Development and internal validation of a machine learning-based clinical decision support tool for surgical intervention in degenerative lumbar spondylolisthesis: a retrospective MRI study

Surgical decision-making for degenerative lumbar spondylolisthesis (DLS) is highly subjective and lacks quantitative, objective criteria. We sought to develop and validate a machine learning (ML) based clinical decision support tool using magnetic resonance imaging (MRI) features to identify appropriate candidates for surgical intervention. We established a retrospective cohort of patients with single-level DLS. Patients were categorized into a surgical candidate group (underwent surgery with satisfactory outcomes) and a nonoperative group (stable or improved with conservative management for at least 6 months). Patients with poor surgical outcomes were excluded to ensure accurate ground truth labeling. A sensitivity analysis including these excluded patients was additionally performed to assess the robustness of the classification framework. Extracted features included clinical demographics and MRI parameters such as Pfirrmann grade, multifidus fatty infiltration, facet joint orientation, and degree of slippage. Four ML algorithms (Logistic Regression, LASSO, Random Forest, and XGBoost) were developed and evaluated using 5-fold cross-validation. Model performance was assessed via the area under the receiver operating characteristic curve (AUC), sensitivity, and specificity. SHapley Additive exPlanations (SHAP) analysis was employed to interpret the best-performing model. Decision curve analysis suggested net clinical benefit across a range of threshold probabilities. The XGBoost algorithm yielded the most accurate predictions, achieving an AUC of 0.86 (95% CI: 0.82–0.90) in the 5-fold cross-validation cohorts. SHAP analysis suggested that the degree of slippage, multifidus fatty infiltration, and Pfirrmann disc degeneration grade were the most influential MRI features associated with surgical necessity in this cohort. Decision curve analysis indicated that the XGBoost model provided a higher net clinical benefit across a range of threshold probabilities compared to default strategies. Sensitivity analysis including the 23 patients with poor surgical outcomes demonstrated a modest reduction in AUC (0.82, 95% CI: 0.77–0.87), confirming the overall robustness of the model. In this single-center retrospective preliminary internal validation study, a machine learning model integrating routine MRI features and clinical parameters demonstrated good performance in distinguishing surgical candidates from patients amenable to conservative management. These findings are hypothesis-generating and require prospective, multi-center external validation before any consideration of clinical deployment.

Degenerative lumbar spondylolisthesis (DLS) is a prevalent condition characterized by the forward translation of one vertebra over the adjacent caudal vertebra with an intact neural arch. It is a leading cause of low back pain, neurogenic claudication, and radiculopathy, particularly among the aging population (1). While conservative management, including physical therapy and epidural steroid injections, remains the first-line treatment, approximately 20%–30% of patients experience persistent or progressive symptoms that ultimately necessitate surgical intervention (2).

The decision to proceed with surgery in DLS is currently complex and highly subjective. Traditional indications rely heavily on the failure of conservative treatment and the severity of patient-reported symptoms (3). However, the lack of standardized, objective criteria often leads to significant variability in surgical decision-making among clinicians. Magnetic resonance imaging (MRI) serves as the gold standard for evaluating DLS, providing comprehensive anatomical details regarding intervertebral disc degeneration, facet joint arthropathy, and paraspinal muscle morphology (4). Despite the wealth of information available in MRI, clinicians typically focus on isolated parameters, such as the degree of slippage or central canal stenosis, often underutilizing the multidimensional interactions of these degenerative features in predicting the failure of conservative care.

Machine learning (ML) offers analytical advantages over traditional statistical approaches in orthopedic and spine surgery. Unlike traditional statistical methods, ML algorithms excel at capturing complex, nonlinear relationships within high-dimensional datasets (5). A recent narrative review highlighted the growing role of ML in spine surgery, from preoperative risk stratification to clinical decision support (6). Previous studies have successfully applied ML to predict postoperative complications, hospital readmissions, and patient-reported outcomes following spine surgery (7, 8). In the specific context of lumbar spondylolisthesis, bibliometric analyses have identified surgical management and minimally invasive techniques as dominant research hotspots, yet quantitative decision support tools remain underexplored (9). However, there is a notable scarcity of ML models specifically designed to act as clinical decision support tools for identifying surgical candidates in DLS prior to intervention. A major barrier to clinical implementation of ML is its “black-box” nature; therefore, model interpretability is essential for clinical trust (10).

