Best Machine Learning Model for Predicting Axial Symptoms After Unilateral Laminoplasty: Based on C2 Spinous Process Muscle Radiomics Features and Sagittal Parameters

Abstract

Study Design

Retrospective study.

Objective

To develop a machine learning model for predicting axial symptoms (AS) after unilateral laminoplasty by integrating C2 spinous process muscle radiomics features and cervical sagittal parameters.

Methods

In this retrospective study of 96 cervical myelopathy patients (30 with AS, 66 without) who underwent unilateral laminoplasty between 2018-2022, we extracted radiomics features from preoperative MRI of C2 spinous muscles using PyRadiomics. Clinical data including C2-C7 Cobb angle, cervical sagittal vertical axis (cSVA), T1 slope (T1S) and C2 muscle fat infiltration are collected for clinical model construction. After LASSO regression feature selection, we constructed six machine learning models (SVM, KNN, Random Forest, ExtraTrees, XGBoost, and LightGBM) and evaluated their performance using ROC curves and AUC.

Results

The AS group demonstrated significantly lower preoperative C2-C7 Cobb angles (12.80° ± 7.49° vs 18.02° ± 8.59°, P = .006), higher cSVA (3.01 cm ± 0.87 vs 2.46 ± 1.19 cm, P = .026), T1S (26.68° ± 5.12° vs 23.66° ± 7.58°, P = .025) and higher C2 muscle fat infiltration (23.73 ± 7.78 vs 20.62 ± 6.93 P = .026). Key radiomics features included local binary pattern texture features and wavelet transform characteristics. The combined model integrating radiomics and clinical parameters achieved the best performance with test AUC of 0.881, sensitivity of 0.833, and specificity of 0.786.

Conclusion

The machine learning model based on C2 spinous process muscle radiomics features and clinical parameters (C2-C7 Cobb angle, cSVA, T1S and C2 muscle infiltration) effectively predicts AS occurrence after unilateral laminoplasty, providing clinicians with a valuable tool for preoperative risk assessment and personalized treatment planning.

Introduction

Unilateral laminoplasty is a common surgical method for treating multi-segment cervical spondylotic myelopathy (CSM).^1-3 However, the incidence of postoperative axial symptoms (AS) ranges from 24% to 32%, manifesting as persistent neck and shoulder pain, stiffness, and discomfort, which severely affects patients’ postoperative quality of life and satisfaction.^4-6

Previous studies indicate that the occurrence of AS after laminoplasty is related to multiple factors, including preoperative spinal canal anterior wall occupation ratio, intraoperative facet joint damage, C7 spinous process muscle attachment point damage, excessive open-door angle, postoperative loss of cervical curvature, and reduced mobility. ⁷ Notably, the integrity of C2 spinous process muscles plays an important role in preventing AS. Kato et al ⁸ demonstrate that preserving C2 spinous process muscle attachment points is an independent protective factor for reducing AS occurrence. Zuo et al find that a larger T1 slope is an independent risk factor for AS, ⁹ while Ruan et al ¹⁰ identify greater preoperative C2-C7 Cobb angle as a protective factor.

The C2 spinous process is an important attachment point for multiple posterior neck muscles, including the semispinalis cervicis, multifidus, and other deep neck extensor muscles, which together form the posterior tension band of the cervical spine, crucial for maintaining cervical lordosis and stability.^11,12 Previous studies show that preserving C2 spinous process muscle attachments significantly reduces the incidence of postoperative axial symptoms, with Kato et al confirming this as an independent protective factor. C2 spinous process muscles are closely related to cervical sagittal balance; dysfunction or damage to these muscles can lead to compensatory changes such as forward head posture and increased cSVA, increasing the risk of AS.^13,14

Radiomics is an emerging technology that extracts large numbers of quantitative features from medical images and transforms them into high-dimensional data for clinical decision-making. By extracting imaging features including shape, texture, and gray-level features, radiomics can reveal information undetectable to the naked eye, providing new perspectives for disease diagnosis and prognosis prediction. Combined with machine learning algorithms, radiomics has shown promising application prospects in multiple medical fields.^15,16

This study attempts for the first time to extract radiomics features from the C2 spinous process muscle level on preoperative MRI, combined with cervical sagittal parameters (C2-C7 Cobb angle, cSVA, and T1S) and C2 muscle infiltration, to establish a machine learning model for predicting AS occurrence after laminoplasty. It aims to provide clinical physicians with a more precise risk assessment tool to guide surgical strategy formulation and postoperative rehabilitation management.

Materials and Methods

The study flow is shown in Figure 1.

Figure 1.

Work Flow of Study

Study Population

This study retrospectively analyzes clinical data of cervical spondylotic myelopathy patients who underwent laminoplasty at our hospital between January 2018 and January 2022. Inclusion criteria ¹ : diagnosed with cervical spondylotic myelopathy and/or ossification of the posterior longitudinal ligament ² ; imaging examinations (cervical X-ray, CT, MRI) showing spinal cord compression ³ ; undergoing C3-C7 laminoplasty surgery ⁴ ; complete follow-up data. Exclusion criteria ¹ : cervical trauma, fracture, or tumor ² ; Prior cervical spine surgery ³ Inadequate imaging quality or missing key sagittal radiographs/axial MRI slices precluding radiomics analysis ⁴ ; postoperative loss to follow-up.

Finally, 96 patients (AS group: 30 cases, non-AS group: 66 cases) are enrolled according to inclusion and exclusion criteria. The entire cohort is then randomly divided into a training set (n = 76) and a test set (n = 20) at an 8:2 ratio.

Surgical Method

All surgeries are performed by the same team of experienced spine surgeons. After general anesthesia, patients are placed in prone position with a midline cervical posterior incision, exposing C3-C7 laminae and bilateral facet joints layer by layer. On the open-door side (more symptomatic side), a groove is created at the medial edge of the facet joints using a high-speed drill, preserving a thin layer of medial cortical bone. A similar groove is created on the hinge side, preserving intact medial cortical bone. The lamina on the open-door side is slowly pushed toward the hinge side to form a certain angle between the lamina and facet joint, and a micro titanium plate is placed to maintain spinal canal opening. Postoperative routine anti-infection, anti-inflammatory, and anti-swelling symptomatic treatments are administered. Patients wear cervical collars for 2 months and regularly follow up postoperatively.

Clinical Data and Sagittal Parameters

General clinical data including age, gender, disease duration, BMI, diabetes history, smoking history, Japanese Orthopaedic Association (JOA) score, and Visual Analogue Scale (VAS) are collected. Cervical sagittal parameters include ¹ :C2-C7 Cobb angle: defined as the angle between the C2 lower endplate and C7 upper endplate ² ; cervical sagittal vertical axis (cSVA): defined as the distance between the C2 midpoint plumb line and the posterosuperior corner of the C7 upper endplate ³ ; T1 slope (T1S): defined as the angle between the T1 upper endplate and the horizontal line. These parameters are measured on preoperative neutral position cervical X-rays by two independent professionals, with the average taken as the final result.

Using ImageJ software (version 1.51α, National Institutes of Health, Bethesda, MD, USA), the fat infiltration ratio (FI%) is defined as the proportion of the fat-infiltrated area to the C2 total cross-sectional area of the corresponding deep extensor muscle.

Radiomics Feature Extraction

MRI Acquisition and Segmentation

All patients undergo preoperative examinations on the same 3T MRI equipment. T2-weighted axial sequence parameters: TR/TE = 3000/120 ms, slice thickness 3 mm, gap 0.5 mm, field of view (FOV) 240 × 240 mm, matrix size 256 × 256. To ensure the homogeneity of muscle selection, the muscle level was selected parallel to the C2 lower endplate. Two experienced spine surgeons manually segment the C2 spinous process muscle region (including semispinalis cervicis, multifidus, and other muscles attached to the C2 spinous process) on T2-weighted axial images using ITK-SNAP software, with ICC >0.9 indicating good comparability.

Radiomics Feature Extraction

PyRadiomics package is used to extract radiomics features from the segmented C2 spinous process muscle region. Images are preprocessed before feature extraction, including resampling to 1 × 1 × 1 mm³ voxel size and intensity normalization. The texture features describe the patterns, or the second- and high-order spatial distributions of the intensities. Here the texture features are extracted using several different methods, including the gray-level co-occurrence matrix (GLCM), gray-level run length matrix (GLRLM), gray level size zone matrix (GLSZM) and neighborhood gray-tone difference matrix (NGTDM) methods. A total of over 1833 radiomics features are extracted for subsequent analysis.

