Sage Journals: Discover world-class research

Abstract

Study Design

Narrative Review.

Objective

Machine learning (ML) is one of the latest advancements in artificial intelligence used in medicine and surgery with the potential to significantly impact the way physicians diagnose, prognose, and treat spine tumors. In the realm of spine oncology, ML is utilized to analyze and interpret medical imaging and classify tumors with incredible accuracy. The authors present a narrative review that specifically addresses the use of machine learning in spine oncology.

Methods

This study was conducted in accordance with the Preferred Reporting Items of Systematic Reviews and Meta-Analysis (PRISMA) methodology. A systematic review of the literature in the PubMed, EMBASE, Web of Science, Scopus, and Cochrane Library databases since inception was performed to present all clinical studies with the search terms ‘[[Machine Learning] OR [Artificial Intelligence]] AND [[Spine Oncology] OR [Spine Cancer]]’. Data included studies that were extracted and included algorithms, training and test size, outcomes reported. Studies were separated based on the type of tumor investigated using the machine learning algorithms into primary, metastatic, both, and intradural. A minimum of 2 independent reviewers conducted the study appraisal, data abstraction, and quality assessments of the studies.

Results

Forty-five studies met inclusion criteria out of 480 references screened from the initial search results. Studies were grouped by metastatic, primary, and intradural tumors. The majority of ML studies relevant to spine oncology focused on utilizing a mixture of clinical and imaging features to risk stratify mortality and frailty. Overall, these studies showed that ML is a helpful tool in tumor detection, differentiation, segmentation, predicting survival, predicting readmission rates of patients with either primary, metastatic, or intradural spine tumors.

Conclusion

Specialized neural networks and deep learning algorithms have shown to be highly effective at predicting malignant probability and aid in diagnosis. ML algorithms can predict the risk of tumor recurrence or progression based on imaging and clinical features. Additionally, ML can optimize treatment planning, such as predicting radiotherapy dose distribution to the tumor and surrounding normal tissue or in surgical resection planning. It has the potential to significantly enhance the accuracy and efficiency of health care delivery, leading to improved patient outcomes.

Keywords

machine learning spine oncology artificial intelligence AI in oncology machine learning in spine

Introduction

Spine tumors, both primary and metastatic, have been steadily increasing in incidence in recent years and are a major burden of oncologic care.^1,2 Primary tumors of the spine are rare, but metastases to the vertebral column are much more prevalent, and present in up to 70% of cancer patients.³ Spine metastases can lead to epidural spinal cord compression causing substantial axial pain, vertebral body fractures, as well as myelopathy or radiculopathy. Spine tumors can be classified using their origin and location within the spine. Metastatic tumors originate outside the spine and are often indicative of disease progression. Primary bone tumors of the spine as well as primary tumors of the spinal cord or exiting nerves originate in the spine. These soft tissue tumors are commonly referred to as intradural or extradural based on their location relative to the dura mater. Being able to distinguish tumor type is critical, as they require different approaches to management and treatment.⁴

The location of primary, metastatic, or intradural lesions can also impact the burden of treatment due to location-related intrinsic surgical morbidities and their obvious influence on outcomes. Additionally, location can impact radiation dose delivery given each pathology has a specific response to radiation.⁵ As the field of spine oncology continues to expand, there is a need for new tools that will aid physicians in creating individualized plans for oncologic care.

Background on Machine Learning

Machine learning (ML) involves the development of algorithms that learn from data and make predictions without explicit programming. ML models have been around for almost 5 decades and can be utilized in health care given the large swaths of patient data.⁶ In recent years, ML has been applied to various fields of medicine, including neurosurgery, with the goal of improving diagnosis, treatment, and patient outcomes. Artificial neural networks (ANN) are the primary ML algorithms utilized within clinical outcomes research.⁷ The common types of neural networks include ANNs, Convolutional Neural Networks (CNN), and Recurrent Neural Networks (RNN) (Figure 1, Table 1). ANN is primarily used for tabular data, CNN for visual data such as radiographs, and RNN for sequence data.⁸

Figure 1.

Convolutional Neural Network (CNN) is a type of deep learning algorithm that is designed to analyze image data. It uses a series of filters to identify patterns and features in the image, allowing it to recognize objects and classify images. Artificial Neural Network (ANN) is a type of machine learning algorithm that is modeled after the structure and function of the human brain. It consists of interconnected nodes, or artificial neurons, that process information and make predictions based on the data input. Recurrent Neural Network (RNN) is a type of neural network that is designed to process sequential data, such as time series data or text. Unlike traditional neural networks, RNNs have memory and can use past information to make predictions about future events. They are often used in applications such as speech recognition and natural language processing.