To address this gap, we developed and internally validated an interpretable ML-based clinical decision support tool. By integrating routine MRI features with baseline clinical data, we aimed to explore whether machine learning algorithms could assist in distinguishing patients who would benefit from surgical intervention from those amenable to conservative management. We further incorporated the SHAP (SHapley Additive exPlanations) framework—utilizing the tree-based SHAP explainer specifically designed for gradient-boosted models—to provide a degree of transparency in the model's decision-making process, with the long-term goal of facilitating more objective and reproducible clinical discussions regarding surgical candidacy.

Study design and patient population

This retrospective cohort study was approved by the Institutional Review Board of our institution, and the requirement for informed consent was waived due to the retrospective nature of the study. We reviewed the electronic medical records and imaging database for 546 patients diagnosed with DLS between January 2018 and December 2022. The inclusion criteria were: (1) diagnosis of single-level DLS at L4-L5 or L5-S1; (2) availability of complete baseline clinical data and preoperative lumbar MRI within 3 months of treatment decision; and (3) a minimum follow-up of 12 months with clear documentation of treatment outcomes. Patients were excluded if they had a history of previous lumbar spine surgery, isthmic spondylolisthesis, concomitant spinal tumors, active spinal infection, or severe spinal deformity (e.g., degenerative scoliosis > 20 degrees). Patients with DLS at other levels (e.g., L3-L4) were excluded to maintain a homogeneous cohort, as the biomechanics and natural history of spondylolisthesis differ by level. Global sagittal alignment parameters (pelvic incidence, pelvic tilt) were not routinely available in our MRI-based dataset, as these measurements require standing full-spine radiographs; this is acknowledged as a limitation.

Definition of the outcome

The primary outcome was the binary classification of surgical necessity. To establish a reliable ground truth reflecting actual clinical need rather than subjective physician preference, patients were categorized into two groups based on their treatment trajectory and outcomes. The “Surgical Candidate Group” was defined as patients who underwent surgical decompression and/or fusion and subsequently achieved a satisfactory clinical outcome (minimum 30% improvement in the Oswestry Disability Index at 1-year follow-up) (11). The 30% threshold was selected based on the established minimal clinically important difference (MCID) for the ODI in lumbar degenerative conditions, which has been reported to range from 12.8 to 15 points or approximately 25%–33% relative improvement in the literature (12). Conversely, the “Nonoperative Group” consisted of patients successfully managed with conservative treatment for at least 6 months, demonstrating stable or improved symptoms without crossover to surgery. “Stable or improved” was operationally defined as no worsening of ODI score by more than 10% from baseline and no escalation of treatment intensity (e.g., no new referral for surgical consultation). Notably, patients who underwent surgery but experienced poor outcomes (defined as <30% ODI improvement at 1-year follow-up; n = 23) were excluded from the primary analysis; this step was crucial to ensure the model learned the characteristics of appropriate surgical candidates rather than simply mimicking past subjective decisions. We acknowledge that this exclusion may introduce survivor bias by removing patients with complex pathology or surgical complications. To address this concern, a sensitivity analysis was performed with these 23 patients re-included in the surgical group (see Section 3.5).

Data collection and feature extraction

Baseline clinical characteristics were extracted from the medical records, including age, sex, body mass index (BMI), duration of symptoms, and relevant comorbidities (e.g., diabetes mellitus, hypertension).

