{"paper_id":"29dec4ec-e756-4466-bbee-b49d069a68e9","body_text":"Deep Learning for Predicting Lumbar Instability Using Neutral Lateral Lumbar Radiographs: A Retrospective Study | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Deep Learning for Predicting Lumbar Instability Using Neutral Lateral Lumbar Radiographs: A Retrospective Study Jiajun Song, Jiawei Du, Shengwei Liu, Junyu Chen, Di Zhang, Shi-Qing Feng This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7410485/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 9 You are reading this latest preprint version Abstract Objective To develop a deep learning model for predicting lumbar segmental instability (LSI) using neutral lateral lumbar radiographs and to identify key radiographic features associated with LSI. Methods A DenseNet121-based stacking ensemble model was integrated with Support Vector Machine, Random Forest, and Softmax classifiers. Model validation employed 10-fold cross-validation, with performance assessed using AUC, accuracy, sensitivity, specificity, and F1-score. Sensitivity analyses evaluated robustness across spinal/non-spinal regions, age/gender subgroups, and feature interactions. Gradient-weighted Class Activation Mapping (Grad-CAM) was used to localize critical anatomical regions, which were further validated in machine learning frameworks. Results The DenseNet121-stacking model achieved an AUC of 0.82, accuracy of 76%, sensitivity of 61%, and specificity of 84%. Calibration curves confirmed strong alignment with clinical outcomes. Grad-CAM identified facet joints (34.1%), intervertebral discs (27.0%), and osteophytes (25.4%) as the predominant contributors. The integration of these features into machine learning models yielded an AUC of 0.749. Subgroup analyses demonstrated consistent performance across age and gender groups. Decision curve analysis confirmed the clinical utility of the model in all cohorts. Conclusion The stacking ensemble model developed in this study effectively predicts LSI based on neutral lateral lumbar radiographs and identifies key imaging biomarkers, including facet joint hypertrophy, disc degeneration, and osteophyte formation. The model demonstrated stable performance across different age and gender groups, indicating strong generalizability and providing a reliable tool for precise clinical screening and individualized decision-making. Low Back Pain Lumbar Vertebrae Radiography Deep Learning Predictive Learning Models Data Visualization Figures Figure 1 Figure 2 Figure 3 Introduction Lumbar segmental instability (LSI) accounts for up to 57% of chronic low back pain and, if untreated, may progress to degeneration requiring fusion[ 1 ]. Diagnosis mainly relies on flexion–extension radiographs measuring intervertebral range of motion (IROM) and sagittal translation (ΔST) to detect abnormal movement[ 2 ]. However, this method depends on manual landmarking, introducing subjectivity and low efficiency. Since the flexion-extension maneuver is difficult to standardize, the reproducibility of results is poor[ 3 ]. Moreover, flexion-extension radiography typically requires patients to visit specialized medical centers equipped with dedicated devices and experienced clinicians, resulting in a high threshold for clinical implementation. Extra radiation further limits early screening and acceptance, reducing early detection opportunities. Therefore, a safe and cost-effective alternative is needed. Nearly all hospitals are equipped to perform routine neutral lateral lumbar radiography. These routinely acquired images may contain latent imaging cues related to LSI that are imperceptible to the human eye but can be extracted without additional cost, and even analyzed retrospectively. Fully leveraging such widely available imaging data could greatly expand population-level screening for spinal instability. However, identifying and quantifying these “micro-instability” features from single static images remains technically challenging and requires advanced image analysis algorithms to achieve reliable clinical diagnostic performance. Deep learning-based computer vision methods offer a promising solution, as they are capable of identifying hidden diagnostic information within medical images. For example, infer cardiovascular risk factors from retinal fundus photograph[ 4 ], estimate cervical spinal-cord compression directly from routine cervical radiography[ 5 ], identify facial phenotypes of genetic disorders from ordinary photographs[ 6 ], and detect osteoporosis on routine X-ray [ 7 ]. Building on these advances, we developed a deep-learning model that analyses neutral lateral lumbar radiographs to predict LSI. Internal and external tests confirmed the algorithm’s feasibility and diagnostic accuracy. Because neutral lateral radiography is routine for back pain and other diseases, patients can be automatically assessed for LSI use neutral lateral radiography without extra cost or radiation. Hospitals can apply the model on archived radiography to spot undiagnosed high-risk patients and warn them early. Materials and Methods Study Sample This single-center retrospective study enrolled patients who underwent lumbar spine radiography at XXX from June 2012 to November 2022. The study was approved by the institutional review board of XXX (Approval No. XXX). Anteroposterior, neutral lateral, hyperextension, and hyperflexion radiographs in DICOM format were collected. Patient identifiers were removed using the MicroDicom anonymization tool. Written informed consent was obtained from all eligible participants. Inclusion criteria: (1) high-quality radiographs, (2) all three views available, and (3) no prior spinal surgery. Exclusion criteria included: (1) poor image quality, (2) severe scoliosis, (3) prior spinal surgery, or (4) anatomical variants like lumbarization or sacralization. Ultimately, 1,588 patients were included and split into a training set (1,107 cases: 315 positive, 792 negative) and a testing set (481 cases: 166 positive, 315 negative). Model training and hyperparameter tuning were performed only on the training set to avoid data leakage. Sample Labeling LSI was defined per Nachemson’s radiological criteria.[ 2 ]. IROM and ΔST were measured at each level from L1–2 to L5–S1 (Supplementary Fig. 1). The method for measuring ΔST followed prior literature[ 8 ]. A histogram presents the distribution of IROM and ΔST across levels. Instability was diagnosed if any segment met either: (i) IROM ≥ 10° (L1–2 to L4–5) or ≥ 20° (L5–S1); or (ii) ΔST ≥ 3 mm (L1–2 to L4–5) or ≥ 4 mm (L5–S1).[ 2 ]. Radiographic diagnoses were conducted by three spine surgeons. Each case was independently evaluated by at least two surgeons. If their assessments differed, the final diagnosis was established by vote after discussion among all three reviewers. Data Preprocessing Neutral lateral radiographs were cropped to include: (1) from the lower one-third of T12 to the upper one-third of S1, (2) anteriorly to the vertebral edge (including osteophytes), and (3) posteriorly to the tip of the spinous process. This region was selected based on anatomical relevance to lumbar instability. Images underwent Z-score normalization (mean = 0, SD = 1) and channel-wise scaling to 0–1 to enhance consistency and model performance. Data augmentation—including horizontal flipping, rotation, shifting, zooming, and shearing—was applied randomly to expand the dataset and reduce overfitting. Analysis 1: Classifying LSI based on DenseNet121 Model In this study, we used DenseNet121[ 9 ], pretrained on ImageNet, as a feature extractor. Features from the fully connected layer were input into three classifiers: Support Vector Machine (SVM), Random Forests (RF), and Softmax, with SVM and RF being widely used in classification tasks. A 10-fold cross-validation was applied within the training set to tune hyperparameters and prevent overfitting. Hyperparameter optimization was performed via grid search with 10-fold cross-validation (see Supplementary Table S1 for hyperparameter). The optimal configuration was selected based on area under the curve (AUC) maximization. Figure 1 illustrates the overall workflow. The final model was tested on a held-out test set. Classifier outputs were in probability form, and a stacking ensemble method with soft voting was implemented by averaging these probabilities. To evaluate model performance, we computed metrics including AUC, accuracy, sensitivity, specificity, precision, recall, and F1-score Calibration plots were also generated to assess the alignment of predicted probabilities with actual outcomes. Analysis 2: Sensitivity analyses To assess robustness of the diagnostic model in analysis 1, we performed four sensitivity analyses to ensure the model overly sensitive to data fluctuations. (1) Spinal region specificity: Models trained on masked spinal vs. non-spinal regions showed whether diagnostic features originated from spinal structures. (2) Age subgroup generalization: After 1:1 age-matching (912 samples), participants were stratified as elderly (≥ 60) or younger (< 60). Models were trained/tested within and across subgroups to assess age effects. Performance was compared by ROC and DeLong test. (3) Gender subgroup generalization: Models were trained/tested within male and female groups, and cross-sex validation examined whether predictions reflected LSI features rather than sex differences. ROC and DeLong test evaluated AUCs. (4) Age-gender interaction: Extreme subgroup combinations (elderly male vs. young female; elderly female vs. young male) further tested population-specific effects on performance, assessed by ROC, DeLong test, and standard metrics. Analysis 3: Grad-CAM-Guided Exploration of Interpretable Features for LSI Classification Grad-CAM (Gradient-weighted Class Activation Mapping) is a visualization technique that utilizes the gradient information from the final convolutional layer of a convolutional neural network (CNN) to generate class activation maps, highlighting image regions most relevant to the classification decision. In Analyses 1 and 2, we applied Grad-CAM to each trained model to visualize the spatial attention patterns and statistically assess whether the highlighted anatomical regions were associated with LSI. This approach also aimed to identify potential radiographic indicators of LSI. Using the resulting heatmaps, we extracted anatomical structures corresponding to the high-activation regions in each case. We quantified their occurrence and evaluated the discriminative power of the three most frequently activated regions in distinguishing between LSI and non-LSI patients. These features were subsequently used as input variables for machine learning classifiers, including SVM, K-Nearest Neighbors (KNN), and Logistic Regression (LR), to further assess their predictive utility in the classification of LSI. Results Baseline Patient Characteristics We assessed 1,588 patients. Table 1 summarizes cohort characteristics. In the LSI group, mean age was 58.25 years and 44% were male. Table 1 , Baseline Clinical Characteristics and Outcomes of the 1588 Patient Cohort. Characteristic Mean ± standard deviation or frequency (proportion) Non-LSI Group LSI Group Age(yr) 50.45 ± 13.85 58.25 ± 13.83 Male sex 540(48.78) 212(44.07) Pain episode (First) 889(80.31) 360(74.84) Pain radiation 277(25.02) 98(20.37) Trauma history 22(1.99) 8(1.66) Drinking 157(14.18) 61(12.68) Smoking status 326(29.45) 137(28.48) Diabetes 153(13.82) 52(10.81) Hypertension 164(14.81) 92(19.13) Heart disease 50(4.52) 45(9.36) Continuous variables are presented as the mean ± standard deviation, while categorical variables are reported as frequencies with corresponding percentages (%). LSI: Lumbar Spine Instability. Non-LSI: Non- Lumbar Spine Instability. LSI Prediction Based on DenseNet121 Model and Stacking Learning The DenseNet121-SVM model achieved an AUC of 0.750, accuracy of 0.740, sensitivity of 0.470, and specificity of 0.880; The DenseNet121-RF model achieved an AUC of 0.755, accuracy of 0.694, sensitivity of 0.741, and specificity of 0.670.; The DenseNet121-Softmax model achieved an AUC of 0.795, accuracy of 0.738, sensitivity of 0.657, and specificity of 0.781. Among these, the DenseNet121-stacking model showed the best performance, with an AUC of 0.820, accuracy of 