Table 1.

Definition of Terms, and ML Models.

ML Model Examples/Key Terms	Definition
Random forest	Using multiple decision trees, combines these results to reach a single consensus. Can be used for both classification and regression models
Support vector machine	SVM finds a hyperplane in an N-dimensional space (n-number of features) that creates the largest margin between 2 data sets. Can be used for both classification and regression models
Decision tree	Starting with a root classifier, further categorizations are based on the answers to the previous set of questions
Bayes Classifier	A type of classification model based around bayes' Theorem involving probability of hypothesis with prior knowledge
LASSO	A regression analysis method that performs variable selection and regularization to enhance prediction accuracy. Can be applied to other forms of regression models
Logistic	Dependent variables or targets are put in a dichotomous system, and the probability of each binary event happening can be predicted
Multivariate logistic	Extension of multiple regression where multiple independent variables and one dependent variable where the number of independent variables is used to try and predict the output
Neural network	Interconnected nodes are created to resemble the neurons of a human brain, and helps models learn from mistakes to improve prediction models

AUC	AUC is a value between 0 and 1 that is interpreted as the probability that a model will assign a larger probability to a random positive example than a random negative example. An AUC value of 1 means that a model can correctly distinguish between positive
C-index	Metric that evaluates the predictions made and how closely they resemble real events
Brier score	Calculates the mean squared error between predicted and expected values, and the score represents the magnitude of error in the probability forecast. The better a model performs, the closer to 0 it scores
Dice similarity coefficient	Measures how similar to independent sets of data are on a scale of 0, no overlap, to 1, complete overlap
Training	Teaching the model what to look for, and what the input features mean in reference to output data
Validation	Once the model has been trained, it is evaluated using a testing data set to determine how well the model works

The use of ML in health care currently provides a supplementary role to physicians in developing prediction models, identifying trends and helping reduce physician workload.⁹ In the field of oncology, ML is being tested to extract magnetic resonance imaging (MRI), or computed tomography (CT) data for diagnosis,¹⁰ and predict survival rates based on individual characteristics.¹¹

A systematic analysis on the global burden of cancer reported 26.3% increase in global cancer cases, 20.9% increase in cancer related deaths, and 16.0% increase in disability adjusted life years from 2010 to 2019.¹² As oncologic therapy continues to improve, we expect to see an increasing incidence of patients with spinal metastases. Prediction models and trend data can assist physicians and patients to create personalized oncologic care plans based on patient and tumor characteristics. Given the increasing number of patients entering oncology clinics, ML has been put on the forefront of innovation in trying to help combat the increasing burden of oncologic care for both patients and physicians.¹³

To date, several systematic reviews have examined the use of machine learning within the field of neurosurgery and its utility in neurosurgical applications; however, there are no current studies that have reviewed the use of machine learning within spine oncology. Therefore, in this study, we present a thorough and comprehensive systematic review of the literature regarding the use of ML in spine oncology.

Methods

Search Strategy

A targeted systematic review was performed and reported in accordance with the PRIMSA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines. A literature search was performed using Embase (Elsevier), PubMed (National Library of Medicine), Web of Science, Scopus, and Cochrane Library databases from inception to April 14, 2023, to identify studies evaluating the use of machine learning in spinal oncology. Search terms utilized to identify relevant studies were ‘[[Machine Learning] OR [Artificial Intelligence]] AND [[Spine Oncology] OR [Spine Cancer]]’.

Inclusion and Exclusion Criteria

Studies were included if they met the following criteria: (1) reported primary data with some form of machine learning used with patients with a spinal lesion; (2) reported on where the machine learning algorithm was used, and how it can be applied. Exclusion criteria were non-English studies, abstracts, oral presentations, cadaveric studies, phantom studies, case reports, review articles, systematic reviews, and studies whose only algorithm used was radiomics as radiomics does not change the surgical strategy in metastatic epidural spinal cord compression. Abstracts were screened independently by 2 authors [SBW, VM], and conflicts were resolved through discussion and decision by a third author [JW]. Potential studies were extracted for full-text evaluation, which was completed by a two-vote system by 4 of the authors [SBW, JW, VM, CSAL], with conflicts resolved by an independent third author.

