Abstract
Introduction
Axial pain is a common complication following expansive unilateral open-door laminoplasty (ELAP); however, traditional statistical methods are unable to effectively predict this complication. This study developed machine learning (ML) models to predict post-ELAP axial pain and identify key predictors.
Methods
This retrospective study enrolled 851 cervical spondylotic myelopathy (CSM) patients undergoing ELAP, split into training (n = 714) and temporal validation sets (n = 137). Demographic, imaging, clinical, and surgical data were collected. Predictive features were selected by the least absolute shrinkage and selection operator (Lasso) regression, followed by ML model development with grid search optimizing hyperparameters. The top-performing model underwent temporal validation, and SHapley Additive exPlanations (SHAP) analyzed predictor contributions.
Results
The training set included 218 axial pain cases; the test set had 47. Key predictors (C7 laminoplasty, cervical kyphosis, etc.) were identified to develop ML model. Post-optimization, extreme gradient boosting (XGBoost) achieved superior performance (internal validation area under the receiver [AUC] = 0.948; 95% confidence interval [CI] 0.918–0.978), maintained in temporal validation (AUC = 0.906; 95% CI 0.858–0.954). Through SHAP analysis, the predictors were ranked in descending order of importance as follows: C7 laminoplasty, quantity-based surgical segment classification, cervical kyphosis, angle of lamina open-door, cervical lordosis, and occupying rate of cervical spinal canal.
Conclusions
ML models coupled with SHAP analysis effectively predict post-ELAP axial pain, identifying the key predictors. Performing segment-selective ELAP, avoiding unnecessary C7 laminoplasty, and maintaining optimal open-door angle are critical factors in avoiding the occurrence of axial pain following ELAP.
Supplementary Information
The online version contains supplementary material available at 10.1007/s40122-025-00802-x.
Keywords: Machine learning, Axial pain, Prognostic prediction, Cervical spondylotic myelopathy, Expansive unilateral open-door laminoplasty
Key Summary Points
| Why carry out this study? |
| Axial pain is a common complication following expansive unilateral open-door laminoplasty (ELAP) for cervical spondylotic myelopathy (CSM), yet conventional statistical methods fail to predict its occurrence effectively. |
| Identifying modifiable surgical factors to mitigate post-ELAP axial pain remains a critical unmet. |
| Can ML models predict post-ELAP axial pain with high accuracy and identify actionable predictors to optimize surgical planning? |
| What was learned from this study? |
| The extreme gradient boosting (XGBoost) model demonstrated exceptional predictive performance (area under the receiver [AUC] 0.948 in internal validation; 0.906 in temporal validation), with C7 laminoplasty, multilevel surgery, cervical kyphosis, lamina open-door angle, cervical lordosis, and spinal canal occupancy rate as key predictors. |
| Segment-selective ELAP (avoiding C7 laminoplasty) and maintaining lamina open-door angles within the optimal range reduces axial pain risk, offering actionable guidance to surgeons. |
Introduction
Axial pain, characterized by persistent cervicalgia and refractory paravertebral rigidity, emerges as predominant morbidity subsequent to conventional expansive unilateral open-door laminoplasty (ELAP) [1]. As a type of chronic pain, the exact mechanisms underlying axial pain remain unclear, and no reliable methods currently exist for its effective prediction. Furthermore, axial pain profoundly impacts postoperative quality of life in many patients undergoing ELAP, which may further compromise physical health and impose significant strain on healthcare systems. These findings underscore the critical need to identify axial pain risk factors for developing targeted prevention and mitigation strategies.
Cervical spondylotic myelopathy (CSM) is characterized by cervical spinal stenosis resulting from cervical degeneration [2], often requiring surgical intervention. As a posterior surgical approach, ELAP is widely employed for cervical spinal canal decompression in the management of CSM. However, the ELAP procedure is frequently associated with various complications, including axial pain, which may compromise postoperative quality of life in patients [3, 4]. Currently, no methods exist for reliably predicting post-ELAP axial pain.
Conventional methods for assessing and predicting surgical outcomes often rely on clinician experience and limited parameters, which may compromise accuracy and specificity in risk stratification. Recent advances have introduced computational tools, such as machine learning (ML), to improve prognostic modeling [5, 6]. ML algorithms excel at analyzing large datasets and identifying complex, non-linear relationships between variables, offering enhanced predictive capabilities for clinical outcomes [7–9]. This study aimed to develop an ML model with optimal performance to predict axial pain at 1 year following ELAP, while identifying key risk factors, thereby providing actionable insights for optimizing surgical approach selection and refining techniques to reduce axial pain incidence.
Methods
Patients and Study Design
The retrospective study received approval from the Ethics Committee in Clinical Research of the First Affiliated Hospital of Wenzhou Medical University (KY2024-R286) and included patients diagnosed with CSM who underwent ELAP between April 2016 and September 2023. All participants or their guardians agreed to the study and signed the informed consent forms. All ELAP surgeries were performed by the same experienced spinal surgeon, who has 20 years of experience in ELAP procedures.
Inclusion criteria were as follows: (1) MRI-confirmed CSM; (2) with complete clinical/imaging data; (3) underwent cervical ELAP; (4) ≥ 1-year follow-up.
Exclusion criteria included (1) with preoperative axial pain; (2) concurrent other cervical disorders, such as cervical deformity and trauma-induced cervical spinal cord injury; (3) with neurological disorders that influence outcome assessment.
Ultimately, a total of 851 patients were enrolled in this study. The entire cohort was split into a training set containing 714 cases (from April 2016 to April 2022) and a test set with 137 cases for temporal validation (from May 2022 to September 2023).
Ethical Approval
All participants or their guardians agreed to the study and signed the informed consent forms. Ethics approval was obtained from the Ethics Committee in Clinical Research of the First Affiliated Hospital of Wenzhou Medical University (KY2024-R286). This study was performed in accordance with the Helsinki Declaration of 1964 and its later amendments.
Population Baseline Data
Population characteristics, including age, sex, weight, height, body mass index (BMI), smoking, drinking, hypertension, and diabetes, were collected for analysis.
Clinical Characteristics
Clinical characteristics were assessed through validated metrics: preoperative Japanese Orthopaedic Association (JOA) and visual analogue scale (VAS) scores. The primary outcome was axial pain after ELAP at 1-year follow-up, which was evaluated by the criteria as below: (1) patients with nuchal pain (posterior neck) and/or shoulder pain (including pain of the suspensory muscles) (VAS ≥ 4), and/or shoulder muscle spasm (stiffness and tension of the suspensory muscles); (2) pain with minimal motion (VAS ≥ 4); (4) no tenderness with palpation (VAS < 4) [10–12]. These criteria were also utilized for the exclusion criteria application. Clinical characteristics were independently assessed by two spine surgeons (both with over 10 years of clinical experience) who underwent standardized assessment training and were blinded to study data. Intra-rater and inter-rater reliability results for clinical-characteristic assessments are shown in Table S1.
