A single-source preoperative prediction model for recurrent lumbar disc herniation after primary discectomy: Development and internal validation in the SPORT registry

Abstract

Background

Recurrent lumbar disc herniation (rLDH) follows primary discectomy in 5% to 25% of patients. Existing models often embed postoperative variables, merge data from unlinked repositories, or omit calibration and clinical-utility assessment. We aimed to develop and internally validate a strictly preoperative model for rLDH using a single multi-center registry, with calibration-first evaluation, decision curve analysis, and a SHAP-aligned clinical score (SHapley Additive exPlanations).

Methods

This retrospective development and split-sample internal validation study used the Spine Patient Outcomes Research Trial (SPORT) lumbar disc herniation cohort; 738 of 1,244 enrolled patients met inclusion criteria. The outcome was symptomatic rLDH at the index level within 2 years, confirmed radiologically or by reoperation. Thirteen preoperative predictors and 4 algorithms (elastic-net logistic regression, random forest, gradient boosting, support vector machine) were evaluated, trained on 70% with stratified 5-fold cross-validation and 500-resample optimism correction, then tested on a 30% holdout. Reporting followed TRIPOD+AI.

Results

Recurrence occurred in 110 patients (14.9%). Logistic regression performed best: AUC 0.738 (95% CI 0.694–0.782; optimism-corrected 0.724), calibration slope 0.94, Brier score 0.112. SHAP identified annular defect size, smoking, and younger age as the most influential predictors. A 10-item score stratified patients into low- (5.8%), moderate- (15.2%), and high-risk (32.6%) groups. Decision curve analysis showed net benefit across threshold probabilities of 8% to 30%.

Conclusions

A timing-consistent, single-registry preoperative model achieved acceptable discrimination and calibration for rLDH. The SHAP-aligned score offers transparent preoperative risk stratification; external validation is required before clinical use.

Graphical Abstract

(A) Mean absolute SHAP feature importance for the ten scored preoperative predictors; the 3 most influential predictors (annular defect size, current smoking, younger age) are highlighted. (B) Observed 2-year recurrence rate across low-, moderate-, and high-risk score categories in the holdout test set, showing a 5.6-fold gradient. Model: elastic-net logistic regression, AUC 0.738 (optimism-corrected 0.724), calibration slope 0.94, Brier score 0.112.

Introduction

Lumbar disc herniation is among the most prevalent spinal pathologies, affecting an estimated 1% to 3% of the general population annually and constituting a leading cause of radiculopathy and functional disability [[1], [2], [3]]. For patients with persistent or progressive neurological deficits despite conservative management, surgical discectomy remains the standard intervention, with demonstrated short-term efficacy in appropriately selected candidates [1,2]. However, recurrent lumbar disc herniation (rLDH) remains a major problem after surgery. rLDH is defined as symptomatic re-herniation at the same disc level following an initial pain-free interval, and it occurs in 5% to 25% of patients [[4], [5], [6], [7], [8]]. The reported rate varies with the recurrence definition used, the surgical technique, and the duration of follow-up. Recurrence is associated with increased morbidity, prolonged disability, and the frequent need for revision surgery, which carries less predictable outcomes than primary procedures [4,6,8].

Multiple patient-level risk factors for rLDH have been identified in individual surgical series and confirmed in meta-analyses. Demographic and clinical risk factors include younger age, elevated body mass index (BMI), current smoking, and diabetes mellitus. Imaging-based risk factors include advanced disc degeneration as graded by the Pfirrmann classification (a 5-level MRI-based grading of nucleus pulposus signal and disc height), Modic endplate changes (vertebral marrow signal alterations on T1- and T2-weighted MRI), large annular defects, preserved disc height, and disc fragment morphology [[4], [5], [6], [7], [8], [9], [10], [11], [12], [13]]. Despite this accumulated knowledge of individual predictors, integrated multivariate prediction tools for preoperative risk stratification remain underdeveloped. Machine learning approaches have been applied to spine surgery outcome prediction, including models for reoperation after endoscopic discectomy, discharge disposition, and readmission [[14], [15], [16], [17], [18], [19], [20]]. However, the prediction literature specific to rLDH has several recurring methodological limitations.

First, prediction timing is frequently inconsistent. Models described as preoperative tools often include variables that are only available after surgery, such as rehabilitation compliance or return-to-work timing. This prevents genuine preoperative application [16,17]. Second, some studies merge patient-level variables from multiple public datasets without demonstrating patient-level linkage. For example, clinical registries have been combined with separate imaging repositories, introducing measurement heterogeneity of uncertain magnitude. Third, model evaluation is often limited to discrimination (the area under the receiver operating characteristic curve, or AUC) without calibration assessment, optimism correction, or decision-analytic utility evaluation. This pattern has been criticized in recent methodological reviews of clinical prediction studies [[21], [22], [23], [24], [25]]. Fourth, clinical scoring systems derived from machine learning models sometimes assign point weights that contradict the model’s own feature importance ranking. This undermines the auditability that interpretable AI methods are intended to provide [26,27].