Data Extraction

Articles that passed our inclusion criteria were divided into subgroups based on tumor types, primary bone (extradural), metastatic, intradural, or a combination of the types. The primary bone tumor subgroup was classified as primary spine tumors such as Ewing sarcoma, chondrosarcoma, osteosarcoma, and sacral tumors. The intradural tumor subgroup consisted of schwannoma, astrocytoma, ependymoma, and meningioma. Articles that reported on multiple tumor types were split into a separate, mixed, subgroup. Data collected included study characteristics (publication year, journal published, sample size for testing and validation groups, type of machine learning algorithm used, inclusion and exclusion criteria), patient demographics (sex, type of cancer, area of spinal lesions), procedure characteristics (machine models input data and output groups, machine algorithms validation method, machine models prediction performance). A full-text evaluation was conducted independently by 3 authors [SBW, JW, CSAL].

Results

A review of the literature revealed a surge in publications on Machine Learning in Spine Oncology in the last 30 years (Figure 2). A total of 480 unique publications matched the search criteria. Four hundred and two publications were excluded during title and abstract screening due to failure to address the use of AI/ML in spine oncology. A full-text review was conducted for 77 studies. Twenty-eight of those studies were excluded due to emphasis on radiomics or were not published as a full text article. Forty-six studies met the final inclusion criteria for this systematic review (Figure 3).

Figure 2.

Bar graph of database search results of studies involving machine learning in spine surgery and spinal oncology since database inception.

Figure 3.

PRISMA flow diagram of the systematic literature search for this systematic review. Data added to the PRISMA template [from Moher D, Liberati A, Tetzlaff J, Altman DG, The PRISMA Group (2009). Reporting Items for Systematic reviews and Meta-Analyses: The PRISMA Statement. PLoS Med. 6 (7):e1000097] under the terms of the Creative Commons Attribution License.

We separated the forty-six included articles based on their investigation of primary spine tumors metastatic disease, intradural tumors, or combinations of these types. Once stratified, twenty-nine publications addressed metastatic spinal tumors, 4 publications addressed primary spinal tumors, 5 publications addressed intradural tumors, and 8 publications addressed a mixture of any of the 3 tumor types.

In the reviewed articles, most algorithms used supervised learning either for classification or regression purposes while a few used unsupervised learning models that focus on clustering outcomes. Classification models include Random Forest (RF), Support Vector Machine (SVM), Decision Tree analysis, Bayes Classifier. These models were most often used in spine oncology to classify variables such as tumor detection or differentiation,^14-27 overall mortality,^28-35 or adverse events related to surgery.³⁶ Additionally, we found that regression algorithms including LASSO, logistic and multivariate logistic were used to predict continuous variables such as overall survival and prognosis.^15,33,37-45 Some studies used pretrained models such as ResNet50, Unet++, and ImageNet as a foundation, which were then further trained with their data through 50-100 epochs (number of complete passes through the training dataset). Large scale comparative statistics were absent due to high variability of study aims, algorithms, and verification metrics.

Metastatic Spine Disease

Twenty-nine (63.0%) studies reported on the use of ML techniques to predict survival, tumor detection, and tumor differentiation in metastatic spine tumors. Data extracted from these articles are found in Table 2. Twelve (41.3%) of these studies focused on predicting mortality or adverse events in patients with metastatic tumors. Ten (34.4%) studies aimed to develop algorithms to detect tumors on MRI or CT images, or differentiate spinal metastases based on primary location. Seven (24.1%) studies addressed prognosis. Input features into these models ranged from MRI data, both TIWI and T2WI, as well as Computed Tomography data, and Clinical Features. Each study used different clinical features including demographic data, lab values, metastatic spinal tumor frailty index (MSTFI), and the psoas muscle area. Many did not report pathologic analysis but those that did described patients with Lung, Thyroid, Liver, Breast, Prostate, Esophagus, Urinary Tract, Kidney, Non-small cell lung cancer (NSCLC), Lymphoma and Multiple Myeloma (MM).^{14,24,36,43,46-49} Key definitions for reported outcome parameters, including AUC, dice similarity coefficient, and Brier score, are described in Table 1.

Table 2.

Summary of Machine Learning Studies Focused on Metastatic Disease in Spine Oncology.