0.760, sensitivity of 0.610, and specificity of 0.840. The calibration curve showed good agreement with the ideal reference line. Decision curve analysis (DCA) demonstrated clinical utility across a range of threshold probabilities, particularly for the stacking model (Table 2 , Fig. 1 ). Table 2 , Comparison of the AUCs, Accuracies, Sensitivities, Specificities, Precisions, Recalls and F1-scores for four DenseNet121-models. ModelName Acc AUC Sensitivity Specificity Precision Recall F1 DenseNet121-SVM 0.74 0.75 0.47 0.883 0.678 0.47 0.555 DenseNet121-RF 0.694 0.755 0.741 0.67 0.542 0.741 0.626 DenseNet121-Softmax 0.738 0.795 0.657 0.781 0.612 0.657 0.634 DenseNet121-Stacking 0.763 0.823 0.614 0.841 0.671 0.614 0.642 AUC: Area Under Curve; SVM: Support Vector Machine; RF: Random Forest. Sensitivity analyses Several analyses tested the robustness of the stacking model. First, models trained only on non-lumbar regions showed worse performance than those trained on lumbar regions (Table 3 , Fig. 1 f–i), indicating classification relied on lumbar anatomy. Second, subgroup analyses showed stable performance across age and sex with no significant AUC differences (DeLong test, P > 0.05; Fig. 2 a–d, Table 4 ). Finally, cross-validation between age and gender groups showed the model still distinguished LSI from controls even when trained on young males and tested on elderly females, and vice versa, confirming classification was not driven by demographic features but by LSI-related structural patterns (Fig. 2 a–d, Table 4 ). Table 3 , Comparison of the AUCs, Accuracies, Sensitivities, Specificities, Precisions, Recalls and F1-scores for four DenseNet121-models without the spinal area. ModelName Acc AUC Sensitivity Specificity Precision Recall F1 DenseNet121-SVM 0.624 0.559 0.289 0.8 0.432 0.289 0.347 DenseNet121-RF 0.503 0.583 0.807 0.343 0.393 0.807 0.529 DenseNet121-Softmax 0.615 0.571 0.199 0.835 0.388 0.199 0.263 DenseNet121-Stacking 0.578 0.583 0.524 0.606 0.412 0.524 0.462 AUC: Area Under Curve; SVM: Support Vector Machine; RF: Random Forest. Table 4 , Performance Metrics of the DenseNet121-Stacking Model in Age and Gender Subgroup Analysis. ModelName Acc AUC Sensitivity Specificity Precision Recall F1 F→M 0.754 0.824 0.705 0.788 0.698 0.705 0.701 M→F 0.77 0.812 0.735 0.791 0.67 0.735 0.701 Y→E 0.716 0.787 0.65 0.757 0.623 0.65 0.636 E→Y 0.729 0.801 0.761 0.704 0.665 0.761 0.71 EF→YM 0.736 0.797 0.581 0.835 0.694 0.581 0.632 EM→YF 0.753 0.806 0.755 0.751 0.668 0.755 0.709 AUC: Area Under Curve. F: Female group, M: Male group, Y: Young group, E: Elderly group, EF: Elderly female group, YM: Young male group, EM: Elderly male group, YF: Young female group. LSI associated structural region identified by Grad-CAM and its predictive utility To interpret the “black box” model, Grad-CAM highlighted key anatomical regions including facet joints, discs, and osteophytes (Fig. 3 a–d). In LSI patients, these occurred most frequently (facet joint 34.1%, disc 27.0%, osteophyte 25.4%). This indicates that radiographic signs in these three regions contributed most to classification. We therefore evaluated three features—facet joint hypertrophy, disc height reduction, and osteophyte formation. Individually, their predictive value was limited, but when combined in machine learning models, classification improved (max AUC 0.749; Fig. 3 f). These findings suggest the coexistence of these signs is strongly associated with LSI. Discussion This study yielded three main findings: (1) Deep learning applied to neutral lateral lumbar radiographs accurately distinguished LSI from non-LSI cases, with the DenseNet121-stacking model performing best (AUC = 0.820, accuracy = 0.760, sensitivity = 0.610, specificity = 0.840); (2) Sensitivity analyses confirmed the model’s robustness across age and gender groups; (3) Grad-CAM revealed that combining three imaging features—facet joint hypertrophy, osteophytes, and reduced disc height—provided added diagnostic value. Prior research has explored diverse LSI predictors to improve diagnostic efficiency. Jesse L. Even et al. reported that interspinous fluid on MRI predicted > 3 mm instability with PPV 69.0%. Conversely, absence of facet joint fluid had PPV 75.6% for ruling out > 3 mm instability.[ 10 ] Tahere Seyedhoseinpoor et al. combined clinical indicators with examination; the best model (weight, lumbar lordosis, prone segmental instability) achieved AUC 0.66[ 11 ]. Although flexion–extension radiography is the gold standard, its use still depends on clinical judgment. In our evaluation, the model achieved AUC 0.85, suggesting neutral-position AI screening may exceed symptom/examination-only assessments and complement traditional evaluation. Radiographs hold broader diagnostic potential, as shown in osteoporosis and fracture studies. In a meta-analysis of osteoporosis diagnostics, 17/40 studies used radiographs[ 12 ]. Sena Goral et al. developed an interpretable capsule-network system to identify 68 vertebral landmarks and compute Cobb angles[ 13 ]. Li-Wei Cheng et al. designed a model capable of detecting and classifying vertebral fractures based on radiograph[ 14 ]. Compared with CT, radiography is cheaper, faster, and lower in radiation. Leveraging these advantages could expand population-level screening. We further utilized Grad-CAM visualization technology to identify three anatomical structures closely associated with LSI. The identification of these structures is consistent with existing clinical evidence, offering strong support for our research. Disc space narrowing is recognized as a key sign in advanced stabilization phases[ 15 ], with disc degeneration linked to instability in biomechanical and clinical studies[ 16 ]. Another notable indirect radiographic marker linked to LSI is the traction spur [ 17 ], typically 2–3 mm from the endplate and oriented horizontally—are thought to arise from tensile forces on the periosteum by Sharpey fibers or the anterior longitudinal ligament during instability episodes. Claw osteophytes, originating from the vertebral apophysis near the disc margin and curving toward the adjacent vertebra, likely reflect compressive responses and stability restoration. These features may represent different stages of a shared pathological process and often coexist along the vertebral rim[ 18 ]. Furthermore, the facet joints guide and limit vertebral movement while aiding in load distribution across the spine[ 19 ], facet joints guide vertebral motion and share load distribution; dysfunction can compromise stability. When multiple anatomical markers were incorporated simultaneously as variables into the machine learning model, the classification performance improved significantly compared to models built on a single anatomical structure alone. This finding suggests that LSI is a multifactorial pathology involving multiple anatomical structures, rather than being caused by changes in any single structure. Compared with traditional methods, our approach offers several advantages. First, it uses routine neutral lateral radiographs that are low-cost and low-dose, and the neutral posture is easier to standardize, enabling large-scale screening. Second, reproducible criteria provide greater consistency than expert manual assessments prone to inter-observer variability. Third, the diagnostic performance of our model in identifying lumbar instability based on neutral radiographs surpasses that of physical examinations, manual image interpretation by clinicians, and other conventional approaches. Finally, full automation enables rapid, scalable screening without manual image processing. LSI diagnostic criteria remain inconsistent. Some define instability as sagittal translation ≥ 4 mm or ≥ 8%, rotation ≥ 10° (L1–L5) or ≥ 20° (L5–S1)[ 20 ]. We adopted Nachemson’s criteria but lowered the L1–L2 to L4–L5 translation threshold to ≥ 3 mm for greater sensitivity, still achieving high accuracy. Grad-CAM confirmed localization to discs, facet joints, and osteophytes. Even 3 mm translation may indicate early degeneration and reduced stability; adopting a more sensitive ≥ 3 mm threshold may enable earlier identification and timely conservative interventions (e.g., core stabilization[ 21 ]), potentially preventing progression. Early identification and targeted conservative treatment of LSI can markedly improve outcomes. Studies report > 50% Oswestry Disability Index scores reduction in 49.1% of patients[ 22 ], and individuals with the greatest improvement in physical function also demonstrated relief from negative cognitive patterns related to pain perception[ 23 ]. Randomized controlled trials show core stabilization reduces VAS by 45% and improves joint play[ 24 ], while in early mild spondylolisthesis (< 20%), core reinforcement or Graf ligamentoplasty yields a 10-year adjacent segment reoperation rate of 7% versus ≥ 18% for fusion[ 25 ]. Early physical therapy in low back pain lowers risks of advanced imaging, surgery, injections, opioid use, and overall costs. These findings further highlight the importance of early detection and tailored conservative treatment strategies in patients with potential instability. This study has several limitations. First, the dataset size should be further expanded in future studies to ensure more robust validation and greater generalizability of the findings. Therefore, we believe that the accuracy of our algorithm could be further improved with a larger training dataset. Second, single-center data may limit generalizability; external validation across institutions and regions is needed. Third, we used imaging alone; integrating symptoms and examination via NLP from electronic records could enable multimodal models and improve diagnostic accuracy and reliability. Conclusion In conclusion, this study developed a deep learning algorithm based on neutral-position lateral lumbar radiographs to predict LSI, achieving high diagnostic accuracy. The model is suitable for early screening of populations at high risk for LSI. By leveraging routinely acquired imaging data, this approach provides an efficient, low-cost, and scalable solution for the early identification of individuals at risk of LSI. Declarations Author Contribution Conceptualization: Jiajun Song, Jiawei Du, Di Zhang, Shiqing Feng.Formal Analysis: Jiajun Song, Jiawei Du.Investigation: Jiajun Song, Shengwei Liu, Junyu Chen.Methodology: Jiajun Song, Jiawei Du, Shengwei Liu, Shiqing Feng.Project Administration: Di Zhang, Shiqing Feng.Writing – Original Draft: Jiajun Song, Jiawei Du, Shengwei Liu, Junyu Chen.Writing – Review & Editing: Jiawei Du, Di Zhang, Shiqing Feng.Jiajun Song and Jiawei Du contributed equally to this work. Acknowledgement AcknowledgementsWe thank the Radiology Department of Tianjin Medical University General Hospital for providing access to the anonymized radiographic datasets. References Huang H, Young W, Skaper S, et al. Clinical Neurorestorative Therapeutic Guidelines for Spinal Cord Injury (IANR/CANR version 2019). Journal of orthopaedic translation. 2020;20:14-24.http://dx.doi.org/10.1016/j.jot.2019.10.006. Nachemson A. The Role of Spine Fusion: Question 8. 1981;6(3):306-7.http://dx.doi.org/. Smith JS, Shaffrey CI, Ames CP, Lenke LG. Treatment of adult thoracolumbar spinal deformity: past, present, and future. Journal of neurosurgery Spine. 2019;30(5):551-67.http://dx.doi.org/10.3171/2019.1.Spine181494. Poplin R, Varadarajan AV, Blumer K, et al. Prediction of cardiovascular risk factors from retinal fundus photographs via deep learning. Nature biomedical engineering. 2018;2(3):158-64.http://dx.doi.org/10.1038/s41551-018-0195-0. Tamai K, Terai H, Hoshino M, et al. Deep Learning Algorithm for Identifying Cervical Cord Compression Due to Degenerative Canal Stenosis on Radiography. Spine. 2023;48(8):519-25.http://dx.doi.org/10.1097/brs.0000000000004595. Gurovich Y, Hanani Y, Bar O, et al. Identifying facial phenotypes of genetic disorders using deep learning. Nature medicine. 2019;25(1):60-4.http://dx.doi.org/10.1038/s41591-018-0279-0. Ho CS, Chen YP, Fan TY, et al. Application of deep learning neural network in predicting bone mineral density from plain X-ray radiography. Archives of osteoporosis. 