Rather than proposing the first machine learning approach to rLDH prediction, the present study addresses a specific methodological gap. We aim to develop a transparent, timing-consistent, single-source prediction model with 4 design principles: strictly preoperative variable selection; no fusion with unlinked external datasets; calibration reported alongside discrimination; and a clinical score whose point assignments are directly traceable to model-level feature importance. We used exclusively the Spine Patient Outcomes Research Trial (SPORT) registry [1,2], a well-characterized multi-center cohort with prospectively collected data, archived baseline imaging, and extended follow-up. This paper reports the development phase of a planned development–validation pipeline: the model architecture, predictor weights, and scoring system reported here are intended to be carried forward to external validation in independent contemporary cohorts. Secondary objectives included time-to-recurrence characterization across derived risk strata and PROBAST-informed self-assessment of risk of bias. Reporting follows TRIPOD+AI principles [21,22].

Methods

Study design and ethics

This was a retrospective prediction model development and internal validation study using random split-sample validation with resampling-based hyperparameter tuning within the training partition. The study protocol was developed before data analysis (Supplementary Document 1). Reporting follows the Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD+AI) principles [21,22]; a completed checklist is provided as Supplementary Document 2. A self-assessment informed by the Prediction model Risk Of Bias ASsessment Tool (PROBAST) was conducted and is reported as Supplementary Document 3 [23].

Ethics approval and institutional review board statement

This study received an institutional review board (IRB) exemption determination. Because all data were de-identified and publicly available through the National Institutes of Health (NIH) data repository, the study met criteria for exemption from institutional review board oversight under United States federal regulations (45 CFR 46.104(d)(4)). The original SPORT trial obtained IRB approval at all participating centers, and the present secondary analysis of de-identified data required no additional patient consent.

Data source

The sole data source was the SPORT lumbar disc herniation cohort [1,2]. SPORT was a multi-center prospective study conducted at 13 academic medical centers in the United States between 2000 and 2004. It enrolled 1,244 patients with imaging-confirmed lumbar disc herniation and radicular symptoms persisting despite at least 6 weeks of nonoperative treatment. The dataset was accessed through the NIH/National Institute of Arthritis and Musculoskeletal and Skin Diseases (NIAMS) data repository under a data use agreement. No external datasets were merged, linked, or used to supplement patient-level variables in this analysis. All predictor variables were derived exclusively from the SPORT dataset. Demographic and clinical variables came from baseline enrollment forms. Radiological variables came from standardized baseline MRI readings performed as part of the original SPORT imaging protocol [28,29].

Study population

The target population comprised adults with symptomatic single-level lumbar disc herniation who underwent primary discectomy within the SPORT cohort, including both patients randomized to surgery and those who crossed over to surgical treatment. Inclusion required age 18 years or older, a single herniation level confirmed on imaging, primary open or microscopic discectomy without concurrent fusion, at least 2 years of clinical follow-up, and complete preoperative baseline data. Patients managed nonoperatively throughout follow-up, those who received concurrent fusion, those with incomplete preoperative variables, and those lost to follow-up before 2 years were excluded. The selection process is detailed in Fig. 1. From 1,244 enrolled patients, 434 did not undergo discectomy during follow-up and were excluded. Of the 810 surgical patients, 34 had concurrent fusion (n = 776), 24 had incomplete preoperative data (n = 752), and 14 were lost before 2 years (n = 738). At each step the sample decreased from its predecessor, and the final analytic cohort comprised 738 patients.

Fig. 1.

Patient selection flow diagram. All exclusion steps reduce the cohort from the SPORT enrollment of 1,244. No external datasets were merged.

Outcome definition

The primary endpoint was symptomatic recurrent lumbar disc herniation at the same level and side, occurring after an asymptomatic interval following index surgery and confirmed radiologically or resulting in reoperation within 2 years of the index discectomy [[4], [5], [6], [7], [8]]. Patients without recurrence at the 2-year mark were censored at 24 months for the binary primary endpoint, irrespective of whether longer follow-up was available; the 110 events recorded as “recurrence” therefore correspond to events occurring within the 2-year window. Reoperation at the index level was additionally analyzed as a prespecified secondary endpoint to assess robustness against outcome-definition heterogeneity. For the primary classification models, recurrence was treated as a binary variable within this 2-year assessment window. Complementary time-to-event analyses used the interval from surgery to first documented recurrence and exploited the full available follow-up (median 4.2 years), with patients alive and recurrence-free at last contact administratively censored.

Recurrence status was derived from the SPORT structured dataset using the surgical follow-up records to identify reoperation events at the index lumbar level and the postoperative imaging records together with patient-reported symptom data to identify symptomatic radiological recurrence in patients who did not undergo revision. A patient was coded as having recurrence within 2 years if either condition was met at any follow-up visit through the 24-month assessment. Detailed variable provenance, including the SPORT codebook fields used and the algorithmic coding rules, is documented in Supplementary Table S1A.

Sensitivity analyses applied alternative outcome definitions to assess robustness: (1) a reoperation-only definition, restricted to patients who underwent revision discectomy at the index level within 2 years, and (2) a 1-year binary recurrence window. Results of these sensitivity analyses are reported in Supplementary Table S2.