Feature Selection and Model Construction

We conduct a Mann–Whitney U test on every radiomic feature and retain only those with P < .05. For features that demonstrate high repeatability, we calculate Spearman’s rank correlation coefficient; when any two features show a coefficient >0.9, we keep one of them. To maximize descriptive capacity, we apply a greedy recursive elimination strategy, deleting the most redundant feature at each step, until 23 features remain.

Next, we use the Least Absolute Shrinkage and Selection Operator (LASSO) regression model on the discovery data set to construct the radiomic signature. LASSO shrinks all regression coefficients toward zero and sets many irrelevant coefficients exactly to zero, depending on the regularization weight λ. We identify the optimal λ by 10-fold cross-validation with the minimum-error criterion. The retained non-zero-coefficient features fit the regression model and combine to form the radiomics signature. We then compute a radiomics score for each patient as a linear combination of these features weighted by their model coefficients.

After LASSO screening, we feed the final feature set into multiple machine-learning algorithms for risk-model construction and adopt five-fold cross-validation to obtain the final Rad-Signature.

Clinical Signature

We build the clinical signature in essentially the same manner. First, we select clinical variables with baseline P < .05, then train the same suite of machine-learning models used for the radiomics signature. Five-fold cross-validation and a fixed test cohort ensure fair comparison.

Machine Learning Modeling

We develop predictive models using the selected radiomic and statistically significant clinical features. We split the dataset into a 80% training set and a 20% test set. Within the training set, we perform five-fold cross-validation. The benchmark algorithms include SVM (Support Vector Machine), KNN (K-Nearest Neighbors), RandomForest (Random Forest), ExtraTrees (Extremely Randomized Trees) ,XGBoost (Extreme Gradient Boosting), LightGBM (Light Gradient Boosting Machine). We tune hyperparameters via grid search and evaluate performance with the area under the receiver-operating-characteristic curve (ROC).

Radiomic Nomogram

Radiomic nomogram is established in combination with radiomic signature and clinical signature. The diagnostic efficacy of radiomic nomogram is tested in test cohort, ROC curves are drawn to evaluate the diagnostic efficacy of nomogram.

Statistical Analysis

Statistical analyses were performed using SPSS version 27.0 (IBM Corp.) and Python 3.7.1 (https://www.python.org). Feature extraction and screening and model development and validation are all completed in Python. SPSS is used to compare the variables between cohorts. All statistical analyses were two-sided, and statistical significance was set at P < .05. Continuous variables are presented with descriptive statistics: standard deviation (SD), mean, median, and quartile spacing, where appropriate. Z score normalization is performed on all radiomics features to improve the comparability of the various features. The DeLong test is used to compare differences between ROC curves

Results

Patient Baseline Characteristics

Comparison of baseline characteristics between AS and non-AS groups is shown in Table 1. There are no significant differences between the two groups in age, BMI, disease duration, spinal cord compression degree, preoperative VAS, and preoperative JOA scores (P > .05). Preoperative C2-C7 Cobb angle in the AS group is significantly lower than the non-AS group (12.80° ± 7.49° vs 18.02° ± 8.59°, P = .006), while preoperative cSVA (3.01 ± 0.87 cm vs 2.46 ± 1.19 cm, P = .026) and T1S (26.68° ± 5.12° vs 23.66° ± 7.58°, P = .025) are significantly higher. C2 muscle fat infiltration (20.62 ± 6.93 vs 23.73 ± 7.78, P = 0.026) is lower in non-AS group. Postoperative follow-up VAS scores in the AS group are significantly higher than the non-AS group (4.77 ± 1.04 vs 2.15 ± 1.71, P < .001). Patient characteristics in training and test sets are shown in Table 2; there are no significant differences between the two sets in all indicators (P > .05), indicating balanced data allocation.

Table 1.

The Comparison of Two Groups’ Characteristics (Non-AS and AS)

	Non-AS	AS	P
Sample size	66	30
Duration	28.42 + 44.97	22.33 + 25.67	P = .25
Age	58.89 ± 11.99	61.30 ± 8.69	P = .33
BMI	25.85 + 3.72	25.43 ± 2.64	P = .58
Pre-C2-C7 cobb	18.02 ± 8.59	12.80 + 7.49	P = .006
Pre-cSVA	2.46 + 1.19	3.0140 ± 0.87	P = .026
Pre-T1S	23.66 + 7.58	26.68 ± 5.12	P = .025
Spinal cord compression	15.35 ± 6.72	15.80 ± 7.35	P = .384
C2 muscle fat infiltration	20.62 ± 6.93	23.73 ± 7.78	P = .026
Pre-VAS	3.88 + 1.97	3.50 + 1.70	P = .183
Pre-JOA	11.09 + 2.42	11.67 ± 2.41	P = .141
Final-follow VAS	2.15 ± 1.71	4.77 ± 1.04	P < .001
Final-follow JOA	14.18 ± 2.76	14.57 ± 3.01	P = .27
Gender(male/female)	30/36	14/16	P = .912
Diabetes (no/yes)	44/22	17/13	P = .345
Smoke (No/Yes)	44/22	23/7	P = .323

Table 2.

The Comparison of Train Set and Test Set

	Train	Test	P
Sample size	76	20
Duration	24.19 ± 37.92	35.15 + 46.43	P = .138
Age	59.18 + 11.40	61.40 ± 9.81	P = .214
BMI	25.86 ± 3.55	25.22 ± 2.83	P = .229
Pre-C2-C7 cobb	15.96 + 8.23	18.12 + 9.78	P = .159
Pre-cSVA	2.64 ± 1.10	2.60 ± 1.25	P = .437
Pre-T1S	24.57 ± 7.11	24.73 + 6.88	P = .465
Spinal cord compression	15.18 ± 6.37	16.65 ± 8.68	P = .2
C2 muscle fat infiltration	21.68 ± 7.48	21.27 ± 6.82	P = .413
Pre-VAS	3.72 + 1.94	3.90 ± 1.71	P = .356
Pre-JOA	11.22 ± 2.38	11.45 2.61	P = .356
Final-follow VAS	2.89 ± 1.96	3.25 + 1.94	P = .236
Final-follow JOA	14.18 ± 2.81	14.75 + 2.94	P = .215
Gender(male/female)	36/40	8/12	P = .556
Diabetes (no/yes)	50/26	11/9	P = .372
Smoke (no/yes)	55/21	12/8	P = .284

Signature Building

Features Statistics: A total of 7 categories, 1833 handcrafted features are extracted, including 360 first-order features, 14 shape features, and the last are texture features, shown in Figure 2. All handcrafted features are extracted with an in-house feature analysis program implemented in Pyradiomics (https://pyradiomics.readthedocs.io).

Figure 2.

Number and Ratio of Handcrafted Features

Lasso Feature Selection

Lasso feature selection: Nonzero coefficients were selected to establish the Rad-score with a least absolute shrinkage and selection operator (LASSO) logistic regression model. Coefficients and MSE(mean standard error) of 10 folds validation is show in Figure 3A and B. From over 1833 radiomics features, 5 most predictive features are selected, including:

• 1. lbp_3D_k_glcm_Correlation: local binary pattern gray-level co-occurrence matrix correlation feature (largest contribution)

• 2. lbp_3D_k_glszm_ZonePercentage: local binary pattern gray-level size zone matrix zone percentage feature

• 3. wavelet_LHH_firstorder_Median: wavelet transform LHH decomposition first-order statistic median feature

• 4. wavelet_HHH_firstorder_Mean: wavelet transform HHH decomposition first-order statistic mean feature

• 5. wavelet_LLL_firstorder_Kurtosis: wavelet transform LLL decomposition first-order statistic kurtosis feature

Figure 3.

Feature Selection With the LASSO Regression Model. (A) The LASSO Model’s Tuning Parameter (B) Selection Used 10-Fold Cross-Validation.jpg

These features reflect the texture characteristics, gray-level distribution, and spatial relationships of C2 spinous process muscles from different angles, providing multi-dimensional information for the prediction model. Figure 4 shows the coefficients value in the final selected none zero features.

Figure 4.

The Bar Graph Features That Achieved Nonzero Coefficients

Radiomics Model Performance

Table 3 summarizes results of all machine learning methods in radiomics model performance. LightGBM models perform best in test set, with AUC of 0.845, indicating that radiomics features have good predictive ability. Figure 5 shows the ROC curves of various machine learning models based on radiomics features on the test set.

Table 3.