The surgical-related data, including quantity-based surgical segment classification (classified as class 1: C3-7; class 2: C3-6 and C4-7; class 3: C3-5, C4-6, and C5-7; and class 4: C5-6 based on the number of laminoplasty levels performed), C6 laminoplasty, C7 laminoplasty, open-door side, angle of lamina open-door (at C5, measured as shown in Fig. 1A, B), operative time, and intraoperative blood loss, were collected.
Fig. 1.
Schematic illustrations of cervical radiographic measurements. A, B The preoperative (A) and postoperative (B) lamina angles of surgical segments were measured utilizing CT transverse section scanning at C5. The lamina angle was defined as the angle formed by the intersection of two lines: one extending along the posterior margin of the cervical vertebra and the other one connecting with the inner border of the lamina in the hinge side. The angle of lamina open-door = postoperative lamina angle (b)—preoperative lamina angle (a), with the average value being utilized for statistical analysis. C A schematic illustration of preoperative lateral cervical radiographic measurements. Based on a neutral lateral radiograph, TIA is defined as the angle between a line perpendicular to the midpoint of the T1 superior endplate and a line connecting the end of the sternum and the midpoint of the upper endplate of the T1 vertebra. NT represents the angle formed by a line connecting the upper end of the sternum and the center of the superior endplate of T1 and a vertical line drawn from the upper end of the sternum. T1S represents the angle between a horizontal line and the superior endplate of T1. C2-7 SVA was measured as the distance from the vertical line from the center of the C2 body and the posterosuperior corner of C7. SCA is defined between the C7 slope and the straight line joining the middle of the C7 inferior endplate and the middle of the sella turcica. C2-7 angle represents the angle between the inferior endplates of C2 and C7 (as shown in this figure, the C2-7 angle demonstrates a positive value), and a positive C2-7 angle corresponds to CL, while patients with a negative C2-7 angle are described as having cervical kyphosis. Additionally, cervical ROM is defined as the sum of the CL at maximum flexion (α) and maximum extension (β) on the lateral radiographs. Cervical fROM represents the change of CL between the neutral and flexion positions and eROM represents the change of CL between the neutral and extension positions. TIA thoracic inlet angle, NT neck tilt, T1S T1 slope, SVA sagittal vertical axis, SCA spine cranial angle, CL cervical lordosis, ROM range of motion, eROM extension ROM, fROM flexion ROM
Preoperative Radiographic Characteristics
Preoperative radiographic characteristics including cervical ROM, cervical eROM (extension ROM), cervical fROM (flexion ROM), TIA (thoracic inlet angle), NT (neck tilt), T1S (T1 slope), C2-7 SVA (sagittal vertical axis), SCA (spine cranial angle), CL (cervical lordosis), cervical instability, cervical kyphosis, occupying rate of cervical spinal canal, and ONL (ossification of the nuchal ligament) were independently measured by two spine surgeons (both with over 10 years of clinical experience), who completed a standardized 12-h training program (50 clinical cases) for radiological feature evaluation and were blinded to study data. Intra-rater and inter-rater reliability results for radiological measurements were detailed in Table S1.
As shown in Fig. 1C, based on a neutral lateral radiograph, TIA, NT, T1S, C2-7 SVA, SCA, and C2-7 angle were measured. Based on the flexion and extension lateral radiographs, cervical ROM, cervical eROM, and cervical fROM were measured. In this study, a positive C2-7 angle corresponded to CL, while patients with a negative C2-7 angle were described as having cervical kyphosis [13]. In addition, using neutral or flexion–extension lateral cervical radiographs for measurements, cervical instability was defined as (1) horizontal displacement exceeding 3.5 mm between the posterior margins of adjacent vertebrae and/or (2) rotation displacement exceeding 11 degrees between the inferior endplate of either adjacent vertebra [14].
Feature Selection
To alleviate the potential bias and multicollinearity, the least absolute shrinkage and selection operator (Lasso) regression was employed to select features [7]. Features were selected based on the λ.1-s.e. (standard error) criterion [15], and λ serves as the penalty parameter, determining the degree to which the function is shrunk. The Lasso regression coefficient of each variable is shown in Table S2.
Model Training and Testing
The training cohort was randomly divided into a Dataset 1 (for training) and a Dataset 2 (for validation) with a ratio of 4:1. The model was established using eight ML algorithms below: LR (logistic regression), XGBoost (extreme gradient boost), LightGBM (light gradient boosting machine), RF (random forest), AdaBoost (adaptive boosting), GBDT (gradient boosting decision tree), SVM (support vector machine) with the RBF (radial basic function) kernel, and KNN (K nearest neighbor). Based on the area under the receiver operating characteristics curve (AUC-ROC), the optimal hyperparameters were detected by grid search to improve the predictive performance of these models [7, 16]. The default and optimized hyperparameters for each model are shown in Table S3.
The Dataset 2 (validation data) was used to evaluate predictive power of each model. These assessments were performed using various metrics, including AUC-ROC, accuracy, sensitivity, specificity, negative predictive value (NPV), positive predictive value (PPV), and F1-score. The discrimination was assessed using AUC-ROC for binary classification [17]. The calibration of the ML models was evaluated by calibration plots. The Brier score was recognized as a metric for assessing the models’ discrimination and calibration [18]. Additionally, the net benefit of each model was analyzed using decision curve [19]. Following temporal validation of the optimally performing model (defined by maximal AUC-ROC and minimal Brier score [20]), predictor contributions were ranked and visualized through SHapley Additive exPlanations (SHAP) analysis [21].
Statistical Analysis
In the study, mean ± SD was used to represent continuous variables and frequency [proportion (%)] was used to represent categorical variables. Intergroup comparisons were employed using unpaired t tests, chi-square tests, and Fisher’s exact test. Comparisons with values of P < 0.05 were considered statistically significant. Lasso regression was conducted using R programming language (4.2.3). LR, RF, AdaBoost, GBDT, SVM, and KNN were trained using Python (scikit-learn = 1.1.3). XGBoost, LightGBM, and SHAP analysis were respectively trained using Python (xgboost = 2.0.1), Python (lightgbm = 3.2.1), and Python (shap = 0.43.0).
Result
Demographic Data and Clinical Characteristics
As delineated in Table 1, demographic baseline data and clinical characteristics demonstrated no statistically significant disparities between the training cohort and temporal validation cohort. Table 2 presents comparative analyses of demographic and clinical parameters between axial pain group and non-axial pain groups within the training cohort. The axial pain group exhibited: (1) significantly smaller CL (7.55 ± 17.67° vs. 13.19 ± 12.37°, P < 0.001); (2) higher incidence of cervical kyphotic deformity (27.06% vs. 9.88%, P < 0.001); (3) increased occupying rate of cervical spinal canal (0.55 ± 0.10 vs. 0.50 ± 0.11, P < 0.001); (4) greater proportion of multilevel laminoplasty [≥ 4 segments (including class 1 and class 2), 89.45% vs. 38.31%]; (5) higher frequency of C7 laminoplasty (76.15% vs. 26.01%); and (6) smaller laminoplasty opening angles (17.77 ± 5.16° vs. 22.25 ± 5.66°).
