Determination of Candidate Predictors for Chiropractic Treatment Outcome of Spinal Pain Using a Machine Learning Framework for Small Datasets

It proposes and demonstrates a three-step machine learning framework for small, phenotypically rich datasets to select candidate predictors of chiropractic treatment outcome in spinal pain from self-reported questionnaires. Relative to prediction-modeling paths that typically rely on large samples, this work combines SHAP-based feature selection, Leave-One-Out Cross-Validation with permutation testing, and generalization testing into a pipeline that can run on a cohort of N = 96. Prospective cohort design with 96 patients (mean age 44.5 ± 16.5 years; 53 female), baseline and 1- and 3-month follow-up, four binary outcomes, and explicitly listed methodological steps.

It reports directional signals for candidate predictors: positive treatment expectations, higher self-efficacy, younger age, lower body mass index, and fewer comorbidities were associated with higher recovery odds, while psychological dysfunction generally hindered recovery. These factors emerge from SHAP interpretation of a multi-domain self-report questionnaire covering pain, mood, sleep, lifestyle, and treatment expectations, rather than from a single clinical variable. Interpretive results based on SHAP analysis; these are exploratory associations, and the text does not report effect sizes or confidence intervals for individual factors.

It finds that pain location (lower back, upper back, or neck) and pain duration did not appear important for prediction, while sleep quality and certain pain features showed varying influence across the four outcomes. By placing location and duration, usually treated as core clinical features, alongside psychosocial and lifestyle factors, it suggests the latter may carry more predictive weight. Negative or inconsistent signals from the same SHAP framework, with a limited sample size; the text frames these as exploratory findings.

Through the performance gap between Leave-One-Out Cross-Validation and ensemble cross-validation, it explicitly positions the current results as hypothesis-generating rather than ready for clinical decision-making. The authors proactively report the AUC drop from 0.93-0.99 to 0.62-0.90 and, on that basis, call for prospective replication in adequately powered cohorts. Both AUC ranges are given directly in the text, and the performance gap is used by the authors as the basis for interpretive caution.

The framework targets settings with small but phenotypically rich self-report data, applicable to spinal pain populations for whom multi-domain information on pain, mood, sleep, lifestyle, and treatment expectations is collected before treatment; its goal is to select candidate predictors and generate hypotheses rather than to deliver a deployable individualized prediction tool. It is of reference value to researchers who wish to conduct exploratory predictive modeling in small cohorts and need a pipeline that includes feature interpretation and validation steps.

Readers may still wonder about the specific sources of the AUC gap between Leave-One-Out Cross-Validation and ensemble cross-validation, which the text does not further decompose; effect sizes and uncertainty intervals for individual candidate predictors are absent; the reasons for inconsistent influence of sleep quality and pain features across outcomes remain to be clarified; whether the negative signals for pain location and duration hold in larger samples is open; and how the framework performs outside chiropractic treatment and outside a Swiss specialist clinic setting is unknown. In addition, the available text is abstract-level content without figures or supplementary material, so if those contain feature-importance rankings or model details, the scope of what can be summarized here is correspondingly limited.