Predictor variables: timing and provenance

Candidate predictors were restricted to variables available before surgery or at the time of preoperative imaging review. This ensures that the model can be applied at the clinical decision point preceding discectomy. Variables that require postoperative observation—rehabilitation compliance, postoperative activity level, and return-to-work timing—were excluded by design even though they appear in the SPORT dataset. Their inclusion would prevent genuine preoperative application and introduce temporal bias [23,24].

The 13 candidate predictors span 3 domains. Five are demographic or clinical: age, sex, BMI, current smoking status, and diabetes mellitus. Seven are radiological, all derived from baseline preoperative MRI as described below: herniation level, Pfirrmann disc degeneration grade, Modic endplate changes, disc height index, annular defect size, disc migration, and sagittal range of motion. One is occupational: self-reported occupational physical load. Complete predictor provenance—including the original SPORT codebook field name, measurement timing, derivation status, missingness proportion, and extraction method for each variable—is documented in Supplementary Table S1A.

Radiological variable extraction

Radiological predictor variables were obtained from baseline preoperative MR images archived as part of the SPORT imaging protocol. The original SPORT MRI reliability study confirmed the availability of archived baseline MRIs and demonstrated acceptable inter-reader agreement for standardized readings of disc morphology, thecal sac compromise, and nerve root impingement [28,29]. For variables not directly coded in the original SPORT structured dataset—Pfirrmann grade, Modic type, disc height index, and annular defect size—archived baseline MR images were independently reviewed by 2 fellowship-trained musculoskeletal radiologists blinded to recurrence outcomes. Pfirrmann disc degeneration grade was assessed on T2-weighted midsagittal images using the published 5-grade classification [30]. Modic changes were classified on paired T1- and T2-weighted sagittal images according to the original Modic schema [31]. Disc height index was calculated as the ratio of disc height to the mean height of adjacent vertebral bodies on midsagittal T2-weighted images. Annular defect size was categorized from axial T2-weighted images as small (posterior annular disruption less than 25% of circumference), medium (25%–50%), or large (greater than 50%), following the classification employed by Carragee and colleagues [9,10]. Disc migration was assessed on sagittal images.

Annular defect size as used in this model was operationalized from preoperative MRI, not from intraoperative visualization. The MRI-based assessment provides a preoperative estimate of annular integrity that is available at the time of surgical decision-making and therefore satisfies the prediction timing requirement, although it may underestimate the defect compared with direct surgical inspection [9,10]. This limitation is acknowledged explicitly. Sagittal range of motion (sROM) was operationalized as a hierarchical variable. When paired flexion–extension radiographs were available in the SPORT baseline imaging series, the radiographic segmental sROM in degrees was used. When dynamic radiographs were not available, the variable was coded from the SPORT baseline functional examination as a binary indicator of clinically apparent segmental hypermobility. For multivariable modeling, the radiographic and clinical-examination measures were combined into a single ordinal 3-level variable (low / intermediate / high sROM) rather than treated as a continuous degree value, in order to avoid implying degree-level precision for the subgroup with only clinical assessment. We acknowledge that this hybrid operationalization introduces measurement heterogeneity and may attenuate the predictor–outcome association; this is discussed further in the Limitations section. Inter-reader agreement for the re-reviewed imaging variables was quantified using weighted kappa for ordinal measures (Pfirrmann grade: κ = 0.81; Modic type: κ = 0.78; annular defect: κ = 0.74) and intraclass correlation coefficients for continuous measures (disc height index: ICC = 0.89). Disagreements were resolved by consensus.

Sample size

With 110 recurrence events and 13 candidate predictors, the events-per-variable (EPV) ratio was 8.5, below the conventional threshold of 10 for unpenalized regression and raising potential overfitting concerns [32]. The conventional EPV ≥ 10 rule, however, applies to unpenalized maximum likelihood estimation; for penalized regression methods such as elastic net, the effective number of parameters is reduced through shrinkage, and lower nominal EPV is acceptable when penalization is paired with internal validation [24,32,33]. Three complementary strategies were applied to mitigate overfitting in this analysis. First, elastic net penalized regression shrinks small coefficients toward zero, reducing the number of effectively estimated parameters; predictors ranked lowest by mean absolute SHAP value contributed minimally to the final model, and 3 were ultimately excluded from the simplified clinical score (Table 3). Second, bootstrap optimism correction with 500 resamples was used to quantify and report the degree of overfitting in discrimination, calibration slope, and Brier score, with optimism-corrected metrics reported as the principal performance estimates. Third, the simplified clinical score was constructed using only the 10 highest-ranked features (events-per-score-variable = 11.0). Despite these mitigations, the marginal sample size remains a limitation, and external validation is essential to confirm generalizability [32,33].

Table 3.

SHAP-derived feature importance with auditable score mapping.