MRI features were evaluated independently by two experienced musculoskeletal radiologists blinded to the patients' clinical outcomes. In cases of discrepancy, a consensus was reached through discussion with a senior spine surgeon. Inter-observer reliability was assessed using Cohen's kappa for categorical variables (Pfirrmann grade, Pathria classification, Goutallier fatty infiltration grade) and intraclass correlation coefficients (ICC, two-way random, absolute agreement) for continuous measurements (slippage percentage, facet joint angle). The inter-observer agreement was substantial to excellent: Cohen's kappa was 0.78 (95% CI: 0.71–0.85) for Pfirrmann grading, 0.74 (95% CI: 0.66–0.82) for Pathria classification, and 0.72 (95% CI: 0.64–0.80) for Goutallier grading. The ICC for slippage percentage was 0.91 (95% CI: 0.88–0.94) and for facet joint angle was 0.87 (95% CI: 0.83–0.91). The extracted MRI parameters included the following:

Degree of Slippage: Measured on sagittal T2-weighted images and expressed as a percentage of the inferior vertebral body width.

Intervertebral Disc Degeneration: Assessed using the Pfirrmann grading system (Grade I–V) at the listhetic level (13).

Facet Joint Parameters: Facet joint orientation (angle relative to the sagittal plane) and facet tropism (asymmetry between left and right facet angles) were measured on axial T2-weighted images. The facet joint angle was defined as the angle between a line drawn along the facet joint interface and the sagittal midline of the vertebral body at the corresponding level. Facet tropism was defined as an absolute difference of ≥10° between bilateral facet angles. Facet joint osteoarthritis was graded using the Pathria classification (14).

Paraspinal Muscle Morphology: The cross-sectional area (CSA) and fatty infiltration (FI) of the multifidus and erector spinae muscles were evaluated at the corresponding disc level using the Goutallier classification system adapted for MRI (15). The Goutallier grading was performed on axial T2-weighted images at the level of the listhetic segment (L4-L5 or L5-S1, as applicable).

Modic Changes: The presence and type of endplate signal changes (Type 1, 2, or 3) were recorded.

Machine learning model development

Data preprocessing was performed prior to model training. Missing clinical variables (accounting for <5% of the dataset) were imputed using the k-nearest neighbors (KNN) algorithm. Continuous variables were standardized using z-score normalization. Given the moderate class imbalance between the surgical and nonoperative groups (43.3% vs. 56.7%), the Synthetic Minority Over-sampling Technique (SMOTE) was applied exclusively within each training fold to prevent data leakage into the validation folds (16). We acknowledge that this degree of class imbalance (43.3% vs. 56.7%) is relatively mild and may not strictly necessitate resampling. However, we elected to apply SMOTE for the following reasons: (1) the minority class (surgical candidates) is the clinically actionable class for which higher sensitivity is desired, and even mild imbalance can bias tree-based models toward the majority class during split optimization; (2) SMOTE was applied strictly within each training fold, thereby avoiding data leakage; and (3) to directly address this concern, we performed a supplementary analysis without SMOTE (see Section 3.5), which demonstrated comparable model performance (XGBoost AUC = 0.84 without SMOTE vs. 0.86 with SMOTE), confirming that SMOTE did not artificially inflate discriminative ability.

Following data preprocessing, an initial feature set of 15 candidate variables was assembled from clinical and MRI parameters. To reduce multicollinearity and retain only informative predictors, we performed a two-step feature selection process: (1) variables with a variance inflation factor (VIF) > 5 were removed (eliminating sex, comorbidities, erector spinae FI, and facet OA grade due to high correlation with other features), and (2) the remaining features were further refined using LASSO regularization with 10-fold cross-validation to identify the optimal lambda (λ = 0.023), yielding a final set of 10 features that entered the modeling pipeline: age, BMI, symptom duration, degree of slippage, Pfirrmann grade, facet joint angle (left), facet joint angle (right), facet tropism, multifidus CSA, and multifidus fatty infiltration (Goutallier grade).

Four distinct ML algorithms were selected for model development: Logistic Regression (LR) serving as the baseline linear model, Least Absolute Shrinkage and Selection Operator (LASSO) for feature selection and regularization, Random Forest (RF), and eXtreme Gradient Boosting (XGBoost). The models were trained to predict the binary outcome of surgical necessity. We employed a 5-fold cross-validation strategy to ensure robust evaluation and prevent overfitting. Hyperparameter tuning for RF and XGBoost was conducted using grid search within the cross-validation loops. The final hyperparameters selected for the XGBoost model were: learning rate = 0.05, max depth = 4, n_estimators = 200, subsample = 0.8, colsample_bytree = 0.8, and min_child_weight = 3. For Random Forest: n_estimators = 300, max_depth = 6, min_samples_split = 5, and min_samples_leaf = 3.

