Explainable Supervised Learning Classification Model Using Diffusion Tensor Imaging Predicts Postoperative Outcomes in Cervical Spondylotic Myelopathy: A Preliminary Study Authors: Yifei Peng, Zixuan Zhang, Zhikun Zhang, Kaiyi Shi, Ruoyu Wang, Yue Liu, Li Zhang Published: 2026 ABSTRACT ABSTRACTBackgroundCervical spondylotic myelopathy (CSM) is common in older adults. Some patients may experience incomplete neurological recovery after surgery, or even deterioration. Accurate prognosis is essential for patients, yet current tools use subjective scores and fail to detect early spinal cord microstructural changes.MethodsCSM patients undergoing preoperative cervical MRI and diffusion tensor imaging (DTI) scans between 2024 and 2025 at a single center were included in this retrospective study. Data were retrieved from medical records, and the determination of whether patients achieved minimal clinically important difference (MCID) was based on the change in modified Japanese Orthopaedic Association (mJOA) scores. Three supervised learning classification models, namely extreme gradient boosting (XGB), logistic regression (LR), and support vector machine (SVM), were constructed based on DTI and clinical risk factors. The performance of these models was evaluated using the area under the receiver operating characteristic curve (AUC) and precision‐recall curve (AP), and decision curve analysis (DCA). The contribution of each feature to model prediction was visualized by SHapley Additive exPlanations (SHAP).ResultsThe training set included 163 patients (mean age: 54.43 ± 10.27 years; 57 males), whereas the testing set included 71 patients (mean age: 54.86 ± 12.25 years; 23 males). The AUCs for XGB, LR, and SVM models in the training set were 0.940, 0.791, and 0.908 respectively (p < 0.05). The XGB model showed the best performance in the training set (AUC = 0.940, AP = 0.948), and the results in the testing set demonstrated moderate discriminative ability with good precision (AUC = 0.754; 95% CI, 0.643–0.857, AP = 0.851; 95% CI, 0.742–0.937). DCA showed that the clinical utility of the XGB model was relatively high.ConclusionAn explainable XGB model based on DTI and clinical risk factors provides a foundation for preoperative risk stratification of MCID achievement in postoperative CSM.Keywords: cervical spondylotic myelopathy, diffusion tensor imaging, minimal clinically important difference, supervised learning classification model Abbreviations apparent diffusion coefficient area under the precision‐recall curve area under the receiver operating characteristic curve body mass index confidence intervals cervical spondylotic myelopathy decision curve analysis diffusion tensor imaging fractional anisotropy intraclass correlation coefficients logistic regression minimal clinically important difference modified Japanese Orthopedic Association machine learning Number of Signal Averages odds ratio SHapley Additive exPlanations support vector machine echo time repetition time extreme gradient boosting Cohen's kappa Introduction Cervical spondylotic myelopathy (CSM) is a prevalent form of chronic spinal cord injury in middle‐aged and elderly people. With the increasing aging of the global population, its incidence is expected to gradually increase [1]. The main cause is cervical spondylosis leading to spinal canal stenosis and spinal cord compression. Early symptoms of CSM are often insidious, and severe motor and sensory neurological deficits appear in the middle and late stages [2, 3]. Cervical decompressive surgery is currently the most commonly used method to delay disease progression and improve patient symptoms [4]. However, due to the complex etiology and long duration of CSM, the recovery of neurological function in some patients after surgery is often not satisfactory, and deterioration may even occur [5]. Therefore, accurate prediction of postoperative prognosis in CSM patients is very important for patient expectation management. Current prognosis evaluation relies on neurological function score improvements, which are subjective and may not reflect true differences [6]. In addition, conventional magnetic resonance imaging is still used for the diagnosis and management of CSM [7]. Although conventional MRI can show the morphology of the compressed spinal cord and intramedullary abnormal signals, it is not sensitive to early microscopic changes [8]. Diffusion tensor imaging (DTI) is an important prognostic biomarker of disease severity in CSM patients. Compared with conventional magnetic resonance imaging, it is more sensitive to subtle pathological changes in the spinal cord [9, 10]. Therefore, DTI can also identify changes in CSM in the early stage, which helps to determine the appropriate timing of surgery before irreversible chronic changes begin [11]. The modified Japanese Orthopaedic Association score (mJOA) is the standard tool for assessing CSM outcomes, with its minimal clinically important difference (MCID) set at approximately 2.5 points [12, 13]. In recent years, machine learning (ML) has been widely used in clinical practice and has made great progress in predicting the prognosis of CSM patients [14], but no study has constructed a machine learning model based on clinical features and DTI features to predict patient prognosis. Therefore, this study aims to construct three explainable supervised learning classification models based on DTI features and clinical risk factors to predict postoperative MCID in CSM patients and evaluate model performance. This approach will enhance patient management and treatment outcomes by offering a more reliable and objective assessment tool. Materials and Methods This study was approved by the institutional review board, and due to its retrospective nature, the informed consent was waived. Patient Identification This retrospective study included patients who underwent cervical decompressive surgery for CSM at the First Hospital of Hebei Medical University between May 2024 and May 2025. We initially identified 328 consecutive patients who underwent surgery for cervical spondylotic myelopathy. Among these, 94 patients were excluded based on predefined criteria and 234 patients were finally included in the study. Inclusion criteria: ① Patients with clinical and radiological diagnosis of CSM who underwent cervical decompressive surgery. Exclusion criteria included: ① Patients with poor image quality. ② Other cervical diseases, including trauma, tumors, infections, rheumatoid arthritis, ankylosing spondylitis, and so forth. ③ Patients with claustrophobia or who were unable to lie flat for examination. ④ Patients with a history of cervical spine and spinal cord surgery. ⑤ Patients with other serious systemic diseases. ⑥ Patients with a history of long‐term alcohol or drug dependence. ⑦ Patients who were followed up for less than 12 months (Figure 1). Flowchart of the study. Data were retrieved from medical records. Baseline data before and after treatment were collected for all patients, including age at first visit, duration of symptoms, gender, body mass index (BMI), past medical history, and baseline mJOA score. All patients underwent cervical decompressive surgery performed by the same surgeon, and the mJOA score was used as the standard for assessing the degree of spinal cord nerve injury. The maximum mJOA score is 18, and the more severe the nerve injury, the lower the score. The minimal clinically important difference (MCID) was used as a marker for assessing whether the disease had achieved clinical remission (ΔmJOA = postoperative mJOA score—preoperative mJOA score). When the patient's ΔmJOA > 2.5, the patient's neurological and motor functions were significantly improved, that is, the disease was considered to be in clinical remission. According to whether the patient's ΔmJOA was greater than 2.5, the patients were divided into two groups, that is, the group that reached MCID (ΔmJOA > 2.5, 57.7%) and the group that did not reach MCID (ΔmJOA < 2.5, 42.3%), representing good prognosis and poor prognosis, respectively. These patients were randomly divided into training (n = 163, 70%) and testing (n = 71, 30%) sets using a computer‐generated randomization sequence. The training set included 95 patients with good prognosis (ΔmJOA > 2.5) and 68 with poor prognosis (ΔmJOA < 2.5), while the testing set included 41 and 30 patients, respectively. MRI Protocol All patients underwent preoperative imaging data collection using a 3.0T MRI scanner (Ingenia CX, Philips Healthcare, Best, Netherlands) with an 8‐channel phased‐array coil. Patients were placed head‐first in a supine position for conventional cervical MRI and axial DTI scans, with conventional MR scan parameters shown in Table S1. The cervical DTI was scanned using a single‐shot echo‐planar imaging sequence, with a diffusion weighting coefficient b value of 0, 800 s/mm2, 15 diffusion‐sensitive gradient directions, TR/TE = 2772/85 ms, slice spacing = 0, slice thickness = 2.5 mm, field of view = 224 × 224 mm, matrix = 92 × 90, NSA = 2. Image Analysis Three radiologists (with 12, 18, and 20 years of experience in cervical spondylosis MRI diagnosis, respectively) independently assessed the preoperative qualitative and quantitative MRI parameters. They were unaware of the patients' clinical data and postoperative outcomes before the assessment. Measurements were averaged across observers for subsequent model development. The conventional MRI parameters recorded in this study are as follows: the narrowest position of the spinal canal was determined on T2WI, the anteroposterior diameter of the spinal canal was measured on sagittal T2WI images, and the cross‐sectional area of the spinal canal was measured on axial T2WI images. The location of stenosis, T2 abnormal signal manifestation, and cervical curvature were judged on T2WI (Figure 2a). Cervical curvature was classified and encoded as follows: normal curvature = 0, straightening of curvature = 1, and kyphosis = 2. T2 abnormal signal manifestations were defined as hyperintense signal within the spinal cord on T2WI and were categorized as: (0) no abnormal signal, (1) abnormal signal (faint, clear, or intense). Measurement (circle highlighted in red) of the cross‐sectional area of the spinal canal (a) in axial T2WI and FA (b) and ADC values (c) in axial DTI for CSM patients. ADC, apparent diffusion coefficient (mm2/s); CSM, cervical spondylotic myelopathy; DTI, diffusion tensor imaging; FA, fractional anisotropy. The conventional DTI parameter recording method is as follows. The image was imported into Neuro Perfusion. The most severely compressed segment of the spinal cord, determined by the maximum canal stenosis on T2WI, was selected. Subsequently, the fractional anisotropy (FA) map and apparent diffusion coefficient (ADC) map were reconstructed on the axial plane. The FA and ADC values were automatically calculated by the scanner based on the direction information of the diffusion tensor described by the deterministic fixed‐step length tracking algorithm (Figure 2b). Construction of Supervised Learning Classification Models Univariate analysis of 12 clinical features and 3 imaging features was first conducted in the training set (n = 163) to screen for potential predictors. Variables with p < 0.05 in univariate analysis were subsequently entered into multivariate logistic regression (LR) analysis in the same training set. Variable selection for supervised learning classification models followed a hybrid approach: (1) variables with p < 0.05 in multivariate analysis were included, and (2) variables with established clinical relevance based on prior literature (e.g., age [15]) were also included regardless of statistical significance to ensure clinical comprehensiveness. Therefore, independent clinical risk factors were used to construct three supervised learning classification models: extreme gradient boosting (XGB), LR, and support vector machine (SVM). Model Performance Evaluation Model performance was evaluated using a comprehensive framework encompassing discrimination, calibration, classification, and clinical utility. Discrimination was assessed by the area under the receiver operating characteristic curve (AUC) and the area under the precision‐recall curve (AP), with bootstrap resampling (1000 iterations) to calculate 95% confidence intervals. The DeLong test was used to compare AUCs between models, with p < 0.05 indicating statistical significance. Calibration was evaluated using calibration curves (quantile‐based binning, n = 5) and Brier scores to assess the agreement between predicted probabilities and observed outcomes. Classification performance was quantified by accuracy, sensitivity, specificity, precision, and F1 score, visualized using confusion matrices. Clinical utility was determined by decision curve analysis (DCA) to estimate net benefit across threshold probabilities. Radar charts were constructed to integrate multiple performance metrics. Additionally, stratified performance analyses were conducted by age (< 60 vs. ≥ 60 years) and baseline mJOA severity (mild [≥ 15] vs. severe [< 15]) to evaluate model robustness across clinically relevant subgroups. Model Interpretation The best‐performing model was selected for interpretation. The SHapley Additive exPlanations (SHAP) values of each feature were calculated to measure the contribution of each feature to model prediction. Features were ranked by importance, with higher‐ranked features contributing more to the model. Additionally, the SHAP dependence plots showed the positive and negative contributions of features. Statistical Analysis Statistical analysis was performed using Python version 3.8 and IBM SPSS version 27.0. Inter‐ and intra‐rater reliability was assessed using intraclass correlation coefficients (ICC) for continuous variables (FA, ADC, cross‐sectional area) and Cohen's kappa (κ) for categorical variables (curvature, T2 hyperintensity). For intra‐rater reliability, one radiologist repeated measurements on 30 randomly selected cases after a 4‐week interval. For inter‐rater reliability, all three radiologists independently assessed the same 30 cases. Quantitative data that followed a normal distribution were expressed as mean and standard deviation, and independent sample t‐tests were performed. Nonnormally distributed quantitative data were expressed as median (interquartile range), and Mann–Whitney U tests were performed. Categorical data were expressed as frequencies, and chi‐square tests were performed. Univariate and multivariate analyses were performed using binary logistic regression with the enter method. Variables with p < 0.05 in univariate analysis were included in the multivariate model. Continuous variables were entered as raw values; categorical variables were dummy‐coded. Results Patient Characteristics A total of 234 CSM patients were analyzed, including 80 females and 154 males, aged 25 to 82 years, with a median age of 54 years. Patient characteristics are shown in Table 1. There were no significant differences between the training and testing groups in all variables (p > 0.05), indicating good balance between the two groups in terms of clinical and radiological features. Inter‐ and intra‐rater reliability analyses demonstrated good to excellent agreement for all imaging parameters (ICC range: 0.91–0.96 for continuous variables; κ range: 0.79–0.89 for categorical variables). TABLE 1Comparison of clinical and imaging parameters between the training and testing groups. Italics are used to denote the significance levels (p‐value).FactorCategoryTraining setTesting set p Gender (%)Male106 (65.03)48 (67.61)0.70Female57 (34.97)23 (32.39)Age (years) — 54.43 ± 10.2754.86 ± 12.250.80BMI (kg/m2) — 23.94 ± 2.0523.67 ± 2.130.36Diabetes (%)No131 (80.37)57 (80.28)0.99Yes32 (19.63)14 (19.72)Smoking (%)No83 (50.92)36 (50.70)0.98Yes80 (49.08)35 (49.30)Duration (%)< 6 months84 (51.53)38 (53.52)0.78≥ 6 months79 (48.47)33 (46.48)Surgical procedure (%)Anterior approach124 (76.07)51 (71.83)0.91Posterior approach29 (17.79)18 (25.35)Combined approach10 (6.13)2 (2.82)Location (%)C3‐420 (12.27)13 (18.31)0.82C4‐547 (28.83)14 (19.72)C5‐666 (40.49)31 (43.66)C6‐730 (18.40)13 (18.31)Diseased segment (%)1105 (64.42)42 (59.15)0.45258 (35.58)29 (40.85)Curvature (%)Normal87 (53.37)37 (52.11)0.93Straightening48 (29.45)22 (30.99)Kyphosis28 (17.18)12 (16.90)Cross‐sectional area (mm2)—84.89 (68.59, 97.16)82.18 ± 17.040.93T2 hyperintensity (%)No110 (67.48)45 (63.38)0.54Yes53 (32.52)26 (36.62)FA*103 — 508.52 ± 36.55514.06 ± 44.480.36ADC (mm2/s) — 0.92 (0.83, 0.98)0.89 ± 0.160.19Baseline mJOA10.00 (8.00, 12.00)10.00 (7.00, 12.00)0.95ΔmJOA3.00 (1.00, 5.00)3.45 ± 2.840.15MCID (%)No73 (44.79)26 (36.62)0.25Yes90 (55.21)45 (63.38)Abbreviations: ΔmJOA = postoperative mJOA score—preoperative mJOA score; ADC, apparent diffusion coefficient (mm2/s); BMI, body mass index (kg/m2); FA, fractional anisotropy; MCID, minimal clinically important difference; mJOA, modified Japanese Orthopaedic Association; Pre, preoperative. Comparison of clinical and imaging parameters between the training and testing groups. Italics are used to denote the significance levels (p‐value). Abbreviations: ΔmJOA = postoperative mJOA score—preoperative mJOA score; ADC, apparent diffusion coefficient (mm2/s); BMI, body mass index (kg/m2); FA, fractional anisotropy; MCID, minimal clinically important difference; mJOA, modified Japanese Orthopaedic Association; Pre, preoperative. Independent Clinical Risk Factors Univariate and multivariate analyses showed that curvature encoding, FA, and baseline mJOA were significantly associated with reaching MCID (p < 0.05). Age was not significant in multivariate analysis (p > 0.05) but was retained in the models based on its established clinical relevance. Therefore, age, curvature, FA, and baseline mJOA were independent clinical predictors of CSM prognosis. Age and FA were significantly positively correlated with reaching MCID (OR > 1), while curvature encoding and baseline mJOA were significantly negatively correlated with reaching MCID (OR < 1) (Table 2). TABLE 2Univariate and multivariate analysis results of clinical and imaging characteristics in the training set. Italics are used to denote the significance levels (p‐values).FactorCategoryUnivariate analysisMultivariate analysisOR (95% CI) p OR (95% CI) p Gender (%)Male1.65 (0.85–3.19)0.14FemaleAge (years)—1.04 (1.00–1.07)0.031.03 (1.00–1.07)0.07BMI (kg/m2)—0.98 (0.85–1.14)0.83Diabetes (%)No1.05 (0.48–2.30)0.90YesSmoking (%)No1.62 (0.87–3.02)0.13YesDuration (%)< 6 months0.47 (0.25–0.87)0.020.57 (0.28–1.16)0.12≥ 6 monthsSurgical procedure (%)Anterior approach0.93 (0.54–1.58)0.77Posterior approachCombined approachLocation (%)C3‐40.99 (0.70–1.38)0.93C4‐5C5‐6C6‐7Diseased segment (%)11.11 (0.58–2.12)0.752Curvature (%)Normal0.55 (0.36–0.85)< 0.010.50 (0.31–0.80)< 0.01StraighteningKyphosisCross‐sectional area (mm2)—0.25 (0.03–2.29)0.22T2 hyperintensity (%)No1.22 (0.63–2.36)0.56YesFA*103 1.02 (1.01–1.03)< 0.011.02 (1.01–1.04)< 0.01ADC (mm2/s)—4.31 (0.56–35.8)0.18Baseline mJOA—0.75 (0.65–0.85)< 0.010.722 (0.61–0.86)< 0.01Abbreviations: ADC, apparent diffusion coefficient (mm2/s); BMI, body mass index (kg/m2); CI, confidence interval; FA, fractional anisotropy; mJOA, modified Japanese Orthopaedic Association; OR, odds ratio. Univariate and multivariate analysis results of clinical and imaging characteristics in the training set. Italics are used to denote the significance levels (p‐values). Abbreviations: ADC, apparent diffusion coefficient (mm2/s); BMI, body mass index (kg/m2); CI, confidence interval; FA, fractional anisotropy; mJOA, modified Japanese Orthopaedic Association; OR, odds ratio. Performance of Supervised Learning Classification Models In the training set, the XGB model showed the highest AUC value (0.940; 95% CI, 0.902–0.971), AP value (0.948; 95% CI, 0.910–0.977), accuracy (0.853; 95% CI, 0.798–0.902), recall (0.911; 95% CI, 0.844–0.968), and F1 value (0.872; 95% CI, 0.817–0.919). In the testing set, the XGB model also showed the highest AP value (0.851) (Figure 3, Table 3). The DeLong test showed that in the training set, both XGB and SVM significantly outperformed logistic regression (both p < 0.01), while there was no significant difference between XGB and SVM. Performance of the three supervised learning classification models, namely XGB, logistic regression, and SVM. (a) ROC curves for training set. (b) ROC curves for testing set. (c) AP curves for training set. (d) AP curves for testing set. AP, precision‐recall; ROC, receiver operating characteristic; SVM, support vector machine; XGB, extreme gradient boosting. TABLE 3Performance of various models in the training and testing sets.Model (CI)AUCAPAccuracyRecallPrecisionF1 