Oral Presentation 50th International Society for the Study of the Lumbar Spine Annual Meeting 2024

INTRODUCTION:

Physical activity has a complex association with pain trajectory in chronic low back pain (cLBP) where biopsychosocial factors makes it challenging to discern causality. Recent findings establish links between day-to-day activity variability and self-reported flare-ups, suggesting that activity measures may be objective biomarkers of pain trajectory. We hypothesized that step count trajectories would outperform patient reported outcomes (PROs) in classifying episodic vs. non-episodic cLBP groups. Further, we hypothesized that local activity variability (day-to-day) would be more predictive of episodic pain than summary measures (average daily step count). To investigate this, we extracted temporal features from six-months of daily step count data, along with PROs from cLBP patients to build a series of logistic regression models predicting episodic pain trajectory and, using regularization, identified features predictive of episodic pain.

METHODS:

238 cLBP patients completed pain, psychosocial, and demographic surveying, including the Visual Trajectories Questionnaire-Pain (VTQ-Pain). Patients selected the visual pain trajectory describing their pain the past 6 months then were binned into ‘episodic’ or ‘non-episodic’ pain groups based on response (Fig. 1). We retrospectively collected daily step counts (iPhone HealthKit) for the six-month period described in VTQ-Pain. We used functional principal component analysis (fPCA) to identify global trends in step count trajectories, then calculated 615 additional generalized time features. We then built three elastic-net logistic regression models (3-fold cross-validation, 25 repeats) to predict pain trajectory classification using, as predictors: 1. only activity features, 2. only PROs, 3. both activity features and PROs. For each, we calculated mean training data prediction accuracy using the best models across folds then used student t-test to compare models. Final model performance was scored using testing data prediction accuracy and the area under the receiver-operating curve (AUC).

RESULTS:

Of 238 patients, 77.3% reported an ‘episodic’ pain trajectory and 22.7% reported a ‘non-episodic’ trajectory with no age or sex differences between groups. Training accuracy did not differ between activity-based (73.4%±5.7%) and PRO-based (74%±3.9%) models. Test accuracy of the activity-based model (66.7%) outperformed the PRO-based model (59.7%) by 7% points. The training accuracy of the combined model did not differ from the activity-based model but was 3.5% points lower (70.5%±4.8%; pval<.01) than the PRO-based model. The test accuracy of the PRO-based model and the combined model were identical. Further, the AUC of the activity-based model (.67) was higher than the PRO-based (.58) and the combined model (.57). The activity features with the largest coefficients (importance) were measures of frequency (fast Fourier transform (FFT) coefficients), energy and local variability (9-day autocorrelation). The most important PROs were related to pain during activity and fear of pain (Fig.1B)

DISCUSSION:

While model performances were low, the activity-based model performed comparably or better than the PRO-based model, suggesting that step count trajectories may be more informative for tracking episodic pain than PROs in later assessments. Features selected by regularization highlight the importance of ‘local’ variability in activity trajectories, which may be better biomarkers of pain trajectory than summary measures. This has important implications for the development of objective, continuous pain monitoring.