Model evaluation and interpretability

The predictive performance of each model was evaluated across the cross-validation folds using standard metrics, including the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, and F1-score. Additionally, the precision-recall AUC (PR-AUC) was calculated to provide a complementary evaluation metric that is less sensitive to class distribution. The model demonstrating the highest average AUC was selected as the final optimal model.

To assess the clinical utility of the optimal model, Decision Curve Analysis (DCA) was performed. DCA quantifies the net benefit of using the prediction model across a range of threshold probabilities, comparing it to default strategies of “treat all” or “treat none” (17). In this context, the “treat all” strategy refers to recommending surgical intervention for all patients (i.e., assuming all patients are surgical candidates), while “treat none” refers to managing all patients conservatively. The net benefit was calculated as: Net Benefit = (True Positives/ n)—(False Positives/ n) × [pt/(1 − pt)], where pt is the threshold probability.

To overcome the “black-box” limitation of advanced ML algorithms, we applied the SHapley Additive exPlanations (SHAP) method to the optimal model. Specifically, we employed the TreeSHAP algorithm (tree-based explainer), which is computationally efficient and exact for tree ensemble models such as XGBoost (18). SHAP values, based on cooperative game theory, provide a unified measure of feature importance. We generated SHAP summary plots to visualize the global contribution and directionality of each feature. Mean absolute SHAP values were calculated for each feature to provide a quantitative ranking of feature importance. Furthermore, SHAP force plots were constructed for individual patients to illustrate how specific MRI and clinical features drove the model's prediction, simulating a transparent clinical decision-making process.

Statistical analysis

Statistical analyses and ML model development were performed using Python (version 3.9) with the scikit-learn and xgboost libraries. Continuous variables were presented as means with standard deviations (SD) and compared using independent t -tests or Mann–Whitney U -tests. Categorical variables were presented as frequencies and percentages, compared using Chi-square tests. The reporting of this study adhered to the Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD + AI) guidelines (19, 20). A completed TRIPOD + AI checklist is provided as Supplementary Material. A two-sided p -value < 0.05 was considered statistically significant.

Baseline characteristics

The patient selection process is detailed in the study flowchart (Figure 1). Of 546 patients initially screened, 181 were excluded based on predefined criteria (multi-level DLS, n = 52; prior lumbar surgery, n = 41; isthmic type, n = 35; incomplete MRI, n = 28; follow-up < 12 months, n = 25). Among the remaining 365 patients, 23 who underwent surgery but failed to achieve satisfactory outcomes (<30% ODI improvement at 1-year follow-up) were further excluded from the primary classification framework to ensure accurate ground truth labeling. A final cohort of 342 patients with single-level DLS was available for model development.

Based on the predefined outcome criteria, 148 patients (43.3%) were assigned to the Surgical Candidate Group, while 194 patients (56.7%) were assigned to the Nonoperative Group. The mean age of the cohort was 65.4 ± 8.2 years, and 62.6% were female. Baseline demographic and clinical characteristics were largely comparable between the two groups, although the surgical group had a slightly higher mean BMI (28.1 vs. 26.5 kg/m 2, p = 0.03).

Univariate analysis of the MRI features revealed significant differences between the cohorts. Patients in the surgical group exhibited a significantly higher mean degree of slippage (18.5% vs. 12.2%, p < 0.001) and a higher prevalence of advanced Pfirrmann disc degeneration (Grade IV/V: 78.4% vs. 51.5%, p < 0.001). Furthermore, severe multifidus fatty infiltration (Goutallier Grade 3/4) was markedly more common in the surgical candidates (64.2% vs. 35.1%, p < 0.001).