Comparison of Machine Learning Performance of Radiomics Model

	Model_	Accuracy	AUC	Sensitivity	Specificity	PPV	NPV	Precision	Recall	F1	Threshold	Task
0	SVM	0.803	0.848	0.792	0.808	0.655	0.894	0.655	0.792	0.717	0.299	Label-train
1	SVM	0.800	0.810	0.500	0.929	0.750	0.812	0.750	0.500	0.600	0.393	Label-test
2	KNN	0.697	0.796	0.125	0.962	0.600	0.704	0.600	0.125	0.207	0.600	Label-train
3	KNN	0.700	0.637	0.333	0.857	0.500	0.750	0.500	0.333	0.400	0.400	Label-test
4	RandomForest	0.895	0.942	0.833	0.923	0.833	0.923	0.833	0.833	0.833	0.327	Label-train
5	RandomForest	0.800	0.810	0.500	0.929	0.750	0.812	0.750	0.500	0.600	0.500	Label-test
6	ExtraTrees	0.816	0.891	0.833	0.808	0.667	0.913	0.667	0.833	0.741	0.309	Label-train
7	ExtraTrees	0.750	0.762	0.500	0.857	0.600	0.800	0.600	0.500	0.545	0.403	Label-test
8	XGBoost	0.921	0.986	0.917	0.923	0.846	0.960	0.846	0.917	0.880	0.313	Label-train
9	XGBoost	0.700	0.762	0.667	0.714	0.500	0.833	0.500	0.667	0.571	0.227	Label-test
10	LightGBM	0.750	0.835	0.750	0.750	0.581	0.867	0.581	0.750	0.655	0.319	Label-train
11	LightGBM	0.700	0.845	0.667	0.714	0.500	0.833	0.500	0.667	0.571	0.319	Label-test

Figure 5.

Radiomics-Model ROC in Test Set

Clinical Model Performance

Figure 6 shows the ROC curves of various machine learning models based on clinical features (C2-C7 cobb, T1s, cSVA, C2 muscle fat infiltration) on the test set. Table 4 summarizes results of all machine learning methods. Among all clinical models, LightGBM performs best, with a test set AUC of 0.744.

Figure 6.

Clinical-Model ROC in Test Set

Table 4.

Comparison of Machine Learning Performance of Clinical Model

	Model	Accuracy	AUC	Sensitivity	Specificity	PPV	NPV	Precision	Recall	F1	Threshold	Task
0	SVM	0.303	0.167	0.958	0.000	0.307	0.000	0.307	0.958	0.465	0.285	Label-train
1	SVM	0.450	0.333	0.500	0.429	0.273	0.667	0.273	0.500	0.353	0.315	Label-test
2	KNN	0.789	0.858	0.667	0.846	0.667	0.846	0.667	0.667	0.667	0.400	Label-train
3	KNN	0.700	0.571	0.167	0.929	0.500	0.722	0.500	0.167	0.250	0.600	Label-test
4	RandomForest	0.961	1.000	0.875	1.000	1.000	0.945	1.000	0.875	0.933	0.500	Label-train
5	RandomForest	0.550	0.738	0.500	0.571	0.333	0.727	0.333	0.500	0.400	0.300	Label-test
6	ExtraTrees	0.684	1.000	0.000	1.000	0.000	0.684	0.000	0.000	NaN	1.000	Label-train
7	ExtraTrees	0.550	0.667	0.833	0.429	0.385	0.857	0.385	0.833	0.526	0.100	Label-test
8	XGBoost	0.987	1.000	0.958	1.000	1.000	0.981	1.000	0.958	0.979	0.494	Label-train
9	XGBoost	0.650	0.732	0.833	0.571	0.455	0.889	0.455	0.833	0.588	0.224	Label-test
10	LightGBM	0.776	0.822	0.708	0.808	0.630	0.857	0.630	0.708	0.667	0.360	Label-train
11	LightGBM	0.600	0.744	0.667	0.571	0.400	0.800	0.400	0.667	0.500	0.260	Label-test

Radiomics-Clinical Model

AUC

Table 5 summarizes results of clinical model, radiomics model and combined model. Figure 7A and B shows the ROC curves of radiomics-model, clinical model and combined model in train set and test set. Both clinical signature and rad signature get the prefect fitting. The combined is preformed to combine clinical signature and rad signature, which shows the best performance with AUC of 0.881 in test set.

Table 5.