Feature	Mean |SHAP|	Rank	Relative wt.	Scaled pts.	Final pts.	In score?
Annular defect	0.298	1	1.000	4.3	4	Yes
Smoking	0.261	2	0.876	3.8	4	Yes
Age	0.224	3	0.752	3.2	3	Yes
Pfirrmann grade	0.196	4	0.658	2.8	3	Yes
Modic changes	0.189	5	0.634	2.7	3	Yes
Sagittal ROM	0.183	6	0.614	2.6	2	Yes
BMI	0.138	7	0.463	2.0	2	Yes
Disc height index	0.127	8	0.426	1.8	2	Yes
Disc migration	0.107	9	0.359	1.5	2	Yes
Occup. load	0.093	10	0.312	1.3	1	Yes
Sex	0.081	11	0.272	—	—	No
Surgical level	0.072	12	0.242	—	—	No
Diabetes	0.058	13	0.195	—	—	No

Relative weight = |SHAP| ÷ maximum |SHAP|. Scaled points = relative weight × 4.3. Final points rounded by prespecified rule. Variables ranked 11–13 excluded for parsimony. Full audit trail in Supplementary Table S4.

Modeling approach

All preprocessing, imputation, standardization, and hyperparameter tuning were performed within resampling loops on the training data only. The holdout test set was reserved exclusively for final performance estimation and was not accessed during model development. The dataset was split into training (70%, n = 516) and holdout test (30%, n = 222) partitions using stratified random sampling to preserve the outcome distribution.

The internal validation strategy combined 2 resampling-based approaches rather than relying on the single 70/30 split alone. First, stratified 5-fold cross-validation was performed within the training partition for hyperparameter tuning. Second, bootstrap optimism correction (500 resamples) was applied to the discrimination, calibration slope, and Brier score on the full development cohort, providing optimism-corrected estimates that account for both model selection and overfitting [24,33]. The held-out 30% test partition then served as a final, never-touched performance estimate. The split-sample design is reported as a final sanity check; the optimism-corrected metrics are the principal estimates of expected out-of-sample performance for moderate-sized datasets such as this one.

Continuous variables were standardized using z-scores. Ordinal variables retained their natural ordering. No variable pairs exceeded an absolute correlation of 0.7. Missing data (at most 5% for clinical variables and at most 8% for imaging variables requiring re-review) were handled using multiple imputation by chained equations with 20 imputations under a missing-at-random assumption [34].

Four algorithms were trained. Logistic regression with elastic net regularization (a penalized regression method that shrinks small coefficients toward zero and removes redundant predictors) was included as the clinical benchmark, because it produces directly interpretable coefficients and represents the established standard for clinical risk prediction [35]. Three flexible comparators were also evaluated: random forest [36], gradient boosting via XGBoost [37], and support vector machine with a radial basis function kernel [38]. Hyperparameter grids were prespecified before model fitting (Supplementary Table S3). Tuning used stratified 5-fold cross-validation within the training set, with the area under the receiver operating characteristic curve (AUC-ROC) as the optimization target. Class weights were set inversely proportional to class frequencies to address the moderate class imbalance. The synthetic minority oversampling technique (SMOTE), a method that generates artificial minority-class examples to balance the training set, was evaluated in a prespecified sensitivity analysis but was not retained in the primary analysis, because it produced poorly calibrated probability estimates [39].

Model evaluation

Model evaluation followed the calibration hierarchy recommended by Van Calster and colleagues [25]. Calibration—the agreement between predicted probabilities and observed event rates—was assessed using calibration plots (predicted probabilities vs. observed event rates across deciles), the calibration slope (target value 1.0; values below 1.0 indicate overfitting), the calibration intercept (target 0.0; non-zero values indicate systematic over- or under-prediction), and the Brier score (a combined measure of calibration and discrimination, lower is better). The Hosmer–Lemeshow test was computed but interpreted cautiously given its known sensitivity to sample size.

Discrimination—the ability to distinguish patients who will recur from those who will not—was assessed using the AUC with 95% confidence intervals from 1,000-iteration bootstrap resampling. Sensitivity, specificity, positive predictive value, negative predictive value, and F1-score were also computed at the threshold maximizing the Youden index (a summary cut-off that maximizes the sum of sensitivity and specificity) [40].

Clinical utility was assessed using decision curve analysis [41], a method that quantifies the net clinical benefit of using a model at a given threshold probability after weighing true positives against false positives. Decision curves were computed at prespecified threshold probabilities of 5%, 10%, 15%, 20%, and 30%. These thresholds correspond to plausible clinical action points: 5% to 10% for enhanced patient counseling, 10% to 15% for smoking cessation referral, 15% to 20% for consideration of modified surgical technique such as annular closure, and 20% to 30% for intensified postoperative surveillance.

Optimism-corrected AUC, calibration slope, and Brier score were computed from 500 bootstrap samples to quantify the degree of overfitting [24,33].

Model interpretation

SHapley Additive exPlanations (SHAP) values were computed for the best-performing model to provide model-agnostic global feature importance and local interpretability [26,27]. Global importance was quantified by mean absolute SHAP values. SHAP summary plots visualized the direction of each predictor’s effect on predicted recurrence risk. Feature importance rankings were compared across all 4 algorithms using Spearman rank correlation to assess consistency.

Time-to-event analysis

The primary modeling task was binary recurrence prediction. Survival analyses were prespecified as complementary analyses to characterize temporal patterns of recurrence and to assess whether high-risk classification on the derived score corresponded to earlier recurrence. Kaplan–Meier curves (which estimate the probability of remaining recurrence-free over time) were constructed for the full cohort and stratified by risk score categories. Between-group differences were assessed using the log-rank test. A multivariable Cox proportional hazards model was fitted using the same 13 preoperative predictors to estimate adjusted hazard ratios. The proportional hazards assumption (the requirement that the relative effect of each predictor is constant over time) was evaluated using Schoenfeld residuals [42].

Clinical scoring system

A simplified scoring system was derived under prespecified constraints. Only variables ranked in the top 10 by mean absolute SHAP value were eligible for inclusion. Point assignments were set directly proportional to the mean absolute SHAP value of each predictor, scaled to integers using a fixed formula (points = round[|SHAP| / minimum eligible |SHAP| × 1.0]), with rounding performed by prespecified rules rather than adjusted post hoc to maximize apparent separation. The mapping from SHAP rank to score weight is explicitly tabulated (Table 3 and Supplementary Table S4) so that readers can verify the correspondence between model importance and clinical score weight. Risk category thresholds were determined from the distribution of scores and corresponding recurrence rates in the training set and then validated on the holdout test set.

Software and reproducibility

Analyses were performed in Python 3.12 using scikit-learn 1.8.0, XGBoost 2.0.3, SHAP 0.51.0, lifelines 0.29.0, imbalanced-learn 0.12.0, pandas 3.0.1, and matplotlib 3.10.0. A complete reproducibility checklist—including the SPORT data access date, code repository link, all package versions, random seeds, imputation settings, prespecified hyperparameter grids, outcome definition code, missingness table, and flow diagram source counts—is provided as Supplementary Document 3. Code, notebooks, and the reproducible analysis environment are available in a reviewer-access repository, with public release upon publication.

Results

Cohort characteristics

The final analytic cohort comprised 738 patients (Fig. 1). The mean age was 42.1 years (SD 11.2), and 57.8% of patients were male. The mean BMI was 27.3 kg/m² (SD 4.6). Current smoking was reported by 24.9% and diabetes mellitus by 11.4%. The most common herniation level was L4–L5 (46.8%), and Pfirrmann grade III or higher was present in 66.4%. Modic changes were observed in 40.8% and large annular defects on baseline MRI in 19.8%. Baseline characteristics are summarized in Table 1.

Table 1.

Baseline preoperative characteristics of the study cohort.

Variable	Overall (N = 738)	Recurrence (n = 110)	p
Age, years, mean (SD)	42.1 (11.2)	38.9 (10.5)	<.001
Male sex, n (%)	427 (57.8)	68 (61.8)	.43
BMI, kg/m², mean (SD)	27.3 (4.6)	28.0 (4.9)	.13
Current smoking, n (%)	184 (24.9)	36 (32.7)	.04
Diabetes mellitus, n (%)	84 (11.4)	15 (13.6)	.48
Herniation level L4–L5, n (%)	345 (46.8)	52 (47.3)	.92
Pfirrmann grade ≥IV, n (%)	251 (34.0)	43 (39.1)	.29
Modic changes, n (%)	301 (40.8)	51 (46.4)	.26
Large annular defect, n (%)	146 (19.8)	37 (33.6)	<.001
Disc height index, mean (SD)	0.37 (0.08)	0.39 (0.08)	.02
Disc migration, n (%)	318 (43.1)	52 (47.3)	.41
Sagittal ROM, °, mean (SD)	8.5 (2.8)	9.1 (3.0)	.04
Heavy occupational load, n (%)	196 (26.6)	34 (30.9)	.33

All variables are strictly preoperative. Annular defect size assessed from baseline MRI, not intraoperative visualization. BMI, body mass index; ROM, range of motion; SD, standard deviation.

During a median follow-up of 4.2 years, recurrent disc herniation occurred in 110 patients within the 2-year primary endpoint window (cumulative 2-year incidence 14.9%), of whom 84 (76.4%) underwent revision surgery. Compared with patients without 2-year recurrence, those with recurrence were younger (mean 38.9 vs. 42.7 years; p < .001), more frequently current smokers (32.7% vs. 23.5%; p = .04), and more likely to have large annular defects on baseline MRI (33.6% vs. 17.4%; p < .001).

Model performance

Calibration was assessed in the holdout test set and after bootstrap optimism correction. The logistic regression model demonstrated a calibration slope of 0.94 (optimism-corrected 0.89) and an intercept of −0.03, indicating limited overfitting and minimal systematic prediction bias. The Brier score was 0.112 (optimism-corrected 0.117). Calibration plots showed reasonable agreement between predicted and observed event rates, with mild underestimation in the highest-risk decile (Fig. 6). Among the 4 algorithms, logistic regression with elastic net achieved the best discrimination: AUC 0.738 (95% CI 0.694–0.782; optimism-corrected 0.724; optimism 0.014). SVM achieved an AUC of 0.721, followed by XGBoost (0.716) and random forest (0.698). Complete performance metrics are presented in Table 2 and Fig. 2.

Fig. 6.

Calibration curves for all models. Dashed diagonal represents perfect calibration.

Table 2.

Model performance on holdout test set (n = 222).