Table 1.
Demographics and clinical characteristics of study in the training and temporal validation cohorts
| Variable | Mean ± SD or frequency [proportion (%)] | P | |
|---|---|---|---|
| Training data (n = 714) | Temporal validation data (n = 137) | ||
| Demographic data | |||
| Age (years) | 58.475 ± 10.203 | 57.891 ± 10.028 | 0.539a |
| Sex | 0.869b | ||
| Male | 511 (71.569) | 99 (72.263) | |
| Female | 203 (28.431) | 38 (27.737) | |
| Weight (kg) | 65.133 ± 9.072 | 65.029 ± 9.676 | 0.904a |
| Height (cm) | 164.039 ± 7.479 | 164.577 ± 7.365 | 0.441a |
| BMI (kg/m2) | 24.159 ± 2.583 | 23.921 ± 2.578 | 0.325a |
| Smoking | 0.165b | ||
| Yes | 121 (16.947) | 30 (21.898) | |
| No | 593 (83.053) | 107 (78.102) | |
| Drinking | 0.712b | ||
| Yes | 140 (19.608) | 25 (18.248) | |
| No | 574 (80.392) | 112 (81.752) | |
| Hypertension | 0.178b | ||
| Yes | 165 (23.109) | 39 (28.467) | |
| No | 549 (76.891) | 98 (71.533) | |
| Diabetes | 0.844b | ||
| Yes | 64 (8.964) | 13 (9.489) | |
| No | 650 (91.036) | 124 (90.511) | |
| Preoperative imaging measurements | |||
| Cervical ROM (°) | 40.198 ± 10.981 | 41.525 ± 11.301 | 0.198a |
| Cervical eROM (°) | 13.642 ± 8.300 | 14.592 ± 10.239 | 0.309a |
| Cervical fROM (°) | 26.560 ± 9.347 | 26.943 ± 8.801 | 0.658a |
| TIA (°) | 77.703 ± 9.611 | 78.710 ± 9.462 | 0.261a |
| NT (°) | 49.693 ± 7.367 | 50.788 ± 8.230 | 0.150a |
| T1S (°) | 28.009 ± 8.329 | 27.920 ± 8.614 | 0.910a |
| C2-7 SVA (mm) | 20.207 ± 10.660 | 20.901 ± 8.900 | 0.421a |
| SCA (°) | 75.757 ± 7.773 | 76.946 ± 8.721 | 0.109a |
| CL (°) | 11.470 ± 14.433 | 11.406 ± 14.239 | 0.962a |
| Cervical instability | 0.901b | ||
| Yes | 184 (25.770) | 36 (26.277) | |
| No | 530 (74.230) | 101 (73.723) | |
| Cervical kyphosis | 0.952b | ||
| Yes | 108 (15.126) | 21 (15.328) | |
| No | 606 (84.874) | 116 (84.672) | |
| Occupying rate of cervical spinal canal | 0.516 ± 0.107 | 0.525 ± 0.107 | 0.383a |
| ONL | 0.485b | ||
| Yes | 203 (28.431) | 43 (31.387) | |
| No | 511 (71.569) | 94 (68.613) | |
| Surgical information | |||
| Quantity-based surgical segment classification | 0.300c | ||
| Class 1 | 12 (1.681) | 1 (0.730) | |
| Class 2 | 373 (52.241) | 83 (60.584) | |
| Class 3 | 327 (45.798) | 53 (38.686) | |
| Class 4 | 2 (0.280) | 0 (0.000) | |
| C6 laminoplasty | 0.630b | ||
| Yes | 653 (91.457) | 127 (92.701) | |
| No | 61 (8.543) | 10 (7.299) | |
| C7 laminoplasty | 0.486b | ||
| Yes | 295 (41.317) | 61 (44.526) | |
| No | 419 (58.683) | 76 (55.474) | |
| Open-door side | 0.719b | ||
| Right | 653 (91.457) | 124 (90.511) | |
| Left | 61 (8.543) | 13 (9.489) | |
| Angle of lamina open-door (°) | 20.881 ± 5.885 | 20.867 ± 6.055 | 0.980a |
| Operative time (min) | 118.326 ± 26.165 | 115.482 ± 21.551 | 0.175a |
| Intraoperative blood loss (ml) | 51.024 ± 36.273 | 48.657 ± 34.837 | 0.482a |
| Preoperative clinical characteristics | |||
| JOA score | 13.416 ± 1.690 | 13.613 ± 1.693 | 0.212a |
| VAS score | 0.947 ± 1.467 | 0.964 ± 1.645 | 0.905a |
| Primary outcome | |||
| Axial pain | 0.382b | ||
| Yes | 218 (30.532) | 47 (34.307) | |
| No | 496 (69.468) | 90 (65.693) | |
Mean ± SD is used to represent continuous variables and frequency [proportion (%)] is used to represent binary categorical variables. The statistical methods indicated by the superscripted letters were as follows: a unpaired t test, b chi-square tests, c Fisher’s exact test. SD standard deviation, BMI body mass index, ROM range of motion, eROM extension ROM, fROM flexion ROM, TIA thoracic inlet angle, NT neck tilt, T1S T1 slope, SVA sagittal vertical axis, SCA spine cranial angle, CL cervical lordosis, ONL ossification of the nuchal ligament, JOA Japanese Orthopaedic Association, VAS visual analogue scale
Table 2.
Patients’ demographic characteristics in the training cohort, n = 714
| Variable | Mean ± SD or frequency [proportion (%)] | P | |
|---|---|---|---|
| Axial pain group (n = 218) | Non-axial pain group (n = 496) | ||
| Demographic data | |||
| Age (years) | 58.133 ± 9.547 | 58.625 ± 10.475 | 0.554a |
| Sex | 0.370b | ||
| Male | 161 (73.853) | 350 (70.565) | |
| Female | 57 (26.147) | 146 (29.435) | |
| Weight (kg) | 64.998 ± 9.866 | 65.191 ± 8.699 | 0.794a |
| Height (cm) | 163.668 ± 7.115 | 164.203 ± 7.628 | 0.379a |
| BMI (kg/m2) | 24.185 ± 2.728 | 24.147 ± 2.516 | 0.859a |
| Smoking | 0.990b | ||
| Yes | 37 (16.972) | 84 (16.935) | |
| No | 181 (83.028) | 412 (83.065) | |
| Drinking | 0.331b | ||
| Yes | 38 (17.431) | 102 (20.565) | |
| No | 180 (82.569) | 394 (79.435) | |
| Hypertension | 0.905b | ||
| Yes | 51 (23.394) | 114 (22.984) | |
| No | 167 (76.606) | 382 (77.016) | |
| Diabetes | 0.878b | ||
| Yes | 19 (8.716) | 45 (9.073) | |
| No | 199 (91.284) | 451 (90.927) | |
| Preoperative imaging measurements | |||
| Cervical ROM (°) | 40.430 ± 9.851 | 40.097 ± 11.441 | 0.693a |
| Cervical eROM (°) | 13.364 ± 6.728 | 13.764 ± 8.901 | 0.510a |
| Cervical fROM (°) | 27.067 ± 7.490 | 26.337 ± 10.047 | 0.283a |
| TIA (°) | 77.741 ± 8.811 | 77.686 ± 9.943 | 0.941a |
| NT (°) | 50.081 ± 6.406 | 49.522 ± 7.746 | 0.352a |
| T1S (°) | 27.660 ± 7.679 | 28.162 ± 8.595 | 0.459a |
| C2-7 SVA (mm) | 20.813 ± 9.165 | 19.940 ± 11.245 | 0.277a |
| SCA (°) | 75.677 ± 6.023 | 75.792 ± 8.427 | 0.837a |
| CL (°) | 7.551 ± 17.669 | 13.192 ± 12.366 | < 0.001a |
| Cervical instability | 0.437b | ||
| Yes | 52 (23.853) | 132 (26.613) | |
| No | 166 (76.147) | 364 (73.387) | |
| Cervical kyphosis | < 0.001b | ||
| Yes | 59 (27.064) | 49 (9.879) | |
| No | 159 (72.936) | 447 (90.121) | |
| Occupying rate of cervical spinal canal | 0.550 ± 0.096 | 0.501 ± 0.108 | < 0.001a |
| ONL | 0.473b | ||
| Yes | 58 (26.606) | 145 (29.234) | |
| No | 160 (73.394) | 351 (70.766) | |
| Surgical information | |||
| Quantity-based surgical segment classification | < 0.001c | ||
| Class 1 | 11 (5.046) | 1 (0.202) | |
| Class 2 | 184 (84.404) | 189 (38.105) | |
| Class 3 | 23 (10.550) | 304 (61.290) | |
| Class 4 | 0 (0.000) | 2 (0.403) | |
| C6 laminoplasty | 0.637b | ||
| Yes | 201 (92.202) | 452 (91.129) | |
| No | 17 (7.798) | 44 (8.871) | |
| C7 laminoplasty | < 0.001b | ||
| Yes | 166 (76.147) | 129 (26.008) | |
| No | 52 (23.853) | 367 (73.992) | |
| Open-door side | 0.118b | ||
| Right | 194 (88.991) | 459 (92.540) | |
| Left | 24 (11.009) | 37 (7.460) | |
| Angle of lamina open-door (°) | 17.766 ± 5.155 | 22.250 ± 5.660 | < 0.001a |
| Operative time (min) | 117.060 ± 23.428 | 118.883 ± 27.262 | 0.364a |
| Intraoperative blood loss (ml) | 51.248 ± 38.403 | 50.925 ± 35.296 | 0.916a |
| Preoperative clinical characteristics | |||
| JOA score | 13.477 ± 1.863 | 13.389 ± 1.608 | 0.523a |
| VAS score | 0.968 ± 1.363 | 0.938 ± 1.510 | 0.799a |
Mean ± SD is used to represent continuous variables and frequency [proportion (%)] is used to represent binary categorical variables. The statistical methods indicated by the superscripted letters were as follows: a Unpaired t test, b chi-square tests, c Fisher’s exact test. SD standard deviation, BMI body mass index, ROM range of motion, eROM extension ROM, fROM flexion ROM, TIA thoracic inlet angle, NT neck tilt, T1S T1 slope, SVA sagittal vertical axis, SCA spine cranial angle, CL cervical lordosis, ONL ossification of the nuchal ligament, JOA Japanese Orthopaedic Association, VAS visual analogue scale
Baseline Predictive Performance
Based on the λ.1-s.e. criterion, six features were selected using Lasso regression (Fig. 2 and Table S2) for the subsequent model developing. The resulting features encompassed C7 laminoplasty, quantity-based surgical segment classification, cervical kyphosis, the occupying rate of cervical spinal canal, angle of lamina open-door, and CL.
Fig. 2.
Feature selection was conducted by the Lasso regression. A The coefficient profiles of the 32 features were analyzed by Lasso regression. B Based on the λ.1-s.e. criterion, there were six features selected using Lasso regression in the study. The two dotted vertical lines were drawn at the optimal scores by minimum criterion (λ.min) and λ.1-s.e. criterion, respectively. Based on the λ.1-s.e. criterion, the selected features included C7 laminoplasty, quantity-based surgical segment classification, cervical kyphosis, occupying rate of cervical spinal canal, angle of lamina open-door, and CL. Lasso the least absolute shrinkage and selection operator, s.e. standard error, CL cervical lordosis
The default hyperparameters for each model are shown in Table S3. The comparative analysis of the baseline predictive performance of ML models subjected to tenfold cross-validation is shown in Figure S1 and Table S4. As shown in Fig. 3 and Table 3, the XGBoost model demonstrated the superior baseline predictive performance, achieving an AUC-ROC [95% confidence interval (CI)] of 0.910 (0.853–0.967), an accuracy of 0.824, a sensitivity of 0.754, a specificity of 0.915, a PPV of 0.920, a NPV of 0.741, an F1-score of 0.829, and a Brier score of 0.135. Furthermore, the calibration curve of the XGBoost model was fitting the perfect situation (Fig. S2). The decision curves indicated that LR, XGBoost, LightGBM, RF, GBDT, SVM, and KNN models performed well in clinical usefulness (Fig. S3).
Fig. 3.
The baseline predictive performance for ML models in validation data. A The ROC curves for ML models B The AUCs for ML models. ML machine learning, ROC receiver operating characteristics, AUC the area under curve, LR logistic regression, XGBoost extreme gradient boost, LightGBM light gradient boosting machine, RF random forest, AdaBoost adaptive boosting, GBDT gradient boosting decision tree, SVM support vector machine, KNN K nearest neighbor
Table 3.
Comparison of Brier score, AUC-ROC, accuracy, sensitivity, specificity, PPV, NPV, and F1-score of baseline prediction performance for ML models in validation data
| Model | Brier score | AUC-ROC (95% CI) | Accuracy | Sensitivity | Specificity | PPV | NPV | F1-score |
|---|---|---|---|---|---|---|---|---|
| LR | 0.160 | 0.876 (0.829–0.924) | 0.810 | 0.743 | 0.879 | 0.862 | 0.770 | 0.798 |
| XGBoost | 0.135 | 0.910 (0.853–0.967) | 0.824 | 0.754 | 0.915 | 0.920 | 0.741 | 0.829 |
| LightGBM | 0.345 | 0.795 (0.698–0.893) | 0.806 | 0.803 | 0.809 | 0.845 | 0.760 | 0.824 |
| RF | 0.146 | 0.838 (0.734–0.942) | 0.836 | 0.529 | 0.940 | 0.750 | 0.855 | 0.621 |
| AdaBoost | 0.230 | 0.899 (0.836–0.962) | 0.852 | 0.869 | 0.830 | 0.869 | 0.830 | 0.869 |
| GBDT | 0.189 | 0.882 (0.816–0.948) | 0.778 | 0.738 | 0.830 | 0.849 | 0.709 | 0.789 |
| SVM | 0.148 | 0.864 (0.791–0.937) | 0.778 | 0.852 | 0.681 | 0.776 | 0.780 | 0.812 |
| KNN | 0.181 | 0.778 (0.686–0.869) | 0.689 | 0.564 | 0.784 | 0.667 | 0.702 | 0.611 |
AUC-ROC the area under the receiver operating characteristics curve, CI confidence interval, NPV negative predictive value, PPV positive predictive value, LR logistic regression, XGBoost extreme gradient boost, LightGBM light gradient boosting machine, RF random forest, AdaBoost adaptive boosting, GBDT gradient boosting decision tree, SVM support vector machine, KNN K nearest neighbor
The Performance of ML Models with Optimal Hyperparameters
The hyperparameters of each model were optimized by grid search (Table S3). The comparative analysis of the hyperparameter-optimized ML predictive performance subjected to tenfold cross-validation is shown in Figure S4 and Table S5. Interestingly, following hyperparameter optimization, the XGBoost model maintained its status as the top-performing model, with an AUC-ROC (95% CI) of 0.948 (0.918–0.978), an accuracy of 0.884, a sensitivity of 0.887, a specificity of 0.881, a PPV of 0.860, a NPV of 0.904, an F1-score of 0.873, and a Brier score of 0.083 (Fig. 4 and Table 4). The calibration curve of optimized XGBoost model was very close to the ideal curve (Figure S5). As shown in Figure S6, hyperparameter-optimized LR, XGBoost, LightGBM, RF, GBDT, SVM, and KNN models possessed excellent clinical benefit.
Fig. 4.
The predictive efficiency of ML models with the optimized hyperparameters using grid search in validation data. A The ROC curves for hyperparameter-optimizing ML models. B The AUCs for hyperparameter-optimizing ML models. ML machine learning, ROC receiver operating characteristics, AUC the area under curve, LR logistic regression, XGBoost extreme gradient boost, LightGBM light gradient boosting machine, RF random forest, AdaBoost adaptive boosting, GBDT gradient boosting decision tree, SVM support vector machine, KNN K nearest neighbor
Table 4.
Comparison of the Brier score, AUC-ROC, accuracy, sensitivity, specificity, PPV, NPV, and F1-score of hyperparameter-optimized ML models in validation data
| Model | Brier score | AUC-ROC (95% CI) | Accuracy | Sensitivity | Specificity | PPV | NPV | F1-score |
|---|---|---|---|---|---|---|---|---|
| LR | 0.127 | 0.907 (0.866–0.948) | 0.830 | 0.921 | 0.737 | 0.782 | 0.901 | 0.845 |
| XGBoost | 0.083 | 0.948 (0.918–0.978) | 0.884 | 0.887 | 0.881 | 0.860 | 0.904 | 0.873 |
| LightGBM | 0.128 | 0.902 (0.841–0.962) | 0.824 | 0.803 | 0.851 | 0.875 | 0.769 | 0.838 |
| RF | 0.111 | 0.928 (0.880–0.975) | 0.833 | 0.846 | 0.821 | 0.815 | 0.852 | 0.830 |
| AdaBoost | 0.230 | 0.923 (0.868–0.978) | 0.870 | 0.885 | 0.851 | 0.885 | 0.851 | 0.885 |
| GBDT | 0.114 | 0.927 (0.877–0.978) | 0.833 | 0.787 | 0.894 | 0.906 | 0.764 | 0.842 |
| SVM | 0.106 | 0.923 (0.873–0.973) | 0.843 | 0.836 | 0.851 | 0.879 | 0.800 | 0.857 |
| KNN | 0.137 | 0.859 (0.789–0.929) | 0.723 | 0.744 | 0.709 | 0.644 | 0.796 | 0.690 |
AUC-ROC the area under the receiver operating characteristics curve, CI confidence interval, NPV negative predictive value, PPV positive predictive value, LR logistic regression, XGBoost extreme gradient boost, LightGBM light gradient boosting machine, RF random forest, AdaBoost adaptive boosting, GBDT gradient boosting decision tree, SVM support vector machine, KNN K nearest neighbor
The performance of the hyperparameter-optimized XGBoost model, trained as described, remained stable in the temporal validation set (AUC: 0.906, 95% CI (0.858–0.954)) (Fig. 5 and Table 5). Moreover, in the temporal validation set, the calibration curve of the hyperparameter-optimized XGBoost model also showed strong alignment with the ideal curve (Figure S7) and the model still possessed excellent clinical benefit (Fig. S8).
Fig. 5.
The ROC curve of the hyperparameter-optimized XGBoost model for predicting axial pain in the temporal validation set. ROC receiver operating characteristics, AUC the area under curve, XGBoost extreme gradient boost, CI confidence interval
Table 5.
Performance of hyperparameter-optimized XGBoost model for predicting axial pain in the temporal validation set
| Brier score | AUC-ROC (95% CI) | Accuracy | Sensitivity | Specificity | PPV | NPV | F1-score |
|---|---|---|---|---|---|---|---|
| 0.126 | 0.906 (0.858–0.954) | 0.803 | 0.766 | 0.822 | 0.692 | 0.871 | 0.727 |
AUC-ROC the area under the receiver operating characteristics curve, CI confidence interval, NPV negative predictive value, PPV positive predictive value, XGBoost extreme gradient boost
In addition, based on the hyperparameter-optimized XGBoost model, which exhibited the highest predictive power, we visualized the importance ranking of selected features using SHAP analysis. The results are shown in descending order as follows: C7 laminoplasty, quantity-based surgical segment classification, cervical kyphosis, angle of lamina open-door, CL, and occupying rate of cervical spinal canal (Fig. 6). In other words, patients with CSM who presented cervical kyphosis, high spinal canal occupancy rate, and/or a small cervical lordosis preoperatively were more prone to developing axial pain following ELAP procedure. Moreover, when ELAP procedure involved an excessively small angle of lamina open-door (measured at C5) and/or multilevel laminoplasty, particularly when C7 laminoplasty was included, the risk of developing axial pain 1 year postoperatively was further elevated.
Fig. 6.