scoreTrainXGB0.940 (0.902, 0.971)0.948 (0.910, 0.977)0.853 (0.798, 0.902)0.911 (0.844, 0.968)0.837 (0.763, 0.900)0.872 (0.817, 0.919)SVM0.908 (0.863, 0.951)0.912 (0.853, 0.960)0.834 (0.810, 0.908)0.833 (0.822, 0.954)0.862 (0.786, 0.928)0.848 (0.825, 0.919)LR0.791 (0.718, 0.858)0.834 (0.763, 0.893)0.724 (0.650, 0.792)0.756 (0.659, 0.846)0.747 (0.656, 0.833)0.751 (0.676, 0.819)TestXGB0.754 (0.643, 0.857)0.851 (0.742, 0.937)0.676 (0.578, 0.789)0.756 (0.628, 0.882)0.739 (0.614, 0.867)0.747 (0.649, 0.835)SVM0.705 (0.575,0.823)0.766 (0.643, 0.911)0.718 (0.578, 0.789)0.756 (0.625, 0.878)0.791 (0.628, 0.878)0.773 (0.649, 0.845)LR0.763 (0.645, 0.877)0.843 (0.741, 0.933)0.704 (0.606, 0.803)0.756 (0.625, 0.875)0.773 (0.644, 0.897)0.764 (0.658, 0.851)Abbreviations: AP, the area under the precision‐recall curve; AUC, the area under the receiver operating characteristic curve; CI, confidence interval; LR, logistic regression; SVM, support vector machine; XGB, extreme gradient boosting. Performance of various models in the training and testing sets. Abbreviations: AP, the area under the precision‐recall curve; AUC, the area under the receiver operating characteristic curve; CI, confidence interval; LR, logistic regression; SVM, support vector machine; XGB, extreme gradient boosting. (p = 0.269). In the testing set, no significant differences were observed between any pair of models (all p > 0.05, Table 4). TABLE 4Performing the DeLong test on the AUC values of the three models. Italics are used to denote the significance levels (p‐values). p‐valueXGB vs. SVMXGB vs. LRSVM vs. LRTrain0.269< 0.01< 0.01Test0.5130.8960.432Abbreviations: LR, logistic regression; SVM, support vector machine; XGB, extreme gradient boosting. Performing the DeLong test on the AUC values of the three models. Italics are used to denote the significance levels (p‐values). Abbreviations: LR, logistic regression; SVM, support vector machine; XGB, extreme gradient boosting. Calibration curves (quantile‐based binning, n = 5) revealed that logistic regression produced the smoothest calibration profile across both datasets. In the training set, XGB exhibited marked calibration scatter with a steep step‐like pattern, deviating substantially from perfect calibration, particularly in the intermediate probability ranges (0.3–0.6). This is consistent with the tendency of gradient boosting algorithms to overfit training data, especially in relatively small samples. In the testing set, XGB calibration improved but still showed non‐negligible fluctuation in the low‐to‐intermediate probability range (0.4–0.6). Overall, XGB maintained the most favorable Brier score despite its calibration scatter (Figure S1, Table 5). TABLE 5Brier scores for the three models in the training and testing sets. Lower Brier scores indicate better probability calibration.ModelXGBSVMLRTrain0.1580.1360.193Test0.2000.2080.209Abbreviations: LR, logistic regression; SVM, support vector machine. Brier scores for the three models in the training and testing sets. Lower Brier scores indicate better probability calibration. Abbreviations: LR, logistic regression; SVM, support vector machine. DCA demonstrated that XGB and SVM achieved a higher net benefit than logistic regression in the training set across threshold probabilities of 0.1–0.8. In the testing set, all three models showed similar net benefit profiles, consistently outperforming both extreme strategies (treat all/none) within the clinical decision range (Figure 4). Decision curve analysis for training set (a) and testing set (b). SVM, support vector machine. Stratified analyses demonstrated XGB's superior robustness across clinically relevant subgroups, maintaining the highest discriminative performance by age and mJOA severity (Figure S2). Confusion matrices illustrated performance differences at the default threshold: in the training set, XGB and SVM correctly classified 82 and 75 good prognosis cases respectively, versus 68 for logistic regression; in the testing set, all models correctly identified 34 good prognosis cases, but XGB misclassified more poor prognosis cases as good (12 vs. 9 for SVM and 10 for logistic regression), accounting for its lower specificity (Figure S3). Among the three models, XGB was selected for subsequent SHAP analysis because it achieved the highest training AUC (0.940), the highest testing AP (0.851), the most favorable Brier score (0.2000), and maintained robust discrimination across age and mJOA severity subgroups, coupled with the superior interpretability of tree‐based models. The radar chart further confirmed the performance of the three models by integrating six indicators (AUC, AP, accuracy, recall, precision, F1). The XGB model performed excellently in the training set and also performed well in the testing set (Figure S4). Model Interpretation We used SHAP to visualize the XGB model (Figure 5). SHAP ranked the features in the prediction, with baseline mJOA having the greatest impact on predicting MCID in CSM patients, followed by FA. SHAP dependence plots revealed distinct patterns of feature effects. Baseline mJOA showed the strongest negative association with MCID prediction (higher preoperative function decreased