set (accuracy = 0.81, F1-score = 0.80). Baseline pain intensity was the most relevant predictor for disability outcomes, while baseline disability was most influential for pain intensity predictions. Psychosocial factors-self-efficacy, kinesiophobia, and psychosocial risk-showed moderate contributions. Age and sex had minimal predictive impact.
Conclusion:
Machine learning models, particularly Decision Tree classifiers, can accurately predict functional disability and pain intensity in patients with subacute LBP using routinely collected clinical and psychosocial data. Such models could support early risk stratification and facilitate the development of clinical decision support tools for personalized care in primary care physiotherapy.
Trial Registration:
ClinicalTrials.gov NCT05860426. Registered on April 26, 2023.