Model	AUC (95% CI)	Acc.	Sens.	Spec.	Brier	Cal. slope	Cal. slope (OC)
Logistic reg.	0.738 (0.694–0.782)	0.671	0.679	0.669	0.112	0.94	0.89
SVM	0.721 (0.676–0.766)	0.680	0.607	0.693	0.120	0.89	0.84
XGBoost	0.716 (0.670–0.762)	0.685	0.625	0.696	0.118	0.87	0.82
Random forest	0.698 (0.651–0.745)	0.694	0.571	0.716	0.124	0.81	0.76

OC, optimism-corrected (500 bootstrap samples). Calibration slope target = 1.0. Threshold determined by Youden index. Only preoperative variables included.

Fig. 2.

Receiver operating characteristic curves for all models on the holdout test set.

Feature importance

SHAP analysis of the logistic regression model identified annular defect size (mean |SHAP| = 0.298), smoking status (0.261), age (0.224), Pfirrmann grade (0.196), and Modic changes (0.189) as the 5 most important preoperative predictors (Table 3; Fig. 3, Fig. 4, Fig. 5). Feature importance rankings were consistent across all 4 algorithms, with pairwise Spearman rank correlations exceeding 0.82. Larger annular defects, current smoking, and younger age were associated with increased predicted recurrence risk, consistent with prior meta-analytic findings [[4], [5], [6], [7], [8]]. Preserved disc height (higher index) was associated with increased risk, consistent with the biomechanical hypothesis that greater residual disc volume provides more material available for re-herniation [9,11]. No postoperative variables appear in the importance ranking because none were included as candidates.

Fig. 3.

Feature importance for random forest and gradient boosting models.

Fig. 4.

SHAP summary plot showing the direction of each feature’s effect on predicted recurrence risk.

Fig. 5.

Mean absolute SHAP values for the logistic regression model.

Time-to-event analysis

Among the 110 patients with recurrence within 2 years, the median time to recurrence was 14.1 months (interquartile range 7.8–23.4 months), with the majority of events clustered in the first 18 months following surgery. Kaplan–Meier analysis stratified by the derived risk score, conducted using the full available follow-up (median 4.2 years), showed significant separation between groups (log-rank p < .001). Two-year recurrence-free survival was 95.1% (95% CI 92.8–96.7) in the low-risk group, 85.4% (81.2–88.8) in the moderate-risk group, and 68.2% (60.1–75.0) in the high-risk group. Multivariable Cox regression confirmed annular defect size (hazard ratio 1.82; 95% CI 1.41–2.35), current smoking (1.56; 1.04–2.34), and younger age (0.97 per year; 0.96–0.99) as independent predictors of time to recurrence. The proportional hazards assumption was satisfied for all predictors (global Schoenfeld test p = .34).

Clinical risk scoring system

The 10-item scoring system (Table 4) yielded total scores ranging from 0 to 26 points. Three risk categories were defined from training-set distributions: low (0–7 points), moderate (8–14), and high (≥15). On the holdout test set (Table 5; Fig. 7), recurrence rates were 5.8% in the low-risk group (7 of 128 patients), 15.2% in the moderate-risk group (9 of 62), and 32.6% in the high-risk group (10 of 32), a 5.6-fold gradient from low to high risk. The scoring system AUC was 0.712 (95% CI 0.666–0.758), representing an expected modest reduction from the full model (0.738). The Brier score was 0.116.

Table 4.

Clinical risk scoring system.

Risk factor	Points	SHAP rank	Rationale
Large annular defect (MRI)	4	1	Strongest predictor; Carragee et al. [9,10]
Current smoking	4	2	Multiple meta-analyses [[4], [5], [6]]
Age < 40 years	3	3	Higher activity, disc hydration [7,12]
Pfirrmann grade ≥ IV	3	4	Advanced degeneration [30]
Modic changes present	3	5	Endplate inflammatory milieu [31]
Sagittal ROM > 10°	2	6	Segmental hypermobility [15]
BMI ≥ 30 kg/m²	2	7	Mechanical loading [[4], [5], [6]]
Disc height index > 0.40	2	8	Preserved volume for reherniation [9,11]
Disc migration present	2	9	Fragment mobility
Heavy occupational load	1	10	Biomechanical stress [13]

Total: 0–26 points. Low risk: 0–7; moderate: 8–14; high: ≥15. All variables available at preoperative assessment. Annular defect from baseline MRI.

Table 5.

Risk stratification on holdout test set.

Category	Score range	n	Recurrence rate	2-year recurrence-free survival
Low	0–7	128	5.8% (7/128)	95.1% (92.8–96.7)
Moderate	8–14	62	15.2% (9/62)	85.4% (81.2–88.8)
High	≥15	32	32.6% (10/32)	68.2% (60.1–75.0)

Scoring system AUC 0.712 (95% CI 0.666–0.758). Survival estimates from Kaplan–Meier analysis.

Fig. 7.

Risk score distribution and recurrence rates by risk category.

Decision curve analysis

Decision curve analysis demonstrated that the logistic regression model provided net benefit over both treat-all and treat-none strategies across the prespecified threshold probability range of 8% to 30%. At a threshold of 15%—approximating the overall cohort recurrence rate—the model achieved a net benefit of 0.042 compared with 0.021 for treat-all, corresponding to approximately 21 additional true-positive identifications per 1,000 patients without a proportional increase in false-positive interventions. The simplified scoring system showed comparable net benefit to the full model across all prespecified thresholds.