Comparative Performance of Clinical, Radiomics, and Combined Models

Model-name	Accuracy	AUC	Sensitivity	Specificity	PPV	NPV	Precision	Recall	F1	Threshold	Task
Clinic signature	0.776	0.822	0.708	0.808	0.630	0.857	0.630	0.708	0.667	0.360	Train
Rad signature	0.750	0.835	0.750	0.750	0.581	0.867	0.581	0.750	0.655	0.319	Train
Nomogram	0.803	0.920	0.917	0.750	0.629	0.951	0.629	0.917	0.746	0.307	Train
Clinic signature	0.600	0.744	0.667	0.571	0.400	0.800	0.400	0.667	0.500	0.260	Test
Rad signature	0.700	0.845	0.667	0.714	0.500	0.833	0.500	0.667	0.571	0.319	Test
Nomogram	0.800	0.881	0.833	0.786	0.625	0.917	0.625	0.833	0.714	0.300	Test

Figure 7.

Radio-Clinical Model ROC (A) Train Set (B) Test Set

Nomogram

We developed a nomogram to visualize the combined model (Figure 8), allowing for the calculation of risk by summing the points for each variable along the corresponding axis.

Figure 8.

Nomogram of Combined Models

Discussion

This study is the first attempt to establish a machine learning model for predicting AS occurrence after laminoplasty based on C2 spinous process muscle radiomics features and cervical sagittal parameters. Results show that the combined model integrating radiomics features and clinical sagittal parameters significantly outperforms models using radiomics features or clinical parameters alone, providing clinical physicians with a more precise risk assessment tool.

Regarding sagittal parameters, we find that preoperative C2-C7 Cobb angle in the AS group is significantly lower than the non-AS group, consistent with Ruan et al’s findings that greater preoperative cervical curvature is a protective factor for AS. When preoperative cervical lordosis is insufficient, patients’ neck and shoulder areas are already in a compensatory state of reduced physiological lordosis, with muscles and ligaments under long-term tension and fatigue, leading to further axial symptoms postoperatively.^17,18 Our sagittal parameter correlation analysis shows that preoperative C2-C7 Cobb angle is negatively correlated with AS, further confirming this finding. Meanwhile, we find that preoperative cSVA and T1S in the AS group are significantly higher than the non-AS group, consistent with Zuo et al’s findings. T1 slope is the only parameter connecting the cervical and thoracic spine. A larger T1 slope usually requires stronger posterior neck muscle strength to maintain normal physiological curvature and spinal balance.^19,20 Laminoplasty surgery itself causes muscle damage, and postoperative insufficient muscle strength fails to maintain sagittal balance, leading to AS occurrence. The fat infiltration ratio of the C2 spinous process muscles serves as a quantifiable indicator of muscle degeneration. FI% reflects the proportion of fat content within the total cross-sectional area of the muscle, representing a comprehensive measure of both muscle quality and metabolic status. Previous studies have shown that fatty infiltration of deep cervical extensor muscles is closely associated with postoperative neck dysfunction, and its severity is negatively correlated with recovery outcomes.^14,21 In our analysis, patients in the AS group exhibited significantly higher FI% compared to the non-AS group, suggesting that a greater degree of preoperative muscle degeneration is associated with an increased risk of postoperative axial symptoms. The clinical model (C2-C7 Cobb, cSVA, T1 S, C2 muscle FI) achieves AUC 0.744, sensitivity 0.778, specificity 0.571. It allows rapid outpatient AS risk triage and complements radiomics by capturing macro sagittal imbalance, while radiomics detects micro-level muscle damage.

Radiomics feature analysis reveals key changes in the microscopic structure of preoperative C2 spinous process muscles, providing a new perspective for predicting AS occurrence after laminoplasty. This study uses LASSO regression to select 5 most predictive radiomics markers from 1833 high-dimensional features extracted from preoperative MRI, including local binary pattern features and wavelet transform features, which reflect the characteristics of preoperative muscle internal structure from different dimensions and can be used for preoperative risk assessment. Among the selected features, lbp_3D_k_glcm_Correlation (Gray Level Co-occurrence Matrix Correlation based on Local Binary Patterns, LBP) exhibited the strongest predictive capability. This feature quantifies the spatial correlation and textural homogeneity of grayscale distribution within muscle tissues. A lower value indicates more disorganized grayscale patterns and disrupted structural continuity, reflecting increased microstructural heterogeneity and disarray. Such subtle alterations are often imperceptible on conventional MRI but may serve as important biological indicators for postoperative axial symptoms (AS), suggesting impaired mechanical continuity and diminished recovery potential of the muscle.