A SHAP analysis of the hyperparameter-optimized XGBoost model. A Visualization for SHAP values of the Lasso-regression-selected features in each sample. The red dots represent the higher SHAP values, while the blue dots represent the lower values. B The selected features are listed in descending order by the average absolute values of SHAP. SHAP SHapley Additive exPlanations, CL cervical lordosis
Discussion
In recent years, researchers have focused on developing clinical prediction models using linear regression and nomograms to predict the prognosis of CSM [22–25]. Numerous studies indicated that ML algorithms exhibited significant advantages in predicting the prognosis of CSM, compared with conventional statistical methods [7, 9]. However, previous studies failed to implement variable selection prior to model development, potentially leading to overfitting and multicollinearity issues. LASSO regression effectively addresses these limitations. In this study, we employed LASSO regression with the λ.1-standard error (λ.1-s.e.) criterion for feature selection, followed by predictive model construction using the identified variables [26, 27]. Our findings identify six significant predictors of postoperative frailty: C7 laminoplasty, quantity-based surgical segment classification, cervical kyphosis, occupying rate of cervical spinal canal, angle of lamina open-door, and CL. This model facilitates personalized postoperative rehabilitation protocols tailored to individual patients, providing evidence-based guidance to interprofessional rehabilitation teams to mitigate axial pain risk in ELAP populations.
Our results indicated that C7 laminoplasty-involved ELAP was identified as the most significant predictor of post-ELAP axial pain. Compared to segment-selective ELAP, conventional long-segment (C3-7) laminoplasty was considered a risk factor for both axial pain and C5 palsy following ELAP [4]. Notably, conventional long-segment laminoplasty typically involves a greater number of laminoplasty levels. Meanwhile, this study revealed that an increased number of laminoplasty levels would elevate the risk of postoperative axial pain. This phenomenon may be attributed to extended surgical segments causing more severe cervical paraspinal muscle injury, which has been demonstrated to correlate with the development of axial pain [28]. Therefore, in clinical practice, segment-selective ELAP should be prioritized whenever feasible to avoid unnecessary C7 laminoplasty and excessive laminoplasty levels. Moreover, our findings showed that an insufficient angle of lamina open-door during laminoplasty was susceptible to axial pain within the first postoperative year. Previous study showed that maintaining the lamina open-door angle between 15 and 30 degrees in ELAP for CSM patients could decrease the incidence of axial pain and C5 palsy postoperatively [29]. Moreover, an excessively large angle of lamina open-door was associated with increased risks of severe axial pain, C5 palsy, and hinge fracture [30, 31]. Moreover, consistent with previous researches [32, 33], cervical kyphosis and a high occupying rate of cervical spinal canal were both risk factors contributing to postoperative axial pain. A previous study demonstrated that cervical kyphosis exceeding 13 degrees was considered a predictor of poor postoperative outcomes following ELAP [33]. In addition, this study revealed that CSM patients with unduly small preoperative cervical lordosis were prone to axial pain following ELAP compared to those with normal alignment. A retrospective study showed that intraoperative anterior vertebral body height augmentation (through anterior cervical fusion) maintained within 2–5 mm could facilitate cervical lordosis preservation, thereby reducing postoperative axial pain incidence [34]. Therefore, CSM patients with diminished cervical lordosis or straightened cervical curvature may consider undergoing anterior cervical fusion or combined anterior–posterior approach surgery to preserve cervical lordosis, which could potentially reduce the incidence of postoperative axial pain. However, validation of this perspective will be conducted in a future study.
Utilizing the hyperparameter-optimized ML models combined with SHAP analysis developed in this research, surgeons can obtain a more intuitive understanding of the significance of key features. This, in turn, can facilitate the preoperative identification of high-risk individuals who are prone to developing postoperative axial pain. Early identification enables the implementation of personalized therapeutic strategies, including targeted preventive measures, patient-specific procedural modifications, and alternative surgical interventions such as cervical microendoscopic laminotomy. Compared with conventional ELAP, this technique demonstrated comparable clinical outcomes in CSM patients, with significantly milder postoperatively axial pain and improved subaxial cervical lordosis [1]. Notably, the current study did not evaluate paraspinal muscle status, which might influence postoperative axial pain outcomes [35, 36]. Future studies incorporating MRI-based paraspinal muscle metrics will be necessary to enhance our predictive model. Taken together, when performing surgery on a CSM patient requiring ELAP treatment who presents preoperative imaging characteristics including cervical kyphosis, high occupying rate of cervical spinal canal, and/or small cervical lordosis, special attention should be given to performing segment-selective laminoplasty, avoiding unnecessary C7 laminoplasty, and maintaining the laminoplasty open-door angle within the optimal range (15–30 degrees as reported). For patients with established axial pain, treatment strategies should be tailored based on specific pain levels.
However, our study had its limitations. First, this study utilized internal temporal validation for model evaluation due to limited access to multicenter data. While temporal validation is widely adopted for predictive model assessment [37–39], future multicenter studies remain imperative. Second, our research was based on a retrospective design, which inherently carries the risk of selection bias. Therefore, conducting a prospective study is imperative to address this concern. Third, although the current cohort satisfies the widely accepted criterion (ten events per parameter [40]) for ML development in terms of sample size and feature-to-event ratio, formal power calculations specific to multivariable predictive models were not conducted. Therefore, future prospective investigations incorporating both methodologically robust sample size justification and larger multicenter cohorts would provide critical validation of the current observations. Consistent with previous studies, this research also directly utilized the AUC of models to rank predictive performance [7, 16, 18]. However, readers should exercise caution when interpreting small AUC differences between ML models. In future investigations, we will employ larger cohorts and implement statistical tests to quantify predictive performance disparities across models. Given the “black-box” characteristic of ML algorithms, improving the model visualization has emerged as a concern that needs to be addressed promptly. Additionally, our study assessed axial pain only at 1 year after ELAP. To enhance understanding of symptom evolution, future studies should prioritize longitudinal assessments to better characterize the progression of axial pain. At last, while we identified key predictors of postoperative axial pain, the underlying mechanisms remain unclear.
Conclusions
Our findings indicated that the application of ML algorithms, particularly the hyperparameter-optimized XGBoost model, can effectively predict post-ELAP axial pain. Furthermore, the key predictors were identified through SHAP value analysis. Performing segment-selective ELAP, avoiding unnecessary C7 laminoplasty, and maintaining optimal open-door angle are critical factors in avoiding the occurrence of axial pain following ELAP.
Supplementary Information
Below is the link to the electronic supplementary material.
Acknowledgements
We thank Dr. Mengxian Jia (affiliated Hangzhou First People’s Hospital, Westlake University) for his kind help in the literature review in this study. No funding was received for this study.