predicted probability of achieving MCID), while FA demonstrated a positive association (higher FA increased the predicted probability). Age exhibited a weaker positive trend, and curvature showed a nonlinear pattern with limited impact. Normal curvature showed slightly positive SHAP values, suggesting a modest increase in predicting MCID probability, while straightening demonstrated neutral effects around zero. Notably, kyphosis displayed negative SHAP values, indicating that patients with cervical kyphosis were less likely to achieve MCID compared to those with normal curvature. SHAP summary and dependence plots of XGB model. (a) SHAP summary plot showing the distribution of the SHAP values for each feature. (b–e) SHAP dependence plots of baseline mJOA, FA, age, and curvature. FA, fractional anisotropy; mJOA, modified Japanese Orthopaedic Association; SHAP, SHapley Additive exPlanation; XGB, extreme gradient boosting. Discussion This study successfully developed and validated an explainable supervised learning classification model based on DTI and clinical risk factors, providing a foundation for preoperative risk stratification of MCID achievement in patients with CSM after surgery. Among the three models evaluated, the XGB model demonstrated superior performance in terms of discriminative ability and clinical utility, which can assist physicians in postoperative management. Traditional assessment methods mostly rely on limited clinical variables or single radiological parameters, which are difficult to fully reflect the multidimensional characteristics of spinal cord lesions [16, 17]. This study analyzed the preoperative radiological features of CSM patients based on multimodal MRI, adding the FA parameter with high predictive potential on the basis of previous studies [18, 19], thereby constructing a more accurate prognostic model. This method can help clinicians more accurately identify patients at different risk levels and facilitate the formulation of personalized postoperative recovery management strategies. In previous studies, T2WI and DTI parameters were commonly used as radiological indicators for prognostication. T2WI can intuitively show the edema state of the spinal cord compression site [20], while DTI can quantify water molecule diffusion and reflect the microscopic structural changes of spinal cord tissue [21]. However, in this study, the predictive role of T2 abnormal signal in the prognosis assessment of CSM patients was not significant, a result also seen in other studies [22, 23]. The possible reason is the limitations of imaging technology, the inconsistent grading criteria for T2WI high signals, and the wide variation range of T2WI, which leads to large errors and research biases in qualitative analysis [24]. Moreover, T2WI only shows high signals when the water content increases to a certain extent, which often indicates that the spinal cord has been severely compressed for a long time, that is, the stage of severe cervical spondylosis or the later stage of spinal cord compression, thus resulting in insufficient sensitivity in this study [25]. Some studies have shown that DTI has a more accurate predictive role than T2WI [26], but ADC in this study was still not significant, possibly due to individual differences or a small sample size, which underestimated its predictive role. This further highlights the superior predictive value of FA. Shabani et al. pointed out that DTI shows the ability to predict the severity of CSM and the recovery of neurological function, indicating its potential in the prognostic assessment of cervical spondylotic myelopathy [27]. Previous studies have applied machine learning to the prognostication of CSM [28]. For example, Sitan Feng et al. developed various machine learning algorithms based on 10 basic predictive variables to predict the one‐year outcome after anterior cervical discectomy and fusion in patients with cervical spondylosis [29]. Meng Ze Zhang et al. developed an additional tree model based on radiomics features to predict postoperative recovery of CSM [30]. This study constructed three supervised learning classification models: XGB, LR, and SVM, based on the above predictive factors, successfully integrating clinical risk factors (age, curvature and baseline mJOA) with DTI features to predict the minimal clinically important difference in CSM patients after surgery, highlighting the predictive value of the integration of imaging and clinical data. We found that the XGB model showed the best performance in the training set and the highest AP value in the testing set, demonstrating the model's good discriminative ability. In addition, the DCA showed that XGB and SVM achieved a higher net benefit than logistic regression in the training set across threshold probabilities of 0.1–0.8, further indicating that the model could provide value for the formulation of personalized treatment plans, preoperative patient communication, and postoperative rehabilitation guidance. We acknowledge the noticeable performance drop between the training set (AUC 0.940) and testing set (AUC 0.754) for the XGB model, which suggests some degree of overfitting. This is likely due to