2021;16(1):153.http://dx.doi.org/10.1007/s11657-021-00985-8. Iguchi T, Kanemura A, Kasahara K, et al. Lumbar instability and clinical symptoms: which is the more critical factor for symptoms: sagittal translation or segment angulation? Journal of spinal disorders & techniques. 2004;17(4):284-90.http://dx.doi.org/10.1097/01.bsd.0000102473.95064.9d. Huang G, Liu Z, Maaten LVD, Weinberger KQ, editors. Densely Connected Convolutional Networks. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR); 2017 21-26 July 2017. Even JL, Chen AF, Lee JY. Imaging characteristics of \"dynamic\" versus \"static\" spondylolisthesis: analysis using magnetic resonance imaging and flexion/extension films. The spine journal : official journal of the North American Spine Society. 2014;14(9):1965-9.http://dx.doi.org/10.1016/j.spinee.2013.11.057. Seyedhoseinpoor T, Dadgoo M, Taghipour M, et al. Combining clinical exams can better predict lumbar spine radiographic instability. Musculoskeletal science & practice. 2022;58:102504.http://dx.doi.org/10.1016/j.msksp.2022.102504. He Y, Lin J, Zhu S, Zhu J, Xu Z. Deep learning in the radiologic diagnosis of osteoporosis: a literature review. The Journal of international medical research. 2024;52(4):3000605241244754.http://dx.doi.org/10.1177/03000605241244754. Goral S, Köse U. Development of A CapsNet and Fuzzy Logic Decision Support System for Diagnosing the Scoliosis and Planning Treatments via Schroth Method. IEEE Access. 2022;10:129055-78.http://dx.doi.org/10.1109/ACCESS.2022.3227763. Cheng L-W, Chou H-H, Cai Y-X, et al. Automated detection of vertebral fractures from X-ray images: A novel machine learning model and survey of the field. Neurocomputing. 2024;566:126946.http://dx.doi.org/https://doi.org/10.1016/j.neucom.2023.126946. Kirkaldy-Willis WH, Farfan HF. Instability of the lumbar spine. Clinical orthopaedics and related research. 1982;(165):110-23.http://dx.doi.org/. Fujiwara A, Lim TH, An HS, et al. The effect of disc degeneration and facet joint osteoarthritis on the segmental flexibility of the lumbar spine. Spine. 2000;25(23):3036-44.http://dx.doi.org/10.1097/00007632-200012010-00011. Pitkänen MT, Manninen HI, Lindgren KA, Sihvonen TA, Airaksinen O, Soimakallio S. Segmental lumbar spine instability at flexion-extension radiography can be predicted by conventional radiography. Clinical radiology. 2002;57(7):632-9.http://dx.doi.org/10.1053/crad.2001.0899. Pate D, Goobar J, Resnick D, Haghighi P, Sartoris DJ, Pathria MN. Traction osteophytes of the lumbar spine: radiographic-pathologic correlation. Radiology. 1988;166(3):843-6.http://dx.doi.org/10.1148/radiology.166.3.3340781. Elder BD, Vigneswaran K, Athanasiou KA, Kim DH. Biomechanical, biochemical, and histological characterization of canine lumbar facet joint cartilage. Journal of neurosurgery Spine. 2009;10(6):623-8.http://dx.doi.org/10.3171/2009.2.Spine08818. Wang Y, Huang K. Research progress of diagnosing methodology for lumbar segmental instability: A narrative review. Medicine. 2022;101(1):e28534.http://dx.doi.org/10.1097/md.0000000000028534. Puntumetakul R, Saiklang P, Tapanya W, et al. The Effects of Core Stabilization Exercise with the Abdominal Drawing-in Maneuver Technique versus General Strengthening Exercise on Lumbar Segmental Motion in Patients with Clinical Lumbar Instability: A Randomized Controlled Trial with 12-Month Follow-Up. International journal of environmental research and public health. 2021;18(15)10.3390/ijerph18157811. Larivière C, Rabhi K, Preuss R, Coutu MF, Roy N, Henry SM. Derivation of clinical prediction rules for identifying patients with non-acute low back pain who respond best to a lumbar stabilization exercise program at post-treatment and six-month follow-up. PloS one. 2022;17(4):e0265970.http://dx.doi.org/10.1371/journal.pone.0265970. Larivière C, Preuss R, Coutu MF, Sullivan MJ, Roy N, Henry SM. Disability reduction following a lumbar stabilization exercise program for low back pain: large vs. small improvement subgroup analyses of physical and psychological variables. BMC musculoskeletal disorders. 2024;25(1):358.http://dx.doi.org/10.1186/s12891-024-07480-4. Kumar SP. Efficacy of segmental stabilization exercise for lumbar segmental instability in patients with mechanical low back pain: A randomized placebo controlled crossover study. North American journal of medical sciences. 2011;3(10):456-61.http://dx.doi.org/10.4297/najms.2011.3456. Kanayama M, Hashimoto T, Shigenobu K, Togawa D, Oha F. A minimum 10-year follow-up of posterior dynamic stabilization using Graf artificial ligament. Spine. 2007;32(18):1992-6; discussion 7.http://dx.doi.org/10.1097/BRS.0b013e318133faae. Additional Declarations No competing interests reported. Supplementary Files SupplementaryMaterials.docx Cite Share Download PDF Status: Under Review Version 1 posted Editorial decision: Revision requested 24 Nov, 2025 Reviews received at journal 14 Nov, 2025 Reviewers agreed at journal 04 Nov, 2025 Reviews received at journal 04 Nov, 2025 Reviewers agreed at journal 03 Nov, 2025 Reviewers invited by journal 03 Nov, 2025 Editor assigned by journal 29 Aug, 2025 Submission checks completed at journal 29 Aug, 2025 First submitted to journal 19 Aug, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {\"props\":{\"pageProps\":{\"initialData\":{\"identity\":\"rs-7410485\",\"acceptedTermsAndConditions\":true,\"allowDirectSubmit\":false,\"archivedVersions\":[],\"articleType\":\"Research Article\",\"associatedPublications\":[],\"authors\":[{\"id\":539891162,\"identity\":\"167ac513-72b7-48f6-80ca-b9008388a9d1\",\"order_by\":0,\"name\":\"Jiajun Song\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"Tianjin Medical University General Hospital\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Jiajun\",\"middleName\":\"\",\"lastName\":\"Song\",\"suffix\":\"\"},{\"id\":539891163,\"identity\":\"bf809244-b200-4392-99f2-1d54df94cdbb\",\"order_by\":1,\"name\":\"Jiawei Du\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"Tianjin Medical University General Hospital\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Jiawei\",\"middleName\":\"\",\"lastName\":\"Du\",\"suffix\":\"\"},{\"id\":539891164,\"identity\":\"adb56772-675b-4df8-8550-a1924340e526\",\"order_by\":2,\"name\":\"Shengwei Liu\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"Tianjin Medical University General Hospital\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Shengwei\",\"middleName\":\"\",\"lastName\":\"Liu\",\"suffix\":\"\"},{\"id\":539891165,\"identity\":\"f1f147a5-a333-419e-b356-29077a45931d\",\"order_by\":3,\"name\":\"Junyu Chen\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"Tianjin Medical University General Hospital\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Junyu\",\"middleName\":\"\",\"lastName\":\"Chen\",\"suffix\":\"\"},{\"id\":539891166,\"identity\":\"b2ab2e02-5428-427d-bfe3-fc82f1cd4992\",\"order_by\":4,\"name\":\"Di Zhang\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"Tianjin Medical University General Hospital\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Di\",\"middleName\":\"\",\"lastName\":\"Zhang\",\"suffix\":\"\"},{\"id\":539891167,\"identity\":\"837c011d-5e56-4a14-9c33-a2cde260a5ed\",\"order_by\":5,\"name\":\"Shi-Qing Feng\",\"email\":\"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAxUlEQVRIiWNgGAWjYFACxgdAwgbC5iFOC7MBkEiTIFnLYRK0yLcnMz4u+HW+TndGAuODt20M8uaEtBicecxsPLPvtoTZjQRmw7ltDIY7Gwhpkcg/Js3bA9bCJs3bxpBgcICQw2Yks//m7TkH0gJkEKOF4UYyGzPPjwNgW5iJ0gLyizRvQ7LktjMPmyXnnJMw3EDQYcAQ+8zzx47f7HjywQ9vymzkCTuMIQGYANpADMYGICFBUD1EC8MfYhSOglEwCkbBiAUAamg95Hd0H0sAAAAASUVORK5CYII=\",\"orcid\":\"\",\"institution\":\"Tianjin Medical University General Hospital\",\"correspondingAuthor\":true,\"prefix\":\"\",\"firstName\":\"Shi-Qing\",\"middleName\":\"\",\"lastName\":\"Feng\",\"suffix\":\"\"}],\"badges\":[],\"createdAt\":\"2025-08-19 16:23:14\",\"currentVersionCode\":1,\"declarations\":\"\",\"doi\":\"10.21203/rs.3.rs-7410485/v1\",\"doiUrl\":\"https://doi.org/10.21203/rs.3.rs-7410485/v1\",\"draftVersion\":[],\"editorialEvents\":[],\"editorialNote\":\"\",\"failedWorkflow\":false,\"files\":[{\"id\":95844643,\"identity\":\"e6945979-d4f8-4c66-bdaa-2658e9482abf\",\"added_by\":\"auto\",\"created_at\":\"2025-11-13 14:45:33\",\"extension\":\"docx\",\"order_by\":0,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":24475291,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"Manuscript.docx\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7410485/v1/c5119896192eb96b764bf0bc.docx\"},{\"id\":95844631,\"identity\":\"981cb586-552f-4cee-a3af-0860f83dbe69\",\"added_by\":\"auto\",\"created_at\":\"2025-11-13 14:45:33\",\"extension\":\"json\",\"order_by\":1,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":7960,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"415aba8de6b64e5a8581bec29023373c.json\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7410485/v1/f5b6892c112df239ae76e529.json\"},{\"id\":96240186,\"identity\":\"ff064e61-a87a-4ed2-a751-9732303ad3b7\",\"added_by\":\"auto\",\"created_at\":\"2025-11-19 07:08:33\",\"extension\":\"docx\",\"order_by\":2,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":1519686,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"SupplementaryMaterials.docx\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7410485/v1/932fa0b9b8a080bb7b865433.docx\"},{\"id\":95844637,\"identity\":\"32e69066-7ff4-4909-aae4-ea58b19e11a5\",\"added_by\":\"auto\",\"created_at\":\"2025-11-13 14:45:33\",\"extension\":\"xml\",\"order_by\":3,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":89188,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"415aba8de6b64e5a8581bec29023373c1enriched.xml\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7410485/v1/ad1352e648d2d577bca061d6.xml\"},{\"id\":96240031,\"identity\":\"dfba5449-aff3-49e0-817f-441582b1abf1\",\"added_by\":\"auto\",\"created_at\":\"2025-11-19 07:08:17\",\"extension\":\"jpeg\",\"order_by\":4,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":12842199,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"floatimage1.jpeg\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7410485/v1/51fe652e0f66d9ad0d1bb39a.jpeg\"},{\"id\":95844632,\"identity\":\"d67b9852-eea3-444d-9cc5-bf4cd2b25b3c\",\"added_by\":\"auto\",\"created_at\":\"2025-11-13 14:45:33\",\"extension\":\"jpeg\",\"order_by\":5,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":5072144,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"floatimage2.jpeg\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7410485/v1/0f835ade298931bc5c92c13f.jpeg\"},{\"id\":96240807,\"identity\":\"73ca0b41-48e4-49c4-a612-7d5177de2463\",\"added_by\":\"auto\",\"created_at\":\"2025-11-19 07:09:36\",\"extension\":\"jpeg\",\"order_by\":6,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":6470746,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"floatimage3.jpeg\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7410485/v1/cabf9f53826256fc8be01afe.jpeg\"},{\"id\":95844630,\"identity\":\"5b610969-8e34-47d1-9adc-56aec6ab4c4a\",\"added_by\":\"auto\",\"created_at\":\"2025-11-13 14:45:32\",\"extension\":\"png\",\"order_by\":7,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":837629,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"Onlinefloatimage1.png\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7410485/v1/6584088d3ab628c524393b33.png\"},{\"id\":96240238,\"identity\":\"fc188411-56bf-4ef8-b94e-287f977bf718\",\"added_by\":\"auto\",\"created_at\":\"2025-11-19 07:08:37\",\"extension\":\"png\",\"order_by\":8,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":564647,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"Onlinefloatimage2.png\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7410485/v1/ac34cd39bee35356d7fa5ce7.png\"},{\"id\":96241363,\"identity\":\"9d5be091-53b5-4b43-8629-946a002f3066\",\"added_by\":\"auto\",\"created_at\":\"2025-11-19 07:10:37\",\"extension\":\"png\",\"order_by\":9,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":934031,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"Onlinefloatimage3.png\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7410485/v1/5bf2e6afb23cf2fc50240ee5.png\"},{\"id\":95844639,\"identity\":\"961d14d5-4b74-4a01-b49a-50f6391b3b5e\",\"added_by\":\"auto\",\"created_at\":\"2025-11-13 14:45:33\",\"extension\":\"xml\",\"order_by\":10,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":84188,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"415aba8de6b64e5a8581bec29023373c1structuring.xml\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7410485/v1/12615603ac6df00bf9a7e7a1.xml\"},{\"id\":95844641,\"identity\":\"8fb50d7b-5717-4428-9efb-e16cc0f3c6b9\",\"added_by\":\"auto\",\"created_at\":\"2025-11-13 14:45:33\",\"extension\":\"html\",\"order_by\":11,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":94001,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"earlyproof.html\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7410485/v1/4a17a808ed517dd0e95c180f.html\"},{\"id\":95844628,\"identity\":\"11e8aeb8-3810-40a1-ba8e-5519dc0afc3c\",\"added_by\":\"auto\",\"created_at\":\"2025-11-13 14:45:32\",\"extension\":\"png\",\"order_by\":1,\"title\":\"Figure 1\",\"display\":\"\",\"copyAsset\":false,\"role\":\"figure\",\"size\":195591,\"visible\":true,\"origin\":\"\",\"legend\":\"\\u003cp\\u003eFlowchart presented the step-by-step procedures in deep learning model construction and the predictive performance for four DenseNet121-models. Panel a. the Flowchart by this study. Panel b. the Receiver Operating Characteristic (ROC) curves for DenseNet121-models. Panel c. the Area Under Curve (AUC) for DenseNet121-models. Panel d. the calibration curves for DenseNet121-models for prediction. Panel e. the Decision Curve Analysis (DCA) for DenseNet121-models. Panel f. the ROC curves for DenseNet121-models without spinal region. Panel g. the AUC for DenseNet121-models without spinal region. Panel h. the calibration curves for DenseNet121-models for prediction. Panel i. the DCA for DenseNet121-models without spinal region.\\u003c/p\\u003e\\n\\u003cp\\u003eSVM: Support Vector Machine; RF: Random Forest. ROC: Receiver Operating Characteristic. AUC: Area Under Curve. DCA: Decision Curve Analysis.\\u003c/p\\u003e\",\"description\":\"\",\"filename\":\"1.png\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7410485/v1/2e1a555d206f2fc39c381ef5.png\"},{\"id\":95844629,\"identity\":\"d20db4d6-5d76-4c72-bb33-e1406d6e4aa3\",\"added_by\":\"auto\",\"created_at\":\"2025-11-13 14:45:32\",\"extension\":\"png\",\"order_by\":2,\"title\":\"Figure 2\",\"display\":\"\",\"copyAsset\":false,\"role\":\"figure\",\"size\":139206,\"visible\":true,\"origin\":\"\",\"legend\":\"\\u003cp\\u003eThe predictive performance for the DenseNet121-Stacking Model in Age and Gender Subgroup Analysis. Panel a. the ROC Curve Analysis for Cross-Training and Prediction Between Older/Younger and Gender Groups. Panel b presents a heatmap illustrating the AUC for different groups during cross-validation. Panel c. the calibration curves for DenseNet121-stacking models across various training-prediction groups. Panel e. the DCA for DenseNet121-stacking model, comparing its effectiveness under various training and testing group combinations.\\u003c/p\\u003e\\n\\u003cp\\u003eROC: Receiver Operating Characteristic. AUC: Area Under Curve. DCA: Decision Curve Analysis. F: Female group, M: Male group, Y: Young group, E: Elderly group, EF: Elderly female group, YM: Young male group, EM: Elderly male group, YF: Young female group.\\u003c/p\\u003e\",\"description\":\"\",\"filename\":\"2.png\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7410485/v1/4fe51417963cd8e3d36fa2c9.png\"},{\"id\":95844627,\"identity\":\"fd074300-db28-4aec-9df8-83d63b77ade5\",\"added_by\":\"auto\",\"created_at\":\"2025-11-13 14:45:32\",\"extension\":\"png\",\"order_by\":3,\"title\":\"Figure 3\",\"display\":\"\",\"copyAsset\":false,\"role\":\"figure\",\"size\":266012,\"visible\":true,\"origin\":\"\",\"legend\":\"\\u003cp\\u003eRegion of Interest (ROI) Analysis Based on Grad-CAM Heatmaps. Panel a. Proportional distribution of anatomical structures in Grad-CAM heatmap regions between lumbar spine instability (LSI) and non-lumbar spine instability (Non-LSI) populations. Panel b. The Grad-CAM heatmap showed activation around the Intervertebral Space. Panel c. The Grad-CAM heatmap showed activation around Osteophytes. Panel d. The Grad-CAM heatmap showed activation around Facet Joints. Panel e. ROC curves based on statistical analysis of anatomical structures in LSI and Non-LSI populations. Panel f. the AUC for ROC Curves in Panel e.\\u003c/p\\u003e\\n\\u003cp\\u003eROI: Region of Interest ROC: Receiver Operating Characteristic. AUC: Area Under Curve. LSI: Lumbar Spine Instability. Non-LSI: Non- Lumbar Spine Instability.\\u003c/p\\u003e\",\"description\":\"\",\"filename\":\"3.png\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7410485/v1/c8eee2d8249e20c63d9fedfb.png\"},{\"id\":96363045,\"identity\":\"b62c72d1-803d-423c-8a24-2dc777231bec\",\"added_by\":\"auto\",\"created_at\":\"2025-11-20 10:03:54\",\"extension\":\"pdf\",\"order_by\":0,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"manuscript-pdf\",\"size\":1362686,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"manuscript.pdf\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7410485/v1/c4f10a20-bc46-4fbb-93cd-0c0f840ab943.pdf\"},{\"id\":96241644,\"identity\":\"d44003b8-acd3-46ef-98c3-a7281db31cf8\",\"added_by\":\"auto\",\"created_at\":\"2025-11-19 07:11:11\",\"extension\":\"docx\",\"order_by\":0,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"supplement\",\"size\":1519686,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"SupplementaryMaterials.docx\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7410485/v1/4abe2851d8f7fba7ada3b7f6.docx\"}],\"financialInterests\":\"No competing interests reported.\",\"formattedTitle\":\"Deep Learning for Predicting Lumbar Instability Using Neutral Lateral Lumbar Radiographs: A Retrospective Study\",\"fulltext\":[{\"header\":\"Introduction\",\"content\":\"\\u003cp\\u003eLumbar segmental instability (LSI) accounts for up to 57% of chronic low back pain and, if untreated, may progress to degeneration requiring fusion[\\u003cspan citationid=\\\"CR1\\\" class=\\\"CitationRef\\\"\\u003e1\\u003c/span\\u003e]. Diagnosis mainly relies on flexion\\u0026ndash;extension radiographs measuring intervertebral range of motion (IROM) and sagittal translation (ΔST) to detect abnormal movement[\\u003cspan citationid=\\\"CR2\\\" class=\\\"CitationRef\\\"\\u003e2\\u003c/span\\u003e]. However, this method depends on manual landmarking, introducing subjectivity and low efficiency. Since the flexion-extension maneuver is difficult to standardize, the reproducibility of results is poor[\\u003cspan citationid=\\\"CR3\\\" class=\\\"CitationRef\\\"\\u003e3\\u003c/span\\u003e]. Moreover, flexion-extension radiography typically requires patients to visit specialized medical centers equipped with dedicated devices and experienced clinicians, resulting in a high threshold for clinical implementation. Extra radiation further limits early screening and acceptance, reducing early detection opportunities.\\u003c/p\\u003e\\u003cp\\u003eTherefore, a safe and cost-effective alternative is needed. Nearly all hospitals are equipped to perform routine neutral lateral lumbar radiography. These routinely acquired images may contain latent imaging cues related to LSI that are imperceptible to the human eye but can be extracted without additional cost, and even analyzed retrospectively. Fully leveraging such widely available imaging data could greatly expand population-level screening for spinal instability. However, identifying and quantifying these \\u0026ldquo;micro-instability\\u0026rdquo; features from single static images remains technically challenging and requires advanced image analysis algorithms to achieve reliable clinical diagnostic performance.\\u003c/p\\u003e\\u003cp\\u003eDeep learning-based computer vision methods offer a promising solution, as they are capable of identifying hidden diagnostic information within medical images. For example, infer cardiovascular risk factors from retinal fundus photograph[\\u003cspan citationid=\\\"CR4\\\" class=\\\"CitationRef\\\"\\u003e4\\u003c/span\\u003e], estimate cervical spinal-cord compression directly from routine cervical radiography[\\u003cspan citationid=\\\"CR5\\\" class=\\\"CitationRef\\\"\\u003e5\\u003c/span\\u003e], identify facial phenotypes of genetic disorders from ordinary photographs[\\u003cspan citationid=\\\"CR6\\\" class=\\\"CitationRef\\\"\\u003e6\\u003c/span\\u003e], and detect osteoporosis on routine X-ray [\\u003cspan citationid=\\\"CR7\\\" class=\\\"CitationRef\\\"\\u003e7\\u003c/span\\u003e].\\u003c/p\\u003e\\u003cp\\u003eBuilding on these advances, we developed a deep-learning model that analyses neutral lateral lumbar radiographs to predict LSI. Internal and external tests confirmed the algorithm\\u0026rsquo;s feasibility and diagnostic accuracy. Because neutral lateral radiography is routine for back pain and other diseases, patients can be automatically assessed for LSI use neutral lateral radiography without extra cost or radiation. Hospitals can apply the model on archived radiography to spot undiagnosed high-risk patients and warn them early.\\u003c/p\\u003e\"},{\"header\":\"Materials and Methods\",\"content\":\"\\u003cdiv id=\\\"Sec3\\\" class=\\\"Section2\\\"\\u003e\\u003ch2\\u003eStudy Sample\\u003c/h2\\u003e\\u003cp\\u003eThis single-center retrospective study enrolled patients who underwent lumbar spine radiography at XXX from June 2012 to November 2022. The study was approved by the institutional review board of XXX (Approval No. XXX). Anteroposterior, neutral lateral, hyperextension, and hyperflexion radiographs in DICOM format were collected. Patient identifiers were removed using the MicroDicom anonymization tool. Written informed consent was obtained from all eligible participants. Inclusion criteria: (1) high-quality radiographs, (2) all three views available, and (3) no prior spinal surgery. Exclusion criteria included: (1) poor image quality, (2) severe scoliosis, (3) prior spinal surgery, or (4) anatomical variants like lumbarization or sacralization. Ultimately, 1,588 patients were included and split into a training set (1,107 cases: 315 positive, 792 negative) and a testing set (481 cases: 166 positive, 315 negative). Model training and hyperparameter tuning were performed only on the training set to avoid data leakage.\\u003c/p\\u003e\\u003c/div\\u003e\\n\\u003ch3\\u003eSample Labeling\\u003c/h3\\u003e\\n\\u003cp\\u003eLSI was defined per Nachemson\\u0026rsquo;s radiological criteria.[\\u003cspan citationid=\\\"CR2\\\" class=\\\"CitationRef\\\"\\u003e2\\u003c/span\\u003e]. IROM and ΔST were measured at each level from L1\\u0026ndash;2 to L5\\u0026ndash;S1 (Supplementary Fig.\\u0026nbsp;1). The method for measuring ΔST followed prior literature[\\u003cspan citationid=\\\"CR8\\\" class=\\\"CitationRef\\\"\\u003e8\\u003c/span\\u003e]. A histogram presents the distribution of IROM and ΔST across levels. Instability was diagnosed if any segment met either: (i) IROM\\u0026thinsp;\\u0026ge;\\u0026thinsp;10\\u0026deg; (L1\\u0026ndash;2 to L4\\u0026ndash;5) or \\u0026ge;\\u0026thinsp;20\\u0026deg; (L5\\u0026ndash;S1); or (ii) ΔST\\u0026thinsp;\\u0026ge;\\u0026thinsp;3 mm (L1\\u0026ndash;2 to L4\\u0026ndash;5) or \\u0026ge;\\u0026thinsp;4 mm (L5\\u0026ndash;S1).[\\u003cspan citationid=\\\"CR2\\\" class=\\\"CitationRef\\\"\\u003e2\\u003c/span\\u003e]. Radiographic diagnoses were conducted by three spine surgeons. Each case was independently evaluated by at least two surgeons. If their assessments differed, the final diagnosis was established by vote after discussion among all three reviewers.\\u003c/p\\u003e\\n\\u003ch3\\u003eData Preprocessing\\u003c/h3\\u003e\\n\\u003cp\\u003eNeutral lateral radiographs were cropped to include: (1) from the lower one-third of T12 to the upper one-third of S1, (2) anteriorly to the vertebral edge (including osteophytes), and (3) posteriorly to the tip of the spinous process. This region was selected based on anatomical relevance to lumbar instability. Images underwent Z-score normalization (mean\\u0026thinsp;=\\u0026thinsp;0, SD\\u0026thinsp;=\\u0026thinsp;1) and channel-wise scaling to 0\\u0026ndash;1 to enhance consistency and model performance. Data augmentation\\u0026mdash;including horizontal flipping, rotation, shifting, zooming, and shearing\\u0026mdash;was applied randomly to expand the dataset and reduce overfitting.\\u003c/p\\u003e\\n\\u003ch3\\u003eAnalysis 1: Classifying LSI based on DenseNet121 Model\\u003c/h3\\u003e\\n\\u003cp\\u003eIn this study, we used DenseNet121[\\u003cspan citationid=\\\"CR9\\\" class=\\\"CitationRef\\\"\\u003e9\\u003c/span\\u003e], pretrained on ImageNet, as a feature extractor. Features from the fully connected layer were input into three classifiers: Support Vector Machine (SVM), Random Forests (RF), and Softmax, with SVM and RF being widely used in classification tasks.\\u003c/p\\u003e\\u003cp\\u003eA 10-fold cross-validation was applied within the training set to tune hyperparameters and prevent overfitting. Hyperparameter optimization was performed via grid search with 10-fold cross-validation (see Supplementary Table \\u003cspan refid=\\\"MOESM1\\\" class=\\\"InternalRef\\\"\\u003eS1\\u003c/span\\u003e for hyperparameter). The optimal configuration was selected based on area under the curve (AUC) maximization. Figure\\u0026nbsp;\\u003cspan refid=\\\"Fig1\\\" class=\\\"InternalRef\\\"\\u003e1\\u003c/span\\u003e illustrates the overall workflow. The final model was tested on a held-out test set. Classifier outputs were in probability form, and a stacking ensemble method with soft voting was implemented by averaging these probabilities.\\u003c/p\\u003e\\u003cp\\u003e\\u003c/p\\u003e\\u003cp\\u003eTo evaluate model performance, we computed metrics including AUC, accuracy, sensitivity, specificity, precision, recall, and F1-score Calibration plots were also generated to assess the alignment of predicted probabilities with actual outcomes.\\u003c/p\\u003e\\n\\u003ch3\\u003eAnalysis 2: Sensitivity analyses\\u003c/h3\\u003e\\n\\u003cp\\u003eTo assess robustness of the diagnostic model in analysis 1, we performed four sensitivity analyses to ensure the model overly sensitive to data fluctuations.\\u003c/p\\u003e\\u003cp\\u003e(1) Spinal region specificity: Models trained on masked spinal vs. non-spinal regions showed whether diagnostic features originated from spinal structures.\\u003c/p\\u003e\\u003cp\\u003e(2) Age subgroup generalization: After 1:1 age-matching (912 samples), participants were stratified as elderly (\\u0026ge;\\u0026thinsp;60) or younger (\\u0026lt;\\u0026thinsp;60). Models were trained/tested within and across subgroups to assess age effects. Performance was compared by ROC and DeLong test.\\u003c/p\\u003e\\u003cp\\u003e(3) Gender subgroup generalization: Models were trained/tested within male and female groups, and cross-sex validation examined whether predictions reflected LSI features rather than sex differences. ROC and DeLong test evaluated AUCs.\\u003c/p\\u003e\\u003cp\\u003e(4) Age-gender interaction: Extreme subgroup combinations (elderly male vs. young female; elderly female vs. young male) further tested population-specific effects on performance, assessed by ROC, DeLong test, and standard metrics.\\u003c/p\\u003e\\u003cdiv id=\\\"Sec8\\\" class=\\\"Section2\\\"\\u003e\\u003ch2\\u003eAnalysis 3: Grad-CAM-Guided Exploration of Interpretable Features for LSI Classification\\u003c/h2\\u003e\\u003cp\\u003eGrad-CAM (Gradient-weighted Class Activation Mapping) is a visualization technique that utilizes the gradient information from the final convolutional layer of a convolutional neural network (CNN) to generate class activation maps, highlighting image regions most relevant to the classification decision.\\u003c/p\\u003e\\u003cp\\u003eIn Analyses 1 and 2, we applied Grad-CAM to each trained model to visualize the spatial attention patterns and statistically assess whether the highlighted anatomical regions were associated with LSI. This approach also aimed to identify potential radiographic indicators of LSI.\\u003c/p\\u003e\\u003cp\\u003eUsing the resulting heatmaps, we extracted anatomical structures corresponding to the high-activation regions in each case. We quantified their occurrence and evaluated the discriminative power of the three most frequently activated regions in distinguishing between LSI and non-LSI patients.\\u003c/p\\u003e\\u003cp\\u003eThese features were subsequently used as input variables for machine learning classifiers, including SVM, K-Nearest Neighbors (KNN), and Logistic Regression (LR), to further assess their predictive utility in the classification of LSI.\\u003c/p\\u003e\\u003c/div\\u003e\"},{\"header\":\"Results\",\"content\":\"\\u003cdiv id=\\\"Sec10\\\" class=\\\"Section2\\\"\\u003e\\n \\u003ch2\\u003eBaseline Patient Characteristics\\u003c/h2\\u003e\\n \\u003cp\\u003eWe assessed 1,588 patients. Table \\u003cspan class=\\\"InternalRef\\\"\\u003e1\\u003c/span\\u003e summarizes cohort characteristics. In the LSI group, mean age was 58.25 years and 44% were male.\\u003c/p\\u003e\\n \\u003cdiv class=\\\"gridtable\\\"\\u003e\\n \\u003ctable id=\\\"Tab1\\\" border=\\\"1\\\"\\u003e\\n \\u003ccaption language=\\\"En\\\"\\u003e\\n \\u003cdiv class=\\\"CaptionNumber\\\"\\u003eTable 1\\u003c/div\\u003e\\n \\u003cdiv class=\\\"CaptionContent\\\"\\u003e\\n \\u003cp\\u003e, Baseline Clinical Characteristics and Outcomes of the 1588 Patient Cohort.\\u003c/p\\u003e\\n \\u003c/div\\u003e\\n \\u003c/caption\\u003e\\n \\u003cthead\\u003e\\n \\u003ctr\\u003e\\n \\u003cth align=\\\"left\\\"\\u003e\\n \\u003cp\\u003eCharacteristic\\u003c/p\\u003e\\n \\u003c/th\\u003e\\n \\u003cth align=\\\"left\\\" colspan=\\\"2\\\"\\u003e\\n \\u003cp\\u003eMean\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;standard deviation or frequency (proportion)\\u003c/p\\u003e\\n \\u003c/th\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003cth align=\\\"left\\\"\\u003e\\u0026nbsp;\\u003c/th\\u003e\\n \\u003cth align=\\\"left\\\"\\u003e\\n \\u003cp\\u003eNon-LSI Group\\u003c/p\\u003e\\n \\u003c/th\\u003e\\n \\u003cth align=\\\"left\\\"\\u003e\\n \\u003cp\\u003eLSI Group\\u003c/p\\u003e\\n \\u003c/th\\u003e\\n \\u003c/tr\\u003e\\n \\u003c/thead\\u003e\\n \\u003ctbody\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd align=\\\"left\\\"\\u003e\\n \\u003cp\\u003eAge(yr)\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"left\\\"\\u003e\\n \\u003cp\\u003e50.45\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;13.85\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"left\\\"\\u003e\\n \\u003cp\\u003e58.25\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;13.83\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd align=\\\"left\\\"\\u003e\\n \\u003cp\\u003eMale sex\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"left\\\"\\u003e\\n \\u003cp\\u003e540(48.78)\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"left\\\"\\u003e\\n \\u003cp\\u003e212(44.07)\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd align=\\\"left\\\"\\u003e\\n \\u003cp\\u003ePain episode (First)\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"left\\\"\\u003e\\n \\u003cp\\u003e889(80.31)\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"left\\\"\\u003e\\n \\u003cp\\u003e360(74.84)\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd align=\\\"left\\\"\\u003e\\n \\u003cp\\u003ePain radiation\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"left\\\"\\u003e\\n \\u003cp\\u003e277(25.02)\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"left\\\"\\u003e\\n \\u003cp\\u003e98(20.37)\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd align=\\\"left\\\"\\u003e\\n \\u003cp\\u003eTrauma history\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"left\\\"\\u003e\\n \\u003cp\\u003e22(1.99)\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"left\\\"\\u003e\\n \\u003cp\\u003e8(1.66)\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd align=\\\"left\\\"\\u003e\\n \\u003cp\\u003eDrinking\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"left\\\"\\u003e\\n \\u003cp\\u003e157(14.18)\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"left\\\"\\u003e\\n \\u003cp\\u003e61(12.68)\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd align=\\\"left\\\"\\u003e\\n \\u003cp\\u003eSmoking status\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"left\\\"\\u003e\\n \\u003cp\\u003e326(29.45)\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"left\\\"\\u003e\\n \\u003cp\\u003e137(28.48)\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd align=\\\"left\\\"\\u003e\\n \\u003cp\\u003eDiabetes\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"left\\\"\\u003e\\n \\u003cp\\u003e153(13.82)\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"left\\\"\\u003e\\n \\u003cp\\u003e52(10.81)\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd align=\\\"left\\\"\\u003e\\n \\u003cp\\u003eHypertension\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"left\\\"\\u003e\\n \\u003cp\\u003e164(14.81)\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"left\\\"\\u003e\\n \\u003cp\\u003e92(19.13)\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd align=\\\"left\\\"\\u003e\\n \\u003cp\\u003eHeart disease\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"left\\\"\\u003e\\n \\u003cp\\u003e50(4.52)\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"left\\\"\\u003e\\n \\u003cp\\u003e45(9.36)\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003c/tbody\\u003e\\n \\u003ctfoot\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd colspan=\\\"3\\\"\\u003eContinuous variables are presented as the mean\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;standard deviation, while categorical variables are reported as frequencies with corresponding percentages (%). LSI: Lumbar Spine Instability. Non-LSI: Non- Lumbar Spine Instability.\\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003c/tfoot\\u003e\\n \\u003c/table\\u003e\\n \\u003c/div\\u003e\\n\\u003c/div\\u003e\\n\\u003cdiv id=\\\"Sec11\\\" class=\\\"Section2\\\"\\u003e\\n \\u003ch2\\u003eLSI Prediction Based on DenseNet121 Model and Stacking Learning\\u003c/h2\\u003e\\n \\u003cp\\u003eThe DenseNet121-SVM model achieved an AUC of 0.750, accuracy of 0.740, sensitivity of 0.470, and specificity of 0.880; The DenseNet121-RF model achieved an AUC of 0.755, accuracy of 0.694, sensitivity of 0.741, and specificity of 0.670.; The DenseNet121-Softmax model achieved an AUC of 0.795, accuracy of 0.738, sensitivity of 0.657, and specificity of 0.781.\\u003c/p\\u003e\\n \\u003cp\\u003eAmong these, the DenseNet121-stacking model showed the best performance, with an AUC of 0.820, accuracy of 0.760, sensitivity of 0.610, and specificity of 0.840. The calibration curve showed good agreement with the ideal reference line. Decision curve analysis (DCA) demonstrated clinical utility across a range of threshold probabilities, particularly for the stacking model (Table \\u003cspan class=\\\"InternalRef\\\"\\u003e2\\u003c/span\\u003e, Fig. \\u003cspan class=\\\"InternalRef\\\"\\u003e1\\u003c/span\\u003e).\\u003c/p\\u003e\\n \\u003cdiv class=\\\"gridtable\\\"\\u003e\\n \\u003ctable id=\\\"Tab2\\\" border=\\\"1\\\"\\u003e\\n \\u003ccaption language=\\\"En\\\"\\u003e\\n \\u003cdiv class=\\\"CaptionNumber\\\"\\u003eTable 2\\u003c/div\\u003e\\n \\u003cdiv class=\\\"CaptionContent\\\"\\u003e\\n \\u003cp\\u003e, Comparison of the AUCs, Accuracies, Sensitivities, Specificities, Precisions, Recalls and F1-scores for four DenseNet121-models.\\u003c/p\\u003e\\n \\u003c/div\\u003e\\n \\u003c/caption\\u003e\\n \\u003cthead\\u003e\\n \\u003ctr\\u003e\\n \\u003cth align=\\\"left\\\"\\u003e\\n \\u003cp\\u003eModelName\\u003c/p\\u003e\\n \\u003c/th\\u003e\\n \\u003cth align=\\\"left\\\"\\u003e\\n \\u003cp\\u003eAcc\\u003c/p\\u003e\\n \\u003c/th\\u003e\\n \\u003cth align=\\\"left\\\"\\u003e\\n \\u003cp\\u003eAUC\\u003c/p\\u003e\\n \\u003c/th\\u003e\\n \\u003cth align=\\\"left\\\"\\u003e\\n \\u003cp\\u003eSensitivity\\u003c/p\\u003e\\n \\u003c/th\\u003e\\n \\u003cth align=\\\"left\\\"\\u003e\\n \\u003cp\\u003eSpecificity\\u003c/p\\u003e\\n \\u003c/th\\u003e\\n \\u003cth align=\\\"left\\\"\\u003e\\n \\u003cp\\u003ePrecision\\u003c/p\\u003e\\n \\u003c/th\\u003e\\n \\u003cth align=\\\"left\\\"\\u003e\\n \\u003cp\\u003eRecall\\u003c/p\\u003e\\n \\u003c/th\\u003e\\n \\u003cth align=\\\"left\\\"\\u003e\\n \\u003cp\\u003eF1\\u003c/p\\u003e\\n \\u003c/th\\u003e\\n \\u003c/tr\\u003e\\n \\u003c/thead\\u003e\\n \\u003ctbody\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd align=\\\"left\\\"\\u003e\\n \\u003cp\\u003eDenseNet121-SVM\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.74\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.75\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.47\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.883\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.678\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.47\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.555\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd align=\\\"left\\\"\\u003e\\n \\u003cp\\u003eDenseNet121-RF\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.694\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.755\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.741\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.67\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.542\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.741\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.626\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd align=\\\"left\\\"\\u003e\\n \\u003cp\\u003eDenseNet121-Softmax\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.738\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.795\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.657\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.781\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.612\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.657\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.634\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd align=\\\"left\\\"\\u003e\\n \\u003cp\\u003eDenseNet121-Stacking\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.763\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.823\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.614\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.841\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.671\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.614\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.642\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003c/tbody\\u003e\\n \\u003ctfoot\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd colspan=\\\"8\\\"\\u003eAUC: Area Under Curve; SVM: Support Vector Machine; RF: Random Forest.\\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003c/tfoot\\u003e\\n \\u003c/table\\u003e\\n \\u003c/div\\u003e\\n\\u003c/div\\u003e\\n\\u003cdiv id=\\\"Sec12\\\" class=\\\"Section2\\\"\\u003e\\n \\u003ch2\\u003eSensitivity analyses\\u003c/h2\\u003e\\n \\u003cp\\u003eSeveral analyses tested the robustness of the stacking model. First, models trained only on non-lumbar regions showed worse performance than those trained on lumbar regions (Table \\u003cspan class=\\\"InternalRef\\\"\\u003e3\\u003c/span\\u003e, Fig. \\u003cspan class=\\\"InternalRef\\\"\\u003e1\\u003c/span\\u003ef\\u0026ndash;i), indicating classification relied on lumbar anatomy. Second, subgroup analyses showed stable performance across age and sex with no significant AUC differences (DeLong test, P\\u0026thinsp;\\u0026gt;\\u0026thinsp;0.05; Fig. \\u003cspan class=\\\"InternalRef\\\"\\u003e2\\u003c/span\\u003ea\\u0026ndash;d, Table \\u003cspan class=\\\"InternalRef\\\"\\u003e4\\u003c/span\\u003e). Finally, cross-validation between age and gender groups showed the model still distinguished LSI from controls even when trained on young males and tested on elderly females, and vice versa, confirming classification was not driven by demographic features but by LSI-related structural patterns (Fig. \\u003cspan class=\\\"InternalRef\\\"\\u003e2\\u003c/span\\u003ea\\u0026ndash;d, Table \\u003cspan class=\\\"InternalRef\\\"\\u003e4\\u003c/span\\u003e).\\u003c/p\\u003e\\n \\u003cdiv class=\\\"gridtable\\\"\\u003e\\n \\u003ctable id=\\\"Tab3\\\" border=\\\"1\\\"\\u003e\\n \\u003ccaption language=\\\"En\\\"\\u003e\\n \\u003cdiv class=\\\"CaptionNumber\\\"\\u003eTable 3\\u003c/div\\u003e\\n \\u003cdiv class=\\\"CaptionContent\\\"\\u003e\\n \\u003cp\\u003e, Comparison of the AUCs, Accuracies, Sensitivities, Specificities, Precisions, Recalls and F1-scores for four DenseNet121-models without the spinal area.\\u003c/p\\u003e\\n \\u003c/div\\u003e\\n \\u003c/caption\\u003e\\n \\u003cthead\\u003e\\n \\u003ctr\\u003e\\n \\u003cth align=\\\"left\\\"\\u003e\\n \\u003cp\\u003eModelName\\u003c/p\\u003e\\n \\u003c/th\\u003e\\n \\u003cth align=\\\"left\\\"\\u003e\\n \\u003cp\\u003eAcc\\u003c/p\\u003e\\n \\u003c/th\\u003e\\n \\u003cth align=\\\"left\\\"\\u003e\\n \\u003cp\\u003eAUC\\u003c/p\\u003e\\n \\u003c/th\\u003e\\n \\u003cth align=\\\"left\\\"\\u003e\\n \\u003cp\\u003eSensitivity\\u003c/p\\u003e\\n \\u003c/th\\u003e\\n \\u003cth align=\\\"left\\\"\\u003e\\n \\u003cp\\u003eSpecificity\\u003c/p\\u003e\\n \\u003c/th\\u003e\\n \\u003cth align=\\\"left\\\"\\u003e\\n \\u003cp\\u003ePrecision\\u003c/p\\u003e\\n \\u003c/th\\u003e\\n \\u003cth align=\\\"left\\\"\\u003e\\n \\u003cp\\u003eRecall\\u003c/p\\u003e\\n \\u003c/th\\u003e\\n \\u003cth align=\\\"left\\\"\\u003e\\n \\u003cp\\u003eF1\\u003c/p\\u003e\\n \\u003c/th\\u003e\\n \\u003c/tr\\u003e\\n \\u003c/thead\\u003e\\n \\u003ctbody\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd align=\\\"left\\\"\\u003e\\n \\u003cp\\u003eDenseNet121-SVM\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.624\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.559\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.289\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.8\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.432\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.289\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.347\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd align=\\\"left\\\"\\u003e\\n \\u003cp\\u003eDenseNet121-RF\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.503\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.583\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.807\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.343\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.393\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.807\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.529\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd align=\\\"left\\\"\\u003e\\n \\u003cp\\u003eDenseNet121-Softmax\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.615\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.571\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.199\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.835\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.388\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.199\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.263\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd align=\\\"left\\\"\\u003e\\n \\u003cp\\u003eDenseNet121-Stacking\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.578\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.583\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.524\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.606\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.412\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.524\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.462\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003c/tbody\\u003e\\n \\u003ctfoot\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd colspan=\\\"8\\\"\\u003eAUC: Area Under Curve; SVM: Support Vector Machine; RF: Random Forest.\\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003c/tfoot\\u003e\\n \\u003c/table\\u003e\\n \\u003c/div\\u003e\\n \\u003cdiv class=\\\"gridtable\\\"\\u003e\\n \\u003cdiv align=\\\"char\\\" class=\\\"colspec\\\"\\u003e\\u003cbr\\u003e\\u003c/div\\u003e\\n \\u003ctable id=\\\"Tab4\\\" border=\\\"1\\\"\\u003e\\n \\u003ccaption language=\\\"En\\\"\\u003e\\n \\u003cdiv class=\\\"CaptionNumber\\\"\\u003eTable 4\\u003c/div\\u003e\\n \\u003cdiv class=\\\"CaptionContent\\\"\\u003e\\n \\u003cp\\u003e, Performance Metrics of the DenseNet121-Stacking Model in Age and Gender Subgroup Analysis.\\u003c/p\\u003e\\n \\u003c/div\\u003e\\n \\u003c/caption\\u003e\\n \\u003cthead\\u003e\\n \\u003ctr\\u003e\\n \\u003cth align=\\\"left\\\"\\u003e\\n \\u003cp\\u003eModelName\\u003c/p\\u003e\\n \\u003c/th\\u003e\\n \\u003cth align=\\\"left\\\"\\u003e\\n \\u003cp\\u003eAcc\\u003c/p\\u003e\\n \\u003c/th\\u003e\\n \\u003cth align=\\\"left\\\"\\u003e\\n \\u003cp\\u003eAUC\\u003c/p\\u003e\\n \\u003c/th\\u003e\\n \\u003cth align=\\\"left\\\"\\u003e\\n \\u003cp\\u003eSensitivity\\u003c/p\\u003e\\n \\u003c/th\\u003e\\n \\u003cth align=\\\"left\\\"\\u003e\\n \\u003cp\\u003eSpecificity\\u003c/p\\u003e\\n \\u003c/th\\u003e\\n \\u003cth align=\\\"left\\\"\\u003e\\n \\u003cp\\u003ePrecision\\u003c/p\\u003e\\n \\u003c/th\\u003e\\n \\u003cth align=\\\"left\\\"\\u003e\\n \\u003cp\\u003eRecall\\u003c/p\\u003e\\n \\u003c/th\\u003e\\n \\u003cth align=\\\"left\\\"\\u003e\\n \\u003cp\\u003eF1\\u003c/p\\u003e\\n \\u003c/th\\u003e\\n \\u003c/tr\\u003e\\n \\u003c/thead\\u003e\\n \\u003ctbody\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd align=\\\"left\\\"\\u003e\\n \\u003cp\\u003eF\\u0026rarr;M\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.754\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.824\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.705\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.788\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.698\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.705\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.701\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd align=\\\"left\\\"\\u003e\\n \\u003cp\\u003eM\\u0026rarr;F\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.77\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.812\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.735\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.791\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.67\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.735\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.701\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd align=\\\"left\\\"\\u003e\\n \\u003cp\\u003eY\\u0026rarr;E\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.716\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.787\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.65\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.757\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.623\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.65\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.636\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd align=\\\"left\\\"\\u003e\\n \\u003cp\\u003eE\\u0026rarr;Y\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.729\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.801\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.761\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.704\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.665\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.761\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.71\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd align=\\\"left\\\"\\u003e\\n \\u003cp\\u003eEF\\u0026rarr;YM\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.736\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.797\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.581\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.835\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.694\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.581\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.632\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd align=\\\"left\\\"\\u003e\\n \\u003cp\\u003eEM\\u0026rarr;YF\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.753\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.806\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.755\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.751\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.668\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.755\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd align=\\\"char\\\"\\u003e\\n \\u003cp\\u003e0.709\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003c/tbody\\u003e\\n \\u003ctfoot\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd colspan=\\\"8\\\"\\u003eAUC: Area Under Curve. F: Female group, M: Male group, Y: Young group, E: Elderly group, EF: Elderly female group, YM: Young male group, EM: Elderly male group, YF: Young female group.\\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003c/tfoot\\u003e\\n \\u003c/table\\u003e\\n \\u003c/div\\u003e\\n\\u003c/div\\u003e\\n\\u003cdiv id=\\\"Sec13\\\" class=\\\"Section2\\\"\\u003e\\n \\u003ch2\\u003eLSI associated structural region identified by Grad-CAM and its predictive utility\\u003c/h2\\u003e\\n \\u003cp\\u003eTo interpret the \\u0026ldquo;black box\\u0026rdquo; model, Grad-CAM highlighted key anatomical regions including facet joints, discs, and osteophytes (Fig. \\u003cspan class=\\\"InternalRef\\\"\\u003e3\\u003c/span\\u003ea\\u0026ndash;d). In LSI patients, these occurred most frequently (facet joint 34.1%, disc 27.0%, osteophyte 25.4%). This indicates that radiographic signs in these three regions contributed most to classification. We therefore evaluated three features\\u0026mdash;facet joint hypertrophy, disc height reduction, and osteophyte formation. Individually, their predictive value was limited, but when combined in machine learning models, classification improved (max AUC 0.749; Fig. \\u003cspan class=\\\"InternalRef\\\"\\u003e3\\u003c/span\\u003ef). These findings suggest the coexistence of these signs is strongly associated with LSI.\\u003c/p\\u003e\\n\\u003c/div\\u003e\"},{\"header\":\"Discussion\",\"content\":\"\\u003cp\\u003eThis study yielded three main findings: (1) Deep learning applied to neutral lateral lumbar radiographs accurately distinguished LSI from non-LSI cases, with the DenseNet121-stacking model performing best (AUC\\u0026thinsp;=\\u0026thinsp;0.820, accuracy\\u0026thinsp;=\\u0026thinsp;0.760, sensitivity\\u0026thinsp;=\\u0026thinsp;0.610, specificity\\u0026thinsp;=\\u0026thinsp;0.840); (2) Sensitivity analyses confirmed the model\\u0026rsquo;s robustness across age and gender groups; (3) Grad-CAM revealed that combining three imaging features\\u0026mdash;facet joint hypertrophy, osteophytes, and reduced disc height\\u0026mdash;provided added diagnostic value.\\u003c/p\\u003e\\u003cp\\u003ePrior research has explored diverse LSI predictors to improve diagnostic efficiency. Jesse L. Even et al. reported that interspinous fluid on MRI predicted\\u0026thinsp;\\u0026gt;\\u0026thinsp;3 mm instability with PPV 69.0%. Conversely, absence of facet joint fluid had PPV 75.6% for ruling out \\u0026gt;\\u0026thinsp;3 mm instability.[\\u003cspan citationid=\\\"CR10\\\" class=\\\"CitationRef\\\"\\u003e10\\u003c/span\\u003e] Tahere Seyedhoseinpoor et al. combined clinical indicators with examination; the best model (weight, lumbar lordosis, prone segmental instability) achieved AUC 0.66[\\u003cspan citationid=\\\"CR11\\\" class=\\\"CitationRef\\\"\\u003e11\\u003c/span\\u003e]. Although flexion\\u0026ndash;extension radiography is the gold standard, its use still depends on clinical judgment. In our evaluation, the model achieved AUC 0.85, suggesting neutral-position AI screening may exceed symptom/examination-only assessments and complement traditional evaluation.\\u003c/p\\u003e\\u003cp\\u003eRadiographs hold broader diagnostic potential, as shown in osteoporosis and fracture studies. In a meta-analysis of osteoporosis diagnostics, 17/40 studies used radiographs[\\u003cspan citationid=\\\"CR12\\\" class=\\\"CitationRef\\\"\\u003e12\\u003c/span\\u003e]. Sena Goral et al. developed an interpretable capsule-network system to identify 68 vertebral landmarks and compute Cobb angles[\\u003cspan citationid=\\\"CR13\\\" class=\\\"CitationRef\\\"\\u003e13\\u003c/span\\u003e]. Li-Wei Cheng et al. designed a model capable of detecting and classifying vertebral fractures based on radiograph[\\u003cspan citationid=\\\"CR14\\\" class=\\\"CitationRef\\\"\\u003e14\\u003c/span\\u003e]. Compared with CT, radiography is cheaper, faster, and lower in radiation. Leveraging these advantages could expand population-level screening.\\u003c/p\\u003e\\u003cp\\u003eWe further utilized Grad-CAM visualization technology to identify three anatomical structures closely associated with LSI. The identification of these structures is consistent with existing clinical evidence, offering strong support for our research. Disc space narrowing is recognized as a key sign in advanced stabilization phases[\\u003cspan citationid=\\\"CR15\\\" class=\\\"CitationRef\\\"\\u003e15\\u003c/span\\u003e], with disc degeneration linked to instability in biomechanical and clinical studies[\\u003cspan citationid=\\\"CR16\\\" class=\\\"CitationRef\\\"\\u003e16\\u003c/span\\u003e]. Another notable indirect radiographic marker linked to LSI is the traction spur [\\u003cspan citationid=\\\"CR17\\\" class=\\\"CitationRef\\\"\\u003e17\\u003c/span\\u003e], typically 2\\u0026ndash;3 mm from the endplate and oriented horizontally\\u0026mdash;are thought to arise from tensile forces on the periosteum by Sharpey fibers or the anterior longitudinal ligament during instability episodes. Claw osteophytes, originating from the vertebral apophysis near the disc margin and curving toward the adjacent vertebra, likely reflect compressive responses and stability restoration. These features may represent different stages of a shared pathological process and often coexist along the vertebral rim[\\u003cspan citationid=\\\"CR18\\\" class=\\\"CitationRef\\\"\\u003e18\\u003c/span\\u003e]. Furthermore, the facet joints guide and limit vertebral movement while aiding in load distribution across the spine[\\u003cspan citationid=\\\"CR19\\\" class=\\\"CitationRef\\\"\\u003e19\\u003c/span\\u003e], facet joints guide vertebral motion and share load distribution; dysfunction can compromise stability. When multiple anatomical markers were incorporated simultaneously as variables into the machine learning model, the classification performance improved significantly compared to models built on a single anatomical structure alone. This finding suggests that LSI is a multifactorial pathology involving multiple anatomical structures, rather than being caused by changes in any single structure.\\u003c/p\\u003e\\u003cp\\u003eCompared with traditional methods, our approach offers several advantages. First, it uses routine neutral lateral radiographs that are low-cost and low-dose, and the neutral posture is easier to standardize, enabling large-scale screening. Second, reproducible criteria provide greater consistency than expert manual assessments prone to inter-observer variability. Third, the diagnostic performance of our model in identifying lumbar instability based on neutral radiographs surpasses that of physical examinations, manual image interpretation by clinicians, and other conventional approaches. Finally, full automation enables rapid, scalable screening without manual image processing.\\u003c/p\\u003e\\u003cp\\u003eLSI diagnostic criteria remain inconsistent. Some define instability as sagittal translation\\u0026thinsp;\\u0026ge;\\u0026thinsp;4 mm or \\u0026ge;\\u0026thinsp;8%, rotation\\u0026thinsp;\\u0026ge;\\u0026thinsp;10\\u0026deg; (L1\\u0026ndash;L5) or \\u0026ge;\\u0026thinsp;20\\u0026deg; (L5\\u0026ndash;S1)[\\u003cspan citationid=\\\"CR20\\\" class=\\\"CitationRef\\\"\\u003e20\\u003c/span\\u003e]. We adopted Nachemson\\u0026rsquo;s criteria but lowered the L1\\u0026ndash;L2 to L4\\u0026ndash;L5 translation threshold to \\u0026ge;\\u0026thinsp;3 mm for greater sensitivity, still achieving high accuracy. Grad-CAM confirmed localization to discs, facet joints, and osteophytes. Even 3 mm translation may indicate early degeneration and reduced stability; adopting a more sensitive\\u0026thinsp;\\u0026ge;\\u0026thinsp;3 mm threshold may enable earlier identification and timely conservative interventions (e.g., core stabilization[\\u003cspan citationid=\\\"CR21\\\" class=\\\"CitationRef\\\"\\u003e21\\u003c/span\\u003e]), potentially preventing progression.\\u003c/p\\u003e\\u003cp\\u003eEarly identification and targeted conservative treatment of LSI can markedly improve outcomes. Studies report\\u0026thinsp;\\u0026gt;\\u0026thinsp;50% Oswestry Disability Index scores reduction in 49.1% of patients[\\u003cspan citationid=\\\"CR22\\\" class=\\\"CitationRef\\\"\\u003e22\\u003c/span\\u003e], and individuals with the greatest improvement in physical function also demonstrated relief from negative cognitive patterns related to pain perception[\\u003cspan citationid=\\\"CR23\\\" class=\\\"CitationRef\\\"\\u003e23\\u003c/span\\u003e]. Randomized controlled trials show core stabilization reduces VAS by 45% and improves joint play[\\u003cspan citationid=\\\"CR24\\\" class=\\\"CitationRef\\\"\\u003e24\\u003c/span\\u003e], while in early mild spondylolisthesis (\\u0026lt;\\u0026thinsp;20%), core reinforcement or Graf ligamentoplasty yields a 10-year adjacent segment reoperation rate of 7% versus \\u0026ge;\\u0026thinsp;18% for fusion[\\u003cspan citationid=\\\"CR25\\\" class=\\\"CitationRef\\\"\\u003e25\\u003c/span\\u003e]. Early physical therapy in low back pain lowers risks of advanced imaging, surgery, injections, opioid use, and overall costs. These findings further highlight the importance of early detection and tailored conservative treatment strategies in patients with potential instability.\\u003c/p\\u003e\\u003cp\\u003eThis study has several limitations. First, the dataset size should be further expanded in future studies to ensure more robust validation and greater generalizability of the findings. Therefore, we believe that the accuracy of our algorithm could be further improved with a larger training dataset. Second, single-center data may limit generalizability; external validation across institutions and regions is needed. Third, we used imaging alone; integrating symptoms and examination via NLP from electronic records could enable multimodal models and improve diagnostic accuracy and reliability.\\u003c/p\\u003e\"},{\"header\":\"Conclusion\",\"content\":\"\\u003cp\\u003eIn conclusion, this study developed a deep learning algorithm based on neutral-position lateral lumbar radiographs to predict LSI, achieving high diagnostic accuracy. The model is suitable for early screening of populations at high risk for LSI. By leveraging routinely acquired imaging data, this approach provides an efficient, low-cost, and scalable solution for the early identification of individuals at risk of LSI.\\u003c/p\\u003e\"},{\"header\":\"Declarations\",\"content\":\"\\u003ch2\\u003eAuthor Contribution\\u003c/h2\\u003e\\u003cp\\u003eConceptualization: Jiajun Song, Jiawei Du, Di Zhang, Shiqing Feng.Formal Analysis: Jiajun Song, Jiawei Du.Investigation: Jiajun Song, Shengwei Liu, Junyu Chen.Methodology: Jiajun Song, Jiawei Du, Shengwei Liu, Shiqing Feng.Project Administration: Di Zhang, Shiqing Feng.Writing \\u0026ndash; Original Draft: Jiajun Song, Jiawei Du, Shengwei Liu, Junyu Chen.Writing \\u0026ndash; Review \\u0026amp; Editing: Jiawei Du, Di Zhang, Shiqing Feng.Jiajun Song and Jiawei Du contributed equally to this work.\\u003c/p\\u003e\\u003ch2\\u003eAcknowledgement\\u003c/h2\\u003e\\u003cp\\u003eAcknowledgementsWe thank the Radiology Department of Tianjin Medical University General Hospital for providing access to the anonymized radiographic datasets.\\u003c/p\\u003e\"},{\"header\":\"References\",\"content\":\"\\u003col\\u003e\\n\\u003cli\\u003eHuang H, Young W, Skaper S, et al. Clinical Neurorestorative Therapeutic Guidelines for Spinal Cord Injury (IANR/CANR version 2019). Journal of orthopaedic translation. 2020;20:14-24.http://dx.doi.org/10.1016/j.jot.2019.10.006.\\u003c/li\\u003e\\n\\u003cli\\u003eNachemson A. The Role of Spine Fusion: Question 8. 1981;6(3):306-7.http://dx.doi.org/.\\u003c/li\\u003e\\n\\u003cli\\u003eSmith JS, Shaffrey CI, Ames CP, Lenke LG. Treatment of adult thoracolumbar spinal deformity: past, present, and future. Journal of neurosurgery Spine. 2019;30(5):551-67.http://dx.doi.org/10.3171/2019.1.Spine181494.\\u003c/li\\u003e\\n\\u003cli\\u003ePoplin R, Varadarajan AV, Blumer K, et al. Prediction of cardiovascular risk factors from retinal fundus photographs via deep learning. Nature biomedical engineering. 2018;2(3):158-64.http://dx.doi.org/10.1038/s41551-018-0195-0.\\u003c/li\\u003e\\n\\u003cli\\u003eTamai K, Terai H, Hoshino M, et al. Deep Learning Algorithm for Identifying Cervical Cord Compression Due to Degenerative Canal Stenosis on Radiography. Spine. 2023;48(8):519-25.http://dx.doi.org/10.1097/brs.0000000000004595.\\u003c/li\\u003e\\n\\u003cli\\u003eGurovich Y, Hanani Y, Bar O, et al. Identifying facial phenotypes of genetic disorders using deep learning. Nature medicine. 2019;25(1):60-4.http://dx.doi.org/10.1038/s41591-018-0279-0.