Authors	Inclusion Criteria	Exclusion Criteria	Input Features	Output	Machine Learning Model	Size of Training Set	Validation Method (NUMBER)	Size Test Set	Prediction Performance (CA/AUC)	Other Performance Metric/Brier Score
Fourman et al, 2023	Patients over 18 years treated surgically for breast cancer spinal metastases from 2008 to 2020 at 2 large tertiary referral cancer centers in the same city	Patients with a diagnosis of breast cancer who underwent spine surgery for trauma, infection, or who survived for >3 months but did not follow-up were excluded	Clinical features	Risk of 90-day adverse events	Unsupervised cluster analysis	-	-	-	-	-
Karhade et al, 2022	(1) Patients 18 years or older; (2) diagnosis of spinal metastatic disease; and (3) undergoing index surgical treatment or radiation (conventional or stereotactic radiosurgery) between 2000 and 2019	Patients were excluded if they underwent intervention for any other condition (such as primary spinal tumors)	Clinical features	6 Week mortality	Elastic-net penalized logistic regression	3001	External data	1303	.81	-
Hallinan et al, 2022	Patients diagnosed with metastatic spinal cord compression between 09/2007 - 09/2017	MRI with prior instrumentation, poor image quality, and non-thoracic regions	MRI (T2W)	Epidural spinal cord compression	ResNet50	177	Fivefold external cross validation	38	.94	-
Zhao et al, 2022	The dataset contains arbitrary MRI images of thoracic, lumbar, and sacrum vertebrae of 6 different field of view from a single institution	-	MRI	Vertebrae	Enhanced-supervision reasoning network self-adaptive reasoning network	3680	Fivefold cross validation	920	.947	-
Chen et al, 2022	(1) Patients diagnosed with spinal MM and metastasis confirmed by histological or cellular pathology; (2) patients examined with preoperative MRI in our hospital to ensure qualified MRI images; and (3) patients with complete clinical information	-	T1W1	Differentiated from MM from lung cancer	ResNet50	-	-	-	.785	-
Elsamadicy et al, 2022	2016 – 2018 of the NRD for all patients with metastatic spinal column tumors (≥18 years old) undergoing surgical intervention	Patients who died and/or had no length of stay for the initial index event were excluded. Patients with insufficient time for 30- or 31-90-day accrual of readmissions were excluded	Clinical features	Cause of 30-day readmission	RF	3477	k-fold cross validation with hyperparameter tuning	869	0.6	-
Netherton et al, 2022	CT scans of patients that underwent radiotherapy	-	CT	Identify vertebral level	Unet++/RF	176	Random forest	22	.82	-
Yen et al, 2022	Patients aged ≥20 years who underwent radiotherapy for image-confirmed spinal metastases at an academic national medical center between 2010 and 2018 were included for analysis	Patients were excluded if they had a primary spinal tumor or received first radiotherapy for spinal metastases at other hospitals, or if the metastatic origin was unidentified due to multiple primary cancers or lack of tissue proof	MRI, lab values, and clinical features	Prognosis	SORG-MLA (stochastic gradient boosting algorithm)	-			90 day .78, 1 year .76	Brier score: 90 day - .16, 1 year .78
Arends et al, 2022	All patients with bone metastases referred to the radiation oncology department in the University medical center Utrecht, a cohort of 60 was randomly chosen with these features. Presence of bone metastases in the trunk (vertebrae, ribs, sternum and/or pelvis); CT slice thickness of 1 mm; visibility of at least 5 thoracic and/or lumbar vertebrae; absence of artefacts	-	CT	Spinal segmentation	Convolutional neural networks	45	Four-fold cross validation	15		Dice similarity coefficient: 96.7% and 94.5%
Hoshiai et al, 2022	CT images covering at least the chest and upper abdomen, past CT images in the same scan range, and CT images with and without contrast enhancement were included	Cases with vertebral metastases on previous CT images were excluded. Cases with >11 metastases were also excluded because it was difficult to evaluate individual metastases, since they might be fused or diffuse	CT	Detecting vertebral metastases	Pooling free residual network	50	-	30	-	Figure of merit: .876
Massaad et al, 2022	Consecutive patients (18 years of age and older) who were surgically treated for spinal metastases between 2010 and 2019 at the Massachusetts General hospital	Patients treated non-surgically	CT	90 day and 1 year mortality risk, major complications	two convolutional neural networks (DenseNet, U-net)
Hu et al, 2022	Adult (≥18 years old) patients undergoing surgical treatment for spinal metastasis between 2012 and 2017 at a tertiary medical center in Taipei, Taiwan (1) age ≥18 years old (2) histopathologically confirmed spinal metastatic disease (3) