Transparent, Reproducible Text-Based Phenotyping of Lumbar Intervertebral Disc Degeneration from 500 Consecutive MRI Reports, with a Pre-Specified Image Analysis Framework.
Authors
Affiliations (7)
Affiliations (7)
- MRI Unit, Radiology Department, Imam Abdulrahman Al Faisal Hospital, First Health Cluster, Riyadh 14723, Saudi Arabia.
- Department of Radiologic Technology, College of Applied Medical Sciences, Qassim University, Buraidah 51452, Saudi Arabia.
- Physics Department, Faculty of Science, Al-Azhar University, Nasr City, Cairo 11884, Egypt.
- MRI Unit, Radiology Department, Buraidah Central Hospital, Qassim Health Cluster, Buraidah 52361, Saudi Arabia.
- MRI Unit, Radiology Department, King Saud Medical City (KSMC), First Health Cluster, Riyadh 12746, Saudi Arabia.
- Department of Adult Neurology and Epilepsy, Imam Abdulrahman Al Faisal Hospital, First Health Cluster, Riyadh 14723, Saudi Arabia.
- Al-Amal Medical Complex, Jazan 86622, Saudi Arabia.
Abstract
<b>Background/Objectives:</b> To characterize the degenerative vocabulary of 500 consecutive lumbar MRI reports using a transparent, reproducible text extraction pipeline; to quantify which descriptors distinguish included from excluded reports and which factors predict text-derived severity; and to specify a reproducible image analysis pipeline whose formal validation against radiologist Pfirrmann grading is defined as the next step. <b>Methods:</b> We analyzed 500 lumbar MRI reports (386 patients; December 2018-November 2025), extracting morphological and severity keywords, segmental levels (L1-L2 to L5-S1), Modic mentions, and a custom text-derived ordinal severity (0-3). Keyword prevalence differences were tested with the Fisher exact test under Benjamini-Hochberg false discovery rate control, with 95% confidence intervals (CIs) for all odds ratios (ORs) and risk differences; predictors of severity were examined by ordinal logistic regression. Robustness to within-patient clustering was assessed with patient-clustered robust estimation and one-report-per-patient analyses. A MATLAB pipeline derived per-disc candidate imaging features (predicted Pfirrmann grade, normalized disc height, T2 signal index, quality control) across 316 studies. <b>Results:</b> In total, 385/500 reports (77.0%) met the inclusion criteria. The structured morphological keyword field was missing in 49.4% of records, varying by year (58.6% in 2020, 68.7% in 2023, 41.9% in 2024, 33.5% in 2025), indicating systematic reporting drift. Bulge was the dominant descriptor (53.8%); L4-L5 and L5-S1 dominated level mentions. After correction, bulge (OR 6.70, 95% CI 3.86-11.65), mild (5.08, 2.39-10.78), dehydration (9.82, 2.36-40.87) and central (8.63, 2.07-36.01) were enriched among the included reports. All four remained significant in clustering-aware sensitivity analyses. Older age independently predicted higher severity (OR 1.36 per 10 years, 95% CI 1.19-1.56). The image pipeline produced per-disc candidate biomarkers across 316 studies. <b>Conclusions:</b> Free-text lumbar MRI reports encode a recognizable but heterogeneous degenerative vocabulary, sufficient for cohort construction yet inconsistent for quantitative grading; the text-derived severity is a noisy proxy. The reproducible image analysis pipeline yields candidate biomarkers whose formal validation against an adjudicated radiologist Pfirrmann reference standard is the explicit, pre-specified next step.