Author Contributions
Conceptualization: Ziwei Fan, Honglin Teng, and Kelun Huang; Methodology: Kelun Huang, Sheng Li, Yile Dai, and Ziwei Fan; Formal analysis and investigation: Kelun Huang, Sheng Li, Yile Dai, Dingjun Xu, Guoliang Chen, and Yaozhi He; Writing—original draft preparation: Kelun Huang, Sheng Li, Yile Dai, and Dingjun Xu; Writing—review and editing: Kelun Huang, Ziwei Fan, and Honglin Teng; Resources: Honglin Teng; Supervision: Ziwei Fan, and Honglin Teng, and each author has approved the final version to be submitted.
Funding
No funding or sponsorship was received for this study or publication of this article. The Rapid Service Fee was funded by the authors.
Data Availability
All data generated or analyzed during this study are included in this published article.
Declarations
Conflict of interest
All authors (Kelun Huang, Sheng Li, Yile Dai, Dingjun Xu, Guoliang Chen, Yaozhi He, Ziwei Fan, and Honglin Teng) declare that they have no competing interests.
Ethical Approval
All participants or their guardians agreed to the study and signed the informed consent forms. Ethics approval was obtained from the Ethics Committee in Clinical Research of the First Affiliated Hospital of Wenzhou Medical University (KY2024-R286). This study was performed in accordance with the Helsinki Declaration of 1964 and its later amendments.
Footnotes
Kelun Huang, Sheng Li, and Yile Dai contributed equally to this work.
Ziwei Fan and Honglin Teng are the co-corresponding authors for this manuscript.
Contributor Information
Ziwei Fan, Email: fzw_formal@163.com.
Honglin Teng, Email: tenghonglin@wzhospital.cn.
References
- 1.Minamide A, Yoshida M, Simpson AK, et al. Microendoscopic laminotomy versus conventional laminoplasty for cervical spondylotic myelopathy: 5-year follow-up study. J Neurosurg Spine. 2017;27(4):403–9. 10.3171/2017.2.SPINE16939. [DOI] [PubMed] [Google Scholar]
- 2.McCormick JR, Sama AJ, Schiller NC, Butler AJ, Donnally CJ. Cervical spondylotic myelopathy: a guide to diagnosis and management. J Am Board Fam Med. 2020;33(2):303–13. 10.3122/jabfm.2020.02.190195. [DOI] [PubMed] [Google Scholar]
- 3.Otani K, Iwabuchi M, Sato K, Konno S, Kikuchi S. Postoperative neck symptoms of posterior approach for cervical compressive myelopathy: expansive open-door laminoplasty vs. segmental partial laminectomy. Fukushima J Med Sci. 2018;64(2):54–9. 10.5387/fms.2017-20. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Tsuji T, Asazuma T, Masuoka K, et al. Retrospective cohort study between selective and standard C3-7 laminoplasty. Minimum 2-year follow-up study. Eur Spine J. 2007;16(12):2072–7. 10.1007/s00586-007-0428-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Yan X, Huang K, Jia M, et al. Construction and verification of a nomogram predicting the risk of preoperative deep vein thrombosis progression after elective spine surgery. Clin Neurol Neurosurg. 2022;222:107439. 10.1016/j.clineuro.2022.107439. [DOI] [PubMed] [Google Scholar]
- 6.Jia M, Lai J, Li K, et al. Optimizing prediction accuracy for early recurrent lumbar disc herniation with a directional mutation-guided SVM model. Comput Biol Med. 2024;173:108297. 10.1016/j.compbiomed.2024.108297. [DOI] [PubMed] [Google Scholar]
- 7.Song J, Li J, Zhao R, Chu X. Developing predictive models for surgical outcomes in patients with degenerative cervical myelopathy: a comparison of statistical and machine learning approaches. Spine J. 2024;24(1):57–67. 10.1016/j.spinee.2023.07.021. [DOI] [PubMed] [Google Scholar]
- 8.Xie W, Li Y, Meng X, Zhao M. Machine learning prediction models and nomogram to predict the risk of in-hospital death for severe DKA: a clinical study based on MIMIC-IV, eICU databases, and a college hospital ICU. Int J Med Inform. 2023;174:105049. 10.1016/j.ijmedinf.2023.105049. [DOI] [PubMed] [Google Scholar]
- 9.Khan O, Badhiwala JH, Akbar MA, Fehlings MG. Prediction of worse functional status after surgery for degenerative cervical myelopathy: a machine learning approach. Neurosurgery. 2021;88(3):584–91. 10.1093/neuros/nyaa477. [DOI] [PubMed] [Google Scholar]
- 10.Hosono N, Yonenobu K, Ono K. Neck and shoulder pain after laminoplasty. A noticeable complication. Spine (Phila Pa 1976). 1996;21(17):1969–73. 10.1097/00007632-199609010-00005. [DOI] [PubMed] [Google Scholar]
- 11.Zhou J, Li L, Li T, Xue Y. Preoperative Modic changes are related to axial symptoms after anterior cervical discectomy and fusion. J Pain Res. 2018;11:2617–23. 10.2147/JPR.S172953. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Takeuchi K, Yokoyama T, Aburakawa S, et al. Axial symptoms after cervical laminoplasty with C3 laminectomy compared with conventional C3 C7 laminoplasty: a modified laminoplasty preserving the semispinalis cervicis inserted into axis. Spine (Phila Pa 1976). 2005;30(22):2544–9. 10.1097/01.brs.0000186332.66490.ba. [DOI] [PubMed] [Google Scholar]
- 13.Lin A, Skaggs DL, Andras LM, Tolo V, Tamrazi B, Illingworth KD. Increasing cervical kyphosis correlates with cervical degenerative disk disease in patients with adolescent idiopathic scoliosis. Spine (Phila Pa 1976). 2024;49(7):486–91. 10.1097/BRS.0000000000004824. [DOI] [PubMed] [Google Scholar]
- 14.White AA, Johnson RM, Panjabi MM, Southwick WO. Biomechanical analysis of clinical stability in the cervical spine. Clin Orthop Relat Res. 1975;109:85–96. 10.1097/00003086-197506000-00011. [DOI] [PubMed] [Google Scholar]
- 15.Zhou D, Liu X, Wang X, et al. A prognostic nomogram based on LASSO Cox regression in patients with alpha-fetoprotein-negative hepatocellular carcinoma following non-surgical therapy. BMC Cancer. 2021;21(1):246. 10.1186/s12885-021-07916-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Gao L, Cao Y, Cao X, et al. Machine learning-based algorithms to predict severe psychological distress among cancer patients with spinal metastatic disease. Spine J. 2023;23(9):1255–69. 10.1016/j.spinee.2023.05.009. [DOI] [PubMed] [Google Scholar]
- 17.Steyerberg EW, Vergouwe Y. Towards better clinical prediction models: seven steps for development and an ABCD for validation. Eur Heart J. 2014;35(29):1925–31. 10.1093/eurheartj/ehu207. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Karhade AV, Shah AA, Bono CM, et al. Development of machine learning algorithms for prediction of mortality in spinal epidural abscess. Spine J. 2019;19(12):1950–9. 10.1016/j.spinee.2019.06.024. [DOI] [PubMed] [Google Scholar]