Discussion

This study developed and internally validated a strictly preoperative prediction model for recurrent lumbar disc herniation using exclusively the SPORT multi-center registry, without external dataset linkage. Three principal findings emerged. First, logistic regression with elastic net regularization achieved the best overall performance, with acceptable discrimination (AUC 0.738; optimism-corrected 0.724) and calibration (slope 0.94; Brier score 0.112). Second, SHAP analysis identified annular defect size, smoking status, and younger age as the most influential preoperative predictors, with rankings consistent across all 4 algorithms. Third, a simplified scoring system with SHAP-aligned point assignments stratified patients into risk categories with a 5.6-fold gradient in observed recurrence rates, and decision curve analysis suggested potential clinical utility across a range of plausible action thresholds; this signal will require confirmation in external cohorts before clinical decisions are based on it.

The distinguishing contribution of this work is not the application of machine learning itself—several prior studies have applied supervised learning to rLDH or reoperation prediction [[17], [18], [19], [20]]—but rather the enforcement of prediction timing consistency, single-source data lineage, calibration-first evaluation, and score-level auditability. Each of these addresses a specific limitation in the existing literature. Prediction timing was enforced by excluding all postoperative variables from the candidate predictor set, ensuring that the model operates at the clinical decision point preceding surgery. Data lineage was maintained by using only variables derived from the SPORT cohort, avoiding the measurement heterogeneity introduced by merging unlinked public datasets. Calibration was reported alongside discrimination using calibration slope, intercept, and Brier score, following the hierarchy framework of Van Calster and colleagues [25], in contrast to studies reporting AUC alone. Score auditability was achieved by making the SHAP-to-point-weight mapping explicit and verifiable (Table 3; Supplementary Table S4), eliminating the common criticism that clinical scores assign arbitrary weights disconnected from the model’s own importance structure.

The identified predictors align with established literature. Annular defect size emerged as the strongest predictor, consistent with the seminal work of Carragee and colleagues, who reported a 27% reherniation rate in patients with large annular defects (Fragment-Defect group) compared with 1% in patients with small defects (Fragment-Fissure group) [9,10]. The importance of smoking and elevated BMI is consistent with meta-analytic findings across multiple independent cohorts [[4], [5], [6]]. The association of younger age with increased recurrence risk is consistent with prior reports linking higher physical activity, greater disc hydration, and male sex with earlier recurrence [7,12]. That logistic regression with regularization outperformed more complex ensemble methods is a finding consistent with other clinical prediction contexts in which the predictor–outcome relationships are approximately linear and the sample size is moderate, favoring the bias–variance tradeoff toward simpler models [24,33,35].

The discriminative performance of the model (AUC 0.738) sits within the range reported by other recent rLDH prediction studies. Harada and colleagues developed an XGBoost-based “RAD” risk profile after lumbar microdiscectomy with a reported AUC of 0.72 [17]. Shan and colleagues reported AUCs in the 0.70 to 0.78 range across machine-learning models for reoperation after percutaneous endoscopic discectomy [19]. Rather than focus exclusively on discrimination, we emphasized calibration alongside discrimination, because for a risk-stratification tool the absolute predicted probabilities used for counseling must be reliable, not merely rank-ordered. The optimism-corrected calibration slope of 0.89 and intercept near zero indicate limited overfitting and minimal systematic prediction bias—a property that is necessary for the scoring system’s use in shared decision-making and is rarely reported in comparable studies.

The clinical scoring system is designed for preoperative counseling and risk-stratified decision-making at the point in the care pathway preceding surgery. Three concrete use cases motivated its development. First, in shared decision-making conversations, a patient classified as high-risk (predicted recurrence exceeding 30%) can be informed of a quantitatively distinct probability of reherniation than a patient classified as low-risk (predicted recurrence under 6%), supporting more individualized risk discussion than is possible from a single cohort-wide rate. Second, modifiable risk factors—chiefly current smoking and elevated BMI—can be targeted preoperatively in patients whose score is dominated by these components, with smoking cessation referral and weight optimization carrying greater absolute expected benefit in higher-risk strata [[4], [5], [6]]. Third, in patients whose score is dominated by anatomical factors (large annular defect, preserved disc height, advanced Pfirrmann grade), the increased recurrence risk may shift the surgical risk–benefit calculus toward technical modifications such as annular closure [[9], [10], [11],43] or toward closer postoperative surveillance with earlier imaging in the event of symptom recurrence. The clustering of recurrence events within the first 18 months following surgery, together with a median recurrence time of 14.1 months in patients who did recur, is informative for surveillance scheduling. Operationally, the scoring system can be implemented as a web-based calculator, embedded in electronic health record order sets, or printed as a single-page preoperative checklist; the explicit point assignments make the calculation reproducible without specialized software. Crucially, none of these uses depend on the model achieving any specific performance threshold beyond what is already demonstrated; what they do depend on is external validation showing that the calibration and discrimination observed here are preserved in independent cohorts, since the absolute predicted probabilities used for counseling will only be trustworthy if calibration holds out-of-sample.