Performance of machine learning models

The predictive performance of the four ML models, evaluated via 5-fold cross-validation, is summarized in Table 1 and visualized in Figure 2. The XGBoost model demonstrated the highest discriminative ability, achieving an average AUC of 0.86 (95% CI: 0.82–0.90). This was followed by Random Forest (AUC = 0.83), LASSO (AUC = 0.78), and Logistic Regression (AUC = 0.75). The XGBoost model also yielded the highest sensitivity (0.81) and specificity (0.79), suggesting a potentially useful capability in correctly identifying both surgical candidates and patients who might remain stable with continued conservative care. Notably, the 95% confidence intervals for the XGBoost AUC (0.82–0.90) did not overlap with those of Logistic Regression (0.71–0.79), suggesting a statistically significant improvement in discriminative performance. The precision-recall AUC (PR-AUC) for the XGBoost model was 0.83, further supporting its discriminative performance in the context of the moderate class distribution.

Model interpretability using SHAP

To provide clinical interpretability, SHAP analysis was performed on the optimal XGBoost model using the TreeSHAP algorithm. The SHAP summary plot (Figure 3) illustrates the global contribution and directionality of each feature. The color scale represents feature values (red = high feature value, blue = low feature value). The degree of slippage emerged as the most influential predictor, with higher values of slippage consistently driving the model toward a surgical prediction. Multifidus fatty infiltration (Goutallier grade) was the second most important feature, with higher grades of fatty infiltration strongly associated with surgical candidacy. Pfirrmann disc degeneration grade ranked third, with advanced degeneration (Grades IV–V) contributing positively to the surgical prediction.

The mean absolute SHAP values (representing the average magnitude of each feature's contribution to the model output across all patients) were as follows: degree of slippage = 0.42, multifidus fatty infiltration = 0.35, Pfirrmann grade = 0.28, facet joint angle (left) = 0.18, symptom duration = 0.15, BMI = 0.12, facet tropism = 0.10, multifidus CSA = 0.09, age = 0.07, and facet joint angle (right) = 0.05.

The SHAP waterfall plots (Figure 4) provide individual-level interpretability. Figure 4A illustrates a true-positive surgical candidate: the waterfall plot visually depicted how the presence of Grade 4 multifidus fatty infiltration and a 22% slippage contributed to surpassing the decision threshold, despite the patient having a relatively low BMI which negatively contributed to the surgical prediction. Conversely, Figure 4B illustrates a patient successfully managed conservatively, where lower slippage and minimal muscle degeneration drove the prediction below the threshold.

Clinical utility

Decision Curve Analysis (DCA) was utilized to evaluate the potential clinical consequences of deploying the XGBoost model (Figure 5). The DCA curve revealed that the XGBoost model provided a higher net clinical benefit compared to the default “treat all” (i.e., recommend surgery for all patients) or “treat none” (i.e., manage all patients conservatively) strategies across a broad range of threshold probabilities (from approximately 15%–75%). Notably, within the clinically most relevant threshold range of 30%–60%—which corresponds to the typical decision uncertainty zone in DLS management—the XGBoost model consistently demonstrated positive net benefit. These findings indicate a potential role for the model in refining surgical indications while potentially reducing unnecessary interventions in patients who might succeed with conservative care.

Sensitivity analyses

To address the potential survivor bias introduced by excluding the 23 patients with poor surgical outcomes, we performed a sensitivity analysis by re-including these patients in the surgical group (yielding a total surgical cohort of n = 171). The XGBoost model was retrained and evaluated using the same 5-fold cross-validation framework. In this expanded cohort, the XGBoost model achieved an AUC of 0.82 (95% CI: 0.77–0.87), sensitivity of 0.76, and specificity of 0.77. While these metrics represent a modest reduction compared to the primary analysis (AUC 0.86), the overall discriminative ability remained in the acceptable range, and the rank ordering of feature importance was preserved. The degree of slippage, multifidus fatty infiltration, and Pfirrmann grade remained the three most influential predictors. This finding suggests that while the exclusion of surgical non-responders did sharpen the classification boundary, the model captures genuine morphological differences between cohorts rather than being solely an artifact of patient selection.