- 19.Vickers AJ, Elkin EB. Decision curve analysis: a novel method for evaluating prediction models. Med Decis Making. 2006;26(6):565–74. 10.1177/0272989X06295361. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Dai Y, Huang W, Xu L, Zhang Q, Huang X. Machine learning models and nomogram based on clinical, laboratory profiles and skeletal muscle index to predict pancreatic fistula after pancreatoduodenectomy. Gland Surg. 2024;13(2):164–77. 10.21037/gs-23-410. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Wang K, Tian J, Zheng C, et al. Interpretable prediction of 3-year all-cause mortality in patients with heart failure caused by coronary heart disease based on machine learning and SHAP. Comput Biol Med. 2021;137:104813. 10.1016/j.compbiomed.2021.104813. [DOI] [PubMed] [Google Scholar]
- 22.Sun LQ, Li M, Li YM. Predictors for surgical outcome of laminoplasty for cervical spondylotic myelopathy. World Neurosurg. 2016;94:89–96. 10.1016/j.wneu.2016.06.092. [DOI] [PubMed] [Google Scholar]
- 23.Avadhani A, Rajasekaran S, Shetty AP. Comparison of prognostic value of different MRI classifications of signal intensity change in cervical spondylotic myelopathy. Spine J. 2010;10(6):475–85. 10.1016/j.spinee.2010.03.024. [DOI] [PubMed] [Google Scholar]
- 24.Wang J, Li H, Yang B. Predictive nomogram for clinical prognosis in cervical spondylotic myelopathy with intramedullary T2-weighted increased signal intensity: a novel digital tool for patient prognosis education. Front Public Health. 2022;10:898242. 10.3389/fpubh.2022.898242. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25.Chen G, Jia M, Connel RK, et al. Nomogram for predicting kyphotic deformity after laminoplasty in cervical spondylotic myelopathy patients without preoperative kyphotic alignment. Clin Neurol Neurosurg. 2020;199:106284. 10.1016/j.clineuro.2020.106284. [DOI] [PubMed] [Google Scholar]
- 26.Doebler P, Doebler A, Buczak P, Groll A. Interactions of scores derived from two groups of variables: alternating lasso regularization avoids overfitting and finds interpretable scores. Psychol Methods. 2023;28(2):422–37. 10.1037/met0000461. [DOI] [PubMed] [Google Scholar]
- 27.McNeish DM. Using Lasso for predictor selection and to assuage overfitting: a method long overlooked in behavioral sciences. Multivar Behav Res. 2015;50(5):471–84. 10.1080/00273171.2015.1036965. [DOI] [PubMed] [Google Scholar]
- 28.Mori E, Ueta T, Maeda T, Yugue I, Kawano O, Shiba K. Effect of preservation of the C-6 spinous process and its paraspinal muscular attachment on the prevention of postoperative axial neck pain in C3-6 laminoplasty. J Neurosurg Spine. 2015;22(3):221–9. 10.3171/2014.11.SPINE131153. [DOI] [PubMed] [Google Scholar]
- 29.Zhang H, Lu S, Sun T, Yadav SK. Effect of lamina open angles in expansion open-door laminoplasty on the clinical results in treating cervical spondylotic myelopathy. J Spinal Disord Tech. 2015;28(3):89–94. 10.1097/BSD.0b013e3182695295. [DOI] [PubMed] [Google Scholar]
- 30.Liu FJ, Ding XK, Chai Y, Qi SH, Li PF. Influence of fixed titanium plate position on the effectiveness of open-door laminoplasty for cervical spondylotic myelopathy. J Orthop Surg Res. 2022;17(1):297. 10.1186/s13018-022-03188-0. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31.Kobayashi Y, Matsumaru S, Kuramoto T, et al. Plate fixation of expansive open-door laminoplasty decreases the incidence of postoperative C5 palsy. Clin Spine Surg. 2019;32(4):E177–82. 10.1097/BSD.0000000000000790. [DOI] [PubMed] [Google Scholar]
- 32.Dong F, Shen C, Jiang S, et al. Measurement of volume-occupying rate of cervical spinal canal and its role in cervical spondylotic myelopathy. Eur Spine J. 2013;22(5):1152–7. 10.1007/s00586-012-2622-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 33.Suda K, Abumi K, Ito M, Shono Y, Kaneda K, Fujiya M. Local kyphosis reduces surgical outcomes of expansive open-door laminoplasty for cervical spondylotic myelopathy. Spine (Phila Pa 1976). 2003;28(12):1258–62. 10.1097/01.BRS.0000065487.82469.D9. [DOI] [PubMed] [Google Scholar]
- 34.Kawakami M, Tamaki T, Yoshida M, Hayashi N, Ando M, Yamada H. Axial symptoms and cervical alignments after cervical anterior spinal fusion for patients with cervical myelopathy. J Spinal Disord. 1999;12(1):50–6. [PubMed] [Google Scholar]
- 35.Eksi MS, Topcu A, Topaloglu F, et al. Fatty infiltration in the multifidus predicts screw-loosening following short-segment decompression and fusion: proof of why we should protect and rehabilitate the paraspinal muscles. Eur Spine J. 2025;34(6):2427–37. 10.1007/s00586-025-08793-1. [DOI] [PubMed] [Google Scholar]
- 36.Xu F, Zhou S, Sun Z, Jiang S, Han G, Li W. Relationship between the postoperative variations of paraspinal muscles and adjacent-segment degeneration in patients with degenerative lumbar spinal stenosis after posterior instrumented lumbar fusion. J Neurosurg Spine. 2024;40(5):551–61. 10.3171/2023.11.SPINE23750. [DOI] [PubMed] [Google Scholar]
- 37.Xu M, Zhu W, Hou S, et al. Development and multicenter validation of machine learning models for predicting postoperative pulmonary complications after neurosurgery. Chin Med J (Engl). 2025;138(17):2170–9. 10.1097/CM9.0000000000003433. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 38.Noda R, Ichikawa D, Shibagaki Y. Machine learning-based diagnostic prediction of IgA nephropathy: model development and validation study. Sci Rep. 2024;14(1):12426. 10.1038/s41598-024-63339-7. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 39.Heber S, Pereyra D, Schrottmaier WC, et al. A model predicting mortality of hospitalized Covid-19 patients four days after admission: development, internal and temporal-external validation. Front Cell Infect Microbiol. 2021;11:795026. 10.3389/fcimb.2021.795026. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 40.Riley RD, Ensor J, Snell KIE, et al. Calculating the sample size required for developing a clinical prediction model. BMJ. 2020;368:m441. 10.1136/bmj.m441. [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
All data generated or analyzed during this study are included in this published article.