Another LBP-derived feature, lbp_3D_k_glszm_ZonePercentage, evaluates muscle homogeneity from the perspective of regional distribution, representing the proportion of small, uniform texture zones within the image. A significant decrease in this feature among patients with AS indicates potential preoperative regional pathological changes, such as focal fatty infiltration or microvascular disturbances, which may compromise the muscle’s tolerance to surgical trauma and lead to postoperative functional impairment.

Additionally, three wavelet-based features — wavelet_LHH_firstorder_Median, wavelet_HHH_firstorder_Mean, and wavelet_LLL_firstorder_Kurtosis — provide insights into the internal muscle structure from a frequency-domain, multiscale perspective. These features capture signal characteristics associated with various biological alterations: a decrease in wavelet_LHH_firstorder_Median suggests reduced contrast at the muscle-fat interface, possibly due to diffuse fatty infiltration or edema, indicative of diminished elastic reserve; an increase in wavelet_HHH_firstorder_Mean reflects high-frequency “rough” signals, suggesting fibrotic bands or microtears and increased muscle stiffness; most notably, a significant increase in wavelet_LLL_firstorder_Kurtosis in the AS group indicates abnormal signal concentration, potentially representing focal “islands” of fat or edema, which may compromise the muscle’s biomechanical adaptability to postoperative loading, resulting in imbalance.

Traditional preoperative imaging evaluation mainly relies on macroscopic morphological indicators of muscles and skeletons, while radiomics, through high-dimensional data mining of preoperative MRI, can comprehensively quantitatively assess preoperative muscle condition from multiple microscopic dimensions such as texture features, spatial relationships, and gray-level distribution, breaking through the limitations of traditional visual assessment.^22,23 This high-precision analysis based on preoperative imaging not only improves the accuracy of postoperative AS risk prediction models but also provides new methodological tools for identifying high-risk patients preoperatively, potentially guiding more precise preoperative risk assessment and individualized surgical strategy formulation, thereby reducing postoperative AS occurrence and improving patient prognosis and quality of life.

Among machine learning models, when using radiomics features alone, LightGBM performs best on the test set (AUC = 0.845); when using clinical features alone, LightGBM also performs best (AUC = 0.744); while the combined model shows the best predictive performance (AUC = 0.881). This indicates complementarity between C2 spinous process muscle radiomics features and traditional clinical parameters, with more comprehensive predictive information obtained by combining both. The advantage of LightGBM in this study stems from the good match between its technical architecture and research data characteristics. Its leaf-wise growth strategy and histogram optimization algorithm efficiently process over 1800 high-dimensional radiomics features extracted from C2 spinous process muscles and accurately capture non-linear patterns in texture features. Facing class imbalance between AS group (30 cases) and non-AS group (66 cases), LightGBM’s weighting mechanism performs excellently, maintaining a test set AUC of 0.845 while showing better predictive performance than other machine learning algorithm. Its excellent noise robustness mitigates the impact of MRI image quality fluctuations and manual segmentation errors, while automatically discovering interactions between muscle microstructure and sagittal parameters, which is crucial for integrating multi-source predictive factors and constructing clinically useful AS risk assessment models. ²⁴

This study has several limitations. It is a single-center retrospective study with a relatively small sample size, which limits the generalizability of the findings. However, standardized surgical techniques and imaging protocols ensure internal data consistency. External validation using multi-center data is warranted but remains unfeasible due to current resource and data access constraints. In addition, manual segmentation of the C2 spinous process muscles introduces potential subjectivity, although we ensure interobserver consistency. Future research focuses on prospective multi-center validation, incorporation of multi-modal imaging, adoption of deep learning for automated segmentation, and integration of perioperative variables to develop more comprehensive and generalizable predictive models.

Conclusion

This research demonstrates that a machine learning model integrating C2 spinous process muscle radiomics features and clinical parameters(C2-C7 Cobb angle, cSVA , T1S values and C2 muscle fat infiltration) effectively predicts axial symptoms after unilateral laminoplasty (AUC = 0.881). This innovative tool enables preoperative risk stratification, supporting personalized surgical planning and rehabilitation protocols to improve patient outcomes.

ORCID iDs

Zhenqi Zhu https://orcid.org/0009-0004-8101-7556

Yan Liang https://orcid.org/0009-0008-7414-9998

Haiying Liu https://orcid.org/0000-0002-5755-7390