Additionally, to evaluate the necessity of SMOTE resampling, we re-ran the XGBoost pipeline using the natural class distribution (43.3% vs. 56.7%) without any oversampling technique. The model achieved an AUC of 0.84 (95% CI: 0.79–0.89), sensitivity of 0.78, and specificity of 0.79. The marginal difference in AUC (0.84 vs. 0.86) was not statistically significant (DeLong test, p = 0.31), indicating that SMOTE did not introduce meaningful interpolation artifacts or artificially inflate model performance. We elected to retain the SMOTE-augmented results as the primary analysis given the slightly higher sensitivity for the clinically actionable surgical class, but we acknowledge that both approaches yield comparable performance.

Main findings

In this study, we developed and internally validated a machine learning-based clinical decision support tool designed to assist in identifying potential surgical candidates among patients with degenerative lumbar spondylolisthesis. By integrating routine clinical data with detailed MRI features, the XGBoost model achieved a good predictive performance (AUC = 0.86), which warrants cautious interpretation given the single-center, retrospective design. These results should be considered preliminary and hypothesis-generating, representing an internal validation that requires confirmation through multi-center external validation before any clinical deployment can be considered. The incorporation of SHAP analysis provided a degree of interpretability, suggesting that the degree of slippage, multifidus fatty infiltration, and Pfirrmann disc degeneration grade were among the primary features associated with surgical necessity in this cohort.

Key imaging predictors

The identification of specific MRI features as key predictors is broadly consistent with the existing biomechanical and clinical literature, though direct comparisons should be made cautiously given differences in study design and patient populations. The degree of slippage has long been recognized as a primary indicator of segmental instability (21). However, the model's reliance on multiple features simultaneously suggests that slippage alone may be insufficient for comprehensive decision-making. The prominent role of multifidus fatty infiltration in the SHAP analysis is noteworthy. The multifidus muscle is considered a crucial dynamic stabilizer of the lumbar spine, and severe fatty infiltration and atrophy have been reported to compromise this stabilizing function, potentially exacerbating mechanical stress on the degenerated disc and facet joints (22). One possible interpretation of our findings is that when dynamic muscular stabilization is compromised (as indicated by high fatty infiltration), conservative therapies may be less likely to achieve satisfactory outcomes, though this hypothesis requires prospective confirmation (23).

Advanced Pfirrmann grading was also a strong predictor for surgery in our cohort. Severe disc degeneration (Grades IV and V) is associated with significant loss of disc height, which may contribute to foraminal stenosis and nerve root compression—symptoms that are often reported as refractory to nonoperative management (24). The model's reliance on these interacting features reflects the complex pathology of DLS. However, we acknowledge that muscular degeneration might also represent a consequence of chronic pain and disuse rather than the primary driver for surgery, highlighting the correlative rather than causative nature of our findings.

Clinical implications

The clinical translation of these findings requires rigorous prospective evaluation. The transition from conservative to surgical management in DLS currently exhibits significant variability, heavily influenced by physician experience and institutional protocols (25). If externally validated, the proposed XGBoost model could potentially provide a supplementary, quantitative reference for assessing the likelihood of failing conservative therapy. The SHAP force plots may allow clinicians to visualize the specific anatomical factors associated with a patient's individual risk profile, potentially facilitating more structured discussions during shared decision-making. For example, a surgeon could reference the model's output to discuss with a patient how their degree of multifidus atrophy and disc degeneration might relate to the expected trajectory of conservative care. It is important to emphasize, however, that such a tool should serve as a complement to—rather than a replacement for—comprehensive clinical judgment, patient preferences, and the surgeon's experience. We envision that, following successful external validation, such a model could be deployed as a web-based clinical calculator or integrated into electronic health record systems to provide real-time decision support at the point of care.

Limitations and future directions

This study has several important limitations that must be acknowledged. A primary limitation is the retrospective, single-center design, which introduces selection bias and substantially limits the generalizability of the findings to other clinical settings and patient populations. The results should therefore be interpreted as hypothesis-generating rather than definitive. All claims regarding clinical utility in this manuscript are confined to internal validation; prospective, multi-center external validation using independent patient cohorts is mandatory before any consideration of clinical deployment.