the relatively small sample size and the complexity of the XGB algorithm, which may memorize training data patterns that cannot generalize perfectly to unseen data. However, the XGB model still maintained moderate discriminative ability in the testing set (AUC 0.754) and demonstrated the highest AP value (0.851). Notably, logistic regression showed more stable performance between training and testing sets, suggesting it may be more robust in generalization in small datasets. Future studies with larger sample sizes are needed to confirm whether the XGB model's superior training performance translates into better external validation. The decision‐making process of machine learning is difficult to interpret, but SHAP analysis helps us better understand the model [31]. The SHAP analysis revealed that baseline mJOA had the greatest impact on predicting MCID, followed by FA as the second most influential feature. This aligns with clinical expectations given the ceiling effect of the mJOA scale—patients with high baseline scores have limited capacity to achieve MCID. In addition, the dependence plot showed that FA and age were positively correlated with SHAP values; that is, the increase in FA value and age tended to increase the model's prediction probability of “good recovery.” This is consistent with the conclusion of Mohammad Mohammadi et al. that FA value is positively correlated with preoperative and postoperative ΔmJOA [32]. However, previous studies have shown that younger patients have better prognosis than older patients [33], which differs from this study. This may be because the dataset in this study included younger patients, and these younger patients had better preoperative neurological function, so the possibility of reaching the minimal clinically important difference after surgery was lower. Besides, Kiyoshi Ito et al. found that there was no statistical difference in the recovery rate between patients < 80 years old and ≥ 80 years old [34]. The dependence plot revealed a non‐linear pattern for cervical curvature with limited overall impact on MCID prediction. Normal curvature (0) showed slightly positive SHAP values, suggesting a modest increase in predicted MCID probability, while straightening (1) demonstrated neutral effects around zero. Notably, kyphosis (2) displayed negative SHAP values, indicating that patients with cervical kyphosis were less likely to achieve MCID compared to those with normal curvature. This finding aligns with previous studies [35]. There are several limitations in this study. First, the limited sample size (n = 234) and single‐center design may affect generalizability; external validation with larger, multicenter datasets is warranted. Second, restricting analysis to patients with complete preoperative MRI/DTI and ≥ 12 months follow‐up may introduce information bias by systematically excluding patients with technically challenging imaging (potentially indicating more severe disease) or early loss to follow‐up (potentially reflecting poor postoperative compliance). While the consecutive nature of initial patient identification minimizes selection bias, this attrition rate represents a limitation. Future prospective studies should implement standardized imaging protocols and active follow‐up strategies to mitigate this bias. Third, while we acknowledge that comprehensive prognostic models should include a more detailed characterization of preoperative symptom severity, grading of spinal cord compression, surgical details (e.g., number of levels treated, whether fusion was performed), predominant symptom types, postoperative rehabilitation, and complications, these were not systematically available in our retrospective dataset. Our deliberate focus on four preoperative predictors (age, curvature, FA, baseline mJOA) addresses a specific clinical need for pre‐surgical risk stratification, but this parsimony may limit the model's comprehensiveness. Performance comparison of different models (Table S2) demonstrated that FA consistently ranked as the second most important predictor across all model complexities. This confirms that FA's contribution was not substantially overestimated. Fourth, the standardized integration of multimodal MRI data remains challenging, particularly in combining DTI with conventional sequences. The DTI analysis was restricted to single‐slice measurements at the most compressed segment, which may not fully capture intra‐segmental heterogeneity or spatial variations along the spinal cord. We emphasize that our goal was to develop a clinically implementable tool using readily available preoperative DTI, rather than to demonstrate the full potential of ML for high‐dimensional radiomics analysis. Future prospective studies incorporating complete perioperative variables and advanced imaging features (radiomics, multislice DTI) will build upon this preliminary framework to achieve more precise prognostic assessment for CSM patients. This study constructed three supervised learning classification models based on age, cervical curvature, FA and baseline mJOA, and evaluated their performance. The XGB model demonstrated the best discriminative ability and clinical utility, and