Berg B, Gorosito MA, Fjeld O, Haugerud H, Storheim K, Solberg TK, Grotle M. Machine Learning Models for Predicting Disability and Pain Following Lumbar Disc Herniation Surgery.
JAMA Netw Open 2024;
7:e2355024. [PMID:
38324310 PMCID:
PMC10851101 DOI:
10.1001/jamanetworkopen.2023.55024]
[Citation(s) in RCA: 0] [Impact Index Per Article: 0] [Reference Citation Analysis] [Abstract] [MESH Headings] [Track Full Text] [Journal Information] [Submit a Manuscript] [Subscribe] [Scholar Register] [Received: 08/08/2023] [Accepted: 12/14/2023] [Indexed: 02/08/2024] Open
Abstract
Importance
Lumber disc herniation surgery can reduce pain and disability. However, a sizable minority of individuals experience minimal benefit, necessitating the development of accurate prediction models.
Objective
To develop and validate prediction models for disability and pain 12 months after lumbar disc herniation surgery.
Design, Setting, and Participants
A prospective, multicenter, registry-based prognostic study was conducted on a cohort of individuals undergoing lumbar disc herniation surgery from January 1, 2007, to May 31, 2021. Patients in the Norwegian Registry for Spine Surgery from all public and private hospitals in Norway performing spine surgery were included. Data analysis was performed from January to June 2023.
Exposures
Microdiscectomy or open discectomy.
Main Outcomes and Measures
Treatment success at 12 months, defined as improvement in Oswestry Disability Index (ODI) of 22 points or more; Numeric Rating Scale (NRS) back pain improvement of 2 or more points, and NRS leg pain improvement of 4 or more points. Machine learning models were trained for model development and internal-external cross-validation applied over geographic regions to validate the models. Model performance was assessed through discrimination (C statistic) and calibration (slope and intercept).
Results
Analysis included 22 707 surgical cases (21 161 patients) (ODI model) (mean [SD] age, 47.0 [14.0] years; 12 952 [57.0%] males). Treatment nonsuccess was experienced by 33% (ODI), 27% (NRS back pain), and 31% (NRS leg pain) of the patients. In internal-external cross-validation, the selected machine learning models showed consistent discrimination and calibration across all 5 regions. The C statistic ranged from 0.81 to 0.84 (pooled random-effects meta-analysis estimate, 0.82; 95% CI, 0.81-0.84) for the ODI model. Calibration slopes (point estimates, 0.94-1.03; pooled estimate, 0.99; 95% CI, 0.93-1.06) and calibration intercepts (point estimates, -0.05 to 0.11; pooled estimate, 0.01; 95% CI, -0.07 to 0.10) were also consistent across regions. For NRS back pain, the C statistic ranged from 0.75 to 0.80 (pooled estimate, 0.77; 95% CI, 0.75-0.79); for NRS leg pain, the C statistic ranged from 0.74 to 0.77 (pooled estimate, 0.75; 95% CI, 0.74-0.76). Only minor heterogeneity was found in calibration slopes and intercepts.
Conclusion
The findings of this study suggest that the models developed can inform patients and clinicians about individual prognosis and aid in surgical decision-making.
Collapse