\\u003c/li\\u003e\\n\\u003cli\\u003eHo CS, Chen YP, Fan TY, et al. Application of deep learning neural network in predicting bone mineral density from plain X-ray radiography. Archives of osteoporosis. 2021;16(1):153.http://dx.doi.org/10.1007/s11657-021-00985-8.\\u003c/li\\u003e\\n\\u003cli\\u003eIguchi T, Kanemura A, Kasahara K, et al. Lumbar instability and clinical symptoms: which is the more critical factor for symptoms: sagittal translation or segment angulation? Journal of spinal disorders \\u0026amp; techniques. 2004;17(4):284-90.http://dx.doi.org/10.1097/01.bsd.0000102473.95064.9d.\\u003c/li\\u003e\\n\\u003cli\\u003eHuang G, Liu Z, Maaten LVD, Weinberger KQ, editors. Densely Connected Convolutional Networks. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR); 2017 21-26 July 2017.\\u003c/li\\u003e\\n\\u003cli\\u003eEven JL, Chen AF, Lee JY. Imaging characteristics of \\u0026quot;dynamic\\u0026quot; versus \\u0026quot;static\\u0026quot; spondylolisthesis: analysis using magnetic resonance imaging and flexion/extension films. The spine journal : official journal of the North American Spine Society. 2014;14(9):1965-9.http://dx.doi.org/10.1016/j.spinee.2013.11.057.\\u003c/li\\u003e\\n\\u003cli\\u003eSeyedhoseinpoor T, Dadgoo M, Taghipour M, et al. Combining clinical exams can better predict lumbar spine radiographic instability. Musculoskeletal science \\u0026amp; practice. 2022;58:102504.http://dx.doi.org/10.1016/j.msksp.2022.102504.\\u003c/li\\u003e\\n\\u003cli\\u003eHe Y, Lin J, Zhu S, Zhu J, Xu Z. Deep learning in the radiologic diagnosis of osteoporosis: a literature review. The Journal of international medical research. 2024;52(4):3000605241244754.http://dx.doi.org/10.1177/03000605241244754.\\u003c/li\\u003e\\n\\u003cli\\u003eGoral S, K\\u0026ouml;se U. Development of A CapsNet and Fuzzy Logic Decision Support System for Diagnosing the Scoliosis and Planning Treatments via Schroth Method. IEEE Access. 2022;10:129055-78.http://dx.doi.org/10.1109/ACCESS.2022.3227763.\\u003c/li\\u003e\\n\\u003cli\\u003eCheng L-W, Chou H-H, Cai Y-X, et al. Automated detection of vertebral fractures from X-ray images: A novel machine learning model and survey of the field. Neurocomputing. 2024;566:126946.http://dx.doi.org/https://doi.org/10.1016/j.neucom.2023.126946.\\u003c/li\\u003e\\n\\u003cli\\u003eKirkaldy-Willis WH, Farfan HF. Instability of the lumbar spine. Clinical orthopaedics and related research. 1982;(165):110-23.http://dx.doi.org/.\\u003c/li\\u003e\\n\\u003cli\\u003eFujiwara A, Lim TH, An HS, et al. The effect of disc degeneration and facet joint osteoarthritis on the segmental flexibility of the lumbar spine. Spine. 2000;25(23):3036-44.http://dx.doi.org/10.1097/00007632-200012010-00011.\\u003c/li\\u003e\\n\\u003cli\\u003ePitk\\u0026auml;nen MT, Manninen HI, Lindgren KA, Sihvonen TA, Airaksinen O, Soimakallio S. Segmental lumbar spine instability at flexion-extension radiography can be predicted by conventional radiography. Clinical radiology. 2002;57(7):632-9.http://dx.doi.org/10.1053/crad.2001.0899.\\u003c/li\\u003e\\n\\u003cli\\u003ePate D, Goobar J, Resnick D, Haghighi P, Sartoris DJ, Pathria MN. Traction osteophytes of the lumbar spine: radiographic-pathologic correlation. Radiology. 1988;166(3):843-6.http://dx.doi.org/10.1148/radiology.166.3.3340781.\\u003c/li\\u003e\\n\\u003cli\\u003eElder BD, Vigneswaran K, Athanasiou KA, Kim DH. Biomechanical, biochemical, and histological characterization of canine lumbar facet joint cartilage. Journal of neurosurgery Spine. 2009;10(6):623-8.http://dx.doi.org/10.3171/2009.2.Spine08818.\\u003c/li\\u003e\\n\\u003cli\\u003eWang Y, Huang K. Research progress of diagnosing methodology for lumbar segmental instability: A narrative review. Medicine. 2022;101(1):e28534.http://dx.doi.org/10.1097/md.0000000000028534.\\u003c/li\\u003e\\n\\u003cli\\u003ePuntumetakul R, Saiklang P, Tapanya W, et al. The Effects of Core Stabilization Exercise with the Abdominal Drawing-in Maneuver Technique versus General Strengthening Exercise on Lumbar Segmental Motion in Patients with Clinical Lumbar Instability: A Randomized Controlled Trial with 12-Month Follow-Up. International journal of environmental research and public health. 2021;18(15)10.3390/ijerph18157811.\\u003c/li\\u003e\\n\\u003cli\\u003eLarivi\\u0026egrave;re C, Rabhi K, Preuss R, Coutu MF, Roy N, Henry SM. Derivation of clinical prediction rules for identifying patients with non-acute low back pain who respond best to a lumbar stabilization exercise program at post-treatment and six-month follow-up. PloS one. 2022;17(4):e0265970.http://dx.doi.org/10.1371/journal.pone.0265970.\\u003c/li\\u003e\\n\\u003cli\\u003eLarivi\\u0026egrave;re C, Preuss R, Coutu MF, Sullivan MJ, Roy N, Henry SM. Disability reduction following a lumbar stabilization exercise program for low back pain: large vs. small improvement subgroup analyses of physical and psychological variables. BMC musculoskeletal disorders. 2024;25(1):358.http://dx.doi.org/10.1186/s12891-024-07480-4.\\u003c/li\\u003e\\n\\u003cli\\u003eKumar SP. Efficacy of segmental stabilization exercise for lumbar segmental instability in patients with mechanical low back pain: A randomized placebo controlled crossover study. North American journal of medical sciences. 2011;3(10):456-61.http://dx.doi.org/10.4297/najms.2011.3456.\\u003c/li\\u003e\\n\\u003cli\\u003eKanayama M, Hashimoto T, Shigenobu K, Togawa D, Oha F. A minimum 10-year follow-up of posterior dynamic stabilization using Graf artificial ligament. Spine. 2007;32(18):1992-6; discussion 7.http://dx.doi.org/10.1097/BRS.0b013e318133faae.\\u003c/li\\u003e\\n\\u003c/ol\\u003e\"}],\"fulltextSource\":\"\",\"fullText\":\"\",\"funders\":[],\"hasAdminPriorityOnWorkflow\":false,\"hasManuscriptDocX\":true,\"hasOptedInToPreprint\":true,\"hasPassedJournalQc\":\"\",\"hasAnyPriority\":false,\"hideJournal\":false,\"highlight\":\"\",\"institution\":\"\",\"isAcceptedByJournal\":false,\"isAuthorSuppliedPdf\":false,\"isDeskRejected\":\"\",\"isHiddenFromSearch\":false,\"isInQc\":false,\"isInWorkflow\":false,\"isPdf\":false,\"isPdfUpToDate\":true,\"isWithdrawnOrRetracted\":false,\"journal\":{\"display\":true,\"email\":\"info@researchsquare.com\",\"identity\":\"european-spine-journal\",\"isNatureJournal\":false,\"hasQc\":true,\"allowDirectSubmit\":false,\"externalIdentity\":\"esjo\",\"sideBox\":\"Learn more about [European Spine Journal](http://link.springer.com/journal/586)\",\"snPcode\":\"586\",\"submissionUrl\":\"https://submission.springernature.com/new-submission/586/3\",\"title\":\"European Spine Journal\",\"twitterHandle\":\"\",\"acdcEnabled\":true,\"dfaEnabled\":true,\"editorialSystem\":\"stoa\",\"reportingPortfolio\":\"Springer Hybrid\",\"inReviewEnabled\":true,\"inReviewRevisionsEnabled\":false},\"keywords\":\"Low Back Pain, Lumbar Vertebrae, Radiography, Deep Learning, Predictive Learning Models, Data Visualization\",\"lastPublishedDoi\":\"10.21203/rs.3.rs-7410485/v1\",\"lastPublishedDoiUrl\":\"https://doi.org/10.21203/rs.3.rs-7410485/v1\",\"license\":{\"name\":\"CC BY 4.0\",\"url\":\"https://creativecommons.org/licenses/by/4.0/\"},\"manuscriptAbstract\":\"\\u003ch2\\u003eObjective\\u003c/h2\\u003e\\u003cp\\u003eTo develop a deep learning model for predicting lumbar segmental instability (LSI) using neutral lateral lumbar radiographs and to identify key radiographic features associated with LSI.\\u003c/p\\u003e\\u003ch2\\u003eMethods\\u003c/h2\\u003e\\u003cp\\u003eA DenseNet121-based stacking ensemble model was integrated with Support Vector Machine, Random Forest, and Softmax classifiers. Model validation employed 10-fold cross-validation, with performance assessed using AUC, accuracy, sensitivity, specificity, and F1-score. Sensitivity analyses evaluated robustness across spinal/non-spinal regions, age/gender subgroups, and feature interactions. Gradient-weighted Class Activation Mapping (Grad-CAM) was used to localize critical anatomical regions, which were further validated in machine learning frameworks.\\u003c/p\\u003e\\u003ch2\\u003eResults\\u003c/h2\\u003e\\u003cp\\u003eThe DenseNet121-stacking model achieved an AUC of 0.82, accuracy of 76%, sensitivity of 61%, and specificity of 84%. Calibration curves confirmed strong alignment with clinical outcomes. Grad-CAM identified facet joints (34.1%), intervertebral discs (27.0%), and osteophytes (25.4%) as the predominant contributors. The integration of these features into machine learning models yielded an AUC of 0.749. Subgroup analyses demonstrated consistent performance across age and gender groups. Decision curve analysis confirmed the clinical utility of the model in all cohorts.\\u003c/p\\u003e\\u003ch2\\u003eConclusion\\u003c/h2\\u003e\\u003cp\\u003eThe stacking ensemble model developed in this study effectively predicts LSI based on neutral lateral lumbar radiographs and identifies key imaging biomarkers, including facet joint hypertrophy, disc degeneration, and osteophyte formation. The model demonstrated stable performance across different age and gender groups, indicating strong generalizability and providing a reliable tool for precise clinical screening and individualized decision-making.\\u003c/p\\u003e\",\"manuscriptTitle\":\"Deep Learning for Predicting Lumbar Instability Using Neutral Lateral Lumbar Radiographs: A Retrospective Study\",\"msid\":\"\",\"msnumber\":\"\",\"nonDraftVersions\":[{\"code\":1,\"date\":\"2025-11-13 14:45:28\",\"doi\":\"10.21203/rs.3.rs-7410485/v1\",\"editorialEvents\":[{\"type\":\"communityComments\",\"content\":0},{\"type\":\"decision\",\"content\":\"Revision requested\",\"date\":\"2025-11-25T04:37:14+00:00\",\"index\":\"\",\"fulltext\":\"\"},{\"type\":\"editorInvitedReview\",\"content\":\"\",\"date\":\"2025-11-14T22:27:17+00:00\",\"index\":\"hide\",\"fulltext\":\"\"},{\"type\":\"reviewerAgreed\",\"content\":\"32016797146500198841387361673430269242\",\"date\":\"2025-11-04T14:38:11+00:00\",\"index\":\"hide\",\"fulltext\":\"\"},{\"type\":\"editorInvitedReview\",\"content\":\"\",\"date\":\"2025-11-04T07:50:14+00:00\",\"index\":\"hide\",\"fulltext\":\"\"},{\"type\":\"reviewerAgreed\",\"content\":\"62932635934838111544513477576895896485\",\"date\":\"2025-11-03T08:19:04+00:00\",\"index\":\"hide\",\"fulltext\":\"\"},{\"type\":\"reviewersInvited\",\"content\":\"\",\"date\":\"2025-11-03T07:45:17+00:00\",\"index\":\"\",\"fulltext\":\"\"},{\"type\":\"editorAssigned\",\"content\":\"\",\"date\":\"2025-08-29T04:32:17+00:00\",\"index\":\"\",\"fulltext\":\"\"},{\"type\":\"checksComplete\",\"content\":\"\",\"date\":\"2025-08-29T04:31:05+00:00\",\"index\":\"\",\"fulltext\":\"\"},{\"type\":\"submitted\",\"content\":\"European Spine Journal\",\"date\":\"2025-08-19T16:11:19+00:00\",\"index\":\"\",\"fulltext\":\"\"}],\"status\":\"published\",\"journal\":{\"display\":true,\"email\":\"info@researchsquare.com\",\"identity\":\"european-spine-journal\",\"isNatureJournal\":false,\"hasQc\":true,\"allowDirectSubmit\":false,\"externalIdentity\":\"esjo\",\"sideBox\":\"Learn more about [European Spine Journal](http://link.springer.com/journal/586)\",\"snPcode\":\"586\",\"submissionUrl\":\"https://submission.springernature.com/new-submission/586/3\",\"title\":\"European Spine Journal\",\"twitterHandle\":\"\",\"acdcEnabled\":true,\"dfaEnabled\":true,\"editorialSystem\":\"stoa\",\"reportingPortfolio\":\"Springer Hybrid\",\"inReviewEnabled\":true,\"inReviewRevisionsEnabled\":false}}],\"origin\":\"\",\"ownerIdentity\":\"f01157a5-b680-4cb4-a2f4-338cb5680bd4\",\"owner\":[],\"postedDate\":\"November 13th, 2025\",\"published\":true,\"recentEditorialEvents\":[],\"rejectedJournal\":[],\"revision\":\"\",\"amendment\":\"\",\"status\":\"under-review\",\"subjectAreas\":[],\"tags\":[],\"updatedAt\":\"2026-01-21T10:57:40+00:00\",\"versionOfRecord\":[],\"versionCreatedAt\":\"2025-11-13 14:45:28\",\"video\":\"\",\"vorDoi\":\"\",\"vorDoiUrl\":\"\",\"workflowStages\":[]},\"version\":\"v1\",\"identity\":\"rs-7410485\",\"journalConfig\":\"researchsquare\"},\"__N_SSP\":true},\"page\":\"/article/[identity]/[[...version]]\",\"query\":{\"redirect\":\"/article/rs-7410485\",\"identity\":\"rs-7410485\",\"version\":[\"v1\"]},\"buildId\":\"XKTyCvWXoU3ODBz1xrDgd\",\"isFallback\":false,\"isExperimentalCompile\":false,\"dynamicIds\":[84888],\"gssp\":true,\"scriptLoader\":[]}","source_license":"CC-BY-4.0","license_restricted":false}