Second, while 5-fold cross-validation was employed to mitigate overfitting, this approach provides only an internal estimate of model performance; external validation using an independent, multi-institutional cohort is indispensable before any consideration of clinical deployment.

Third, the model relied exclusively on static, supine MRI features. This represents a critical limitation because surgical indications for DLS rely heavily on dynamic segmental instability observed on flexion-extension radiographs and standing spinopelvic alignment parameters (e.g., pelvic incidence, lumbar lordosis, pelvic tilt, sagittal vertical axis). Occult mechanical instability—which may be evident only on dynamic imaging—cannot be captured by supine MRI and may lead to misclassification of patients who appear stable on static imaging but demonstrate clinically significant translational motion under physiological loading. The absence of these dynamic parameters restricts the model's capacity to evaluate the complete biomechanical profile of the patient (26). Future multi-modal iterations of this model should integrate flexion-extension radiographic data and erect full-spine standing films to capture dynamic instability and global spinopelvic malalignment, which are increasingly recognized as essential determinants of surgical planning in DLS.

Fourth, our definition of the “ground truth” (successful surgery vs. successful conservative care) introduces a potential risk of circular reasoning. By categorizing patients based on post-treatment outcomes to predict pre-treatment decisions, the model inherently learns past successful decision patterns rather than absolute biological indications for surgery. Patients who underwent conservative treatment but might have benefited from surgery, or those who had poor surgical outcomes, were not explicitly modeled. The sensitivity analysis including the 23 surgical non-responders (Section 3.5) partially addresses this concern by demonstrating that the model retains acceptable discriminative ability even when the ground truth boundary is less distinct. Nevertheless, this outcome-based classification limits the model's ability to independently define surgical necessity and should be interpreted accordingly.

Fifth, the final sample size of 342 patients (from an initial pool of 546), while adequate for the number of features evaluated, remains modest for complex ML algorithms, and the results may not be stable across different sampling conditions.

Sixth, although inter-observer reliability for the MRI feature extraction was substantial to excellent (Section 2.3), the assessment was limited to two radiologists at a single institution. The generalizability of these measurements across different imaging protocols, scanner manufacturers, and reader experience levels remains to be established. Furthermore, patient-reported outcomes beyond ODI (e.g., VAS for leg and back pain, EQ−5D quality of life) were not incorporated into the outcome definition and might provide additional dimensions for characterizing treatment success.

Future iterations of the model should aim to integrate multi-modal imaging data, including dynamic radiographs and full-spine standing films, to capture a more comprehensive biomechanical profile. Prospective, multi-center validation studies with larger sample sizes are the critical next step. Additionally, the development of automated, deep learning-based segmentation tools for rapid extraction of muscle and disc parameters from MRI would be important to facilitate the potential integration of such prediction models into routine clinical workflows (27). Recent advances in deep learning-based object detection algorithms have demonstrated the feasibility of automated spondylolisthesis detection and localization on lumbar radiographs, suggesting that fully automated imaging pipelines may eventually complement clinical decision support systems such as the one proposed in this study (28).

In this retrospective single-center preliminary internal validation study, we developed and internally validated an interpretable machine learning model that demonstrated good performance in distinguishing surgical candidates from patients amenable to conservative management among those with degenerative lumbar spondylolisthesis. SHAP analysis suggested that the degree of slippage, multifidus fatty infiltration, and Pfirrmann disc degeneration grade were the most influential imaging features in this cohort. While these preliminary findings are encouraging, the clinical utility of this tool remains to be established through prospective, multi-center external validation. Until such validation is completed, this model should be regarded as a research tool that generates hypotheses regarding the morphological determinants of surgical candidacy, rather than a clinically deployable decision aid. If confirmed through rigorous external validation, this decision support tool may serve as a supplementary reference to assist clinicians in optimizing patient selection and facilitating shared decision-making in the management of DLS.