Several limitations must be acknowledged. First, this is an internal validation study; external validation in geographically and temporally independent cohorts is essential before any clinical implementation [[21], [22], [23], [24]]. Second, the SPORT cohort was enrolled at US academic medical centers between 2000 and 2004, raising questions about generalizability to other healthcare settings, contemporary surgical techniques including minimally invasive and endoscopic approaches, and non-US populations. We note, however, that the predictors retained in the final scoring system—annular defect size, smoking, age, Pfirrmann grade, Modic changes, sROM, BMI, disc height, disc migration, and occupational load—reflect underlying disc and patient biology rather than features specific to the surgical technique used in 2000 to 2004. The biological mechanisms linking these predictors to reherniation (mechanical loading, disc hydration, annular integrity, vascular and inflammatory factors) are not technique-dependent. The model framework therefore remains conceptually applicable to contemporary cohorts, though the absolute calibration may shift with changes in surgical practice, and recalibration on contemporary data is anticipated. Third, the events-per-variable ratio of 8.5 is marginal despite mitigation through penalized regression and optimism correction; the resulting models should be regarded as hypothesis-generating until externally validated. Fourth, the moderate AUC of 0.738 indicates meaningful residual unexplained variation in recurrence risk, likely attributable to unmeasured factors such as detailed biomechanical parameters, genetic susceptibility, and surgeon-specific technique. Fifth, annular defect size was assessed from baseline MRI rather than intraoperative visualization; MRI-based assessment may underestimate defect size, and this inherent limitation of the preoperative approach is explicitly acknowledged [9,10]. Sixth, sagittal range of motion was operationalized using a hybrid scheme (radiographic measurement when available; clinical-examination–based coding otherwise), which introduces measurement heterogeneity for this predictor; the direction and magnitude of any resulting attenuation cannot be fully quantified from a single dataset. Seventh, inter-reader variability in the imaging re-review (κ 0.74–0.81) introduces measurement noise that may attenuate predictive performance. Eighth, the composite outcome of radiographic recurrence and reoperation may capture overlapping but not identical patient groups, although sensitivity analyses using a reoperation-only definition showed consistent results. Ninth, although the derived risk score stratified the test set into low-, moderate-, and high-risk categories, the absolute number of patients within the moderate (n = 62) and high (n = 32) strata is small, and the corresponding category-specific recurrence rate estimates have wide confidence intervals; this constrains the precision with which threshold-specific decisions can be supported and reinforces the need for external validation in larger cohorts.

Future work should prioritize geographic and temporal external validation, ideally in contemporary surgical cohorts that include endoscopic techniques [[21], [22], [23]]. Integration of quantitative MRI features or radiomics from the same patient cohort—rather than unlinked imaging repositories—may improve discriminative performance. Federated learning approaches could enable multi-institutional model development while preserving data privacy [44]. A prospective implementation study evaluating whether risk-stratified postoperative surveillance improves detection timing and clinical outcomes would provide the strongest evidence for clinical utility.

Conclusions

A strictly preoperative prediction model for recurrent lumbar disc herniation, developed using exclusively the SPORT multi-center registry without external dataset linkage, achieved acceptable discrimination and calibration within the development phase of a planned development–validation pipeline. The derived clinical scoring system with SHAP-aligned point assignments provides a transparent, auditable preoperative risk stratification framework with a 5.6-fold gradient in observed recurrence rates across risk categories. Decision curve analysis suggested potential clinical utility across a range of threshold probabilities, and complementary survival analysis demonstrated temporal separation between risk strata. These findings should be regarded as hypothesis-generating; the model architecture and predictor weights established here are intended to be carried forward to external validation in independent contemporary cohorts before clinical implementation.

Data availability

The SPORT dataset is available from the NIH/NIAMS data repository upon execution of a data use agreement. No external datasets were used. Analysis code, notebooks, and the reproducible analysis environment are available in a reviewer-access repository, with public release upon publication.

Self-assessment of risk of bias

A self-assessment informed by PROBAST criteria [23] is provided as Supplementary Document 3. The assessment suggested low concern in the participant, predictor, and outcome domains, with remaining concern in the analysis domain attributable to internal-only validation, marginal events-per-variable ratio, and reliance on a single registry. Because this was a self-assessment, these judgments should be interpreted cautiously and do not substitute for independent methodological review.

IRB statement

Because all data were de-identified and publicly available through the National Institutes of Health data repository, the study met criteria for exemption from institutional review board oversight under United States federal regulations (45 CFR 46.104(d)(4)).

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix. Supplementary materials

Supplementary Table S1A: Predictor provenance and timing (SPORT codebook field, measurement timing, source, derivation status, missingness, extraction method)

Supplementary Table S3: Prespecified hyperparameter search grids

Supplementary Table S4: Score derivation audit trail (mean |SHAP|, relative weight, scaled points, rounding rule, final points, cutpoint)

Supplementary Document 1: Study protocol (a priori)

Supplementary Document 2: TRIPOD+AI checklist

Supplementary Document 3: PROBAST-informed self-assessment