JMIR Publications

ABSTRACT

Background:

Background:

Patients with persistent spinal pain after spine surgery may undergo prolonged medication escalation and repeated procedures before referral for spinal cord stimulation (SCS) or neuromodulation evaluation.

Objective:

Objective:

To develop and internally evaluate referral-support models for future SCS/neuromodulation documentation among patients with failed back surgery syndrome/persistent spinal pain syndrome type 2 in a tertiary pain clinic.

Methods:

Methods:

We performed a retrospective patient-level electronic health-record analysis using two prespecified horizons. The index model used features available at or before the earliest FBSS/post-spine-surgery pain-clinic presentation. The 90-day landmark model used features accumulated within the first 90 days of tertiary pain-clinic care. Penalized logistic regression was the primary model; random forest and histogram gradient boosting were exploratory. Leakage-prone future variables, no-variance predictors, and exact duplicates were excluded. Performance was assessed using AUROC, AUPRC, Brier score, calibration, threshold metrics, feature importance, and decision curves.

Results:

Results:

The index cohort included 572 patients; 116 (20.3%) had future SCS/neuromodulation documentation. The primary index model achieved random test AUROC 0.633 (95% CI, 0.476-0.772), AUPRC 0.310, and Brier score 0.161. The 90-day landmark cohort included 557 patients; 101 (18.1%) had future SCS documentation. The primary 90-day model achieved random test AUROC 0.772 (95% CI, 0.636-0.876), AUPRC 0.423, and Brier score 0.129. Temporal tests were underpowered, with only 10 and 6 events.

Conclusions:

Conclusions:

A 90-day landmark referral-support model demonstrated encouraging internal random-test performance for future SCS pathway entry, but external and prospective validation are required before workflow implementation.