has the potential to become an important auxiliary tool for physicians in decision‐making regarding postoperative management. Notably, the parsimonious model based on only four preoperative predictors outperformed the full‐feature model in the testing set (AUC 0.754 vs. 0.689), supporting the use of a simplified feature set for better generalization and clinical implementability (Table S2). However, given the wide confidence intervals in the testing set and the single‐center nature of this study, external validation with larger multicenter datasets is necessary before clinical implementation. Author Contributions Yifei Peng: conceptualization, methodology, software, formal analysis, writing – original draft, and writing – review and editing; Zixuan Zhang: data curation, software, investigation, and visualization; Zhikun Zhang: data curation, investigation, and visualization; Kaiyi Shi: data curation, formal analysis, and validation; Ruoyu Wang: investigation and resources; Yue Liu: methodology and project administration; Li Zhang: conceptualization, supervision, funding acquisition, and writing – review and editing. Funding The authors have nothing to report. Conflicts of Interest The authors declare no conflicts of interest. Supporting information Table S1: Conventional MR scanning sequences and parameters.TR, repetition time; TE, echo time; mDixon, dixon method; SAG, sagittal; TSE, turbo spin echo. Table S2: Performance comparison of different models. Model A contains 4 features: Age, Curvature, FA, and baseline mJOA; Model B contains 7 features: Age, Curvature, Cross‐Sectional Area, T2 hyperintensity, ADC, FA, and baseline mJOA; Model C contains 12 features: Gender, Age, BMI, Diabetes, Smoking, Duration, Curvature, Cross‐Sectional Area, T2 hyperintensity, ADC, FA, and baseline mJOA; Model D contains all 15 features: Gender, Age, BMI, Diabetes, Smoking, Duration, Surgical procedure, Location, Diseased segment, Curvature, Cross‐Sectional Area, T2 hyperintensity, ADC, FA, and baseline mJOA. Figure S1: Calibration curves for XGBoost, SVM, and logistic regression models in the training set (top row) and testing set (bottom row). SVM, support vector machine. Figure S2: Stratified performance of XGBoost, SVM, and logistic regression models across clinically relevant subgroups. (a) AUC by age groups (< 60 vs. ≥ 60 years). (b) AUC by baseline mJOA severity (mild ≥ 15 vs. severe < 15). (c) Accuracy by age groups. (d) Accuracy by baseline mJOA severity. AUC, area under the receiver operating characteristic curve; mJOA, modified Japanese Orthopedic Association; SVM, support vector machine. Figure S3: Confusion matrices of XGBoost, SVM, and logistic regression models in the training (top) and testing (bottom) sets. NPV, negative predictive value; PPV, positive predictive value; SVM, support vector machine. Figure S4: Radar charts comparing six performance metrics across XGBoost, SVM, and logistic regression models. (a) Training set performance. (b) Testing set performance. SVM, support vector machine. Acknowledgments The authors have nothing to report. Data Availability Statement The data that support the findings of this study are available from the corresponding author upon reasonable request. References Associated Data Supplementary Materials Table S1: Conventional MR scanning sequences and parameters.TR, repetition time; TE, echo time; mDixon, dixon method; SAG, sagittal; TSE, turbo spin echo. Table S2: Performance comparison of different models. Model A contains 4 features: Age, Curvature, FA, and baseline mJOA; Model B contains 7 features: Age, Curvature, Cross‐Sectional Area, T2 hyperintensity, ADC, FA, and baseline mJOA; Model C contains 12 features: Gender, Age, BMI, Diabetes, Smoking, Duration, Curvature, Cross‐Sectional Area, T2 hyperintensity, ADC, FA, and baseline mJOA; Model D contains all 15 features: Gender, Age, BMI, Diabetes, Smoking, Duration, Surgical procedure, Location, Diseased segment, Curvature, Cross‐Sectional Area, T2 hyperintensity, ADC, FA, and baseline mJOA. Figure S1: Calibration curves for XGBoost, SVM, and logistic regression models in the training set (top row) and testing set (bottom row). SVM, support vector machine. Figure S2: Stratified performance of XGBoost, SVM, and logistic regression models across clinically relevant subgroups. (a) AUC by age groups (< 60 vs. ≥ 60 years). (b) AUC by baseline mJOA severity (mild ≥ 15 vs. severe < 15). (c) Accuracy by age groups. (d) Accuracy by baseline mJOA severity. AUC, area under the receiver operating characteristic curve; mJOA, modified Japanese Orthopedic Association; SVM, support vector machine. Figure S3: Confusion matrices of XGBoost, SVM, and logistic regression models in the training (top) and testing (bottom) sets. NPV, negative predictive value; PPV, positive predictive value; SVM, support vector machine. Figure S4: Radar charts comparing six performance metrics across XGBoost, SVM, and logistic regression models. (a) Training set performance. (b) Testing set performance. SVM, support vector machine. Data Availability Statement The data that support the findings of this study are available from the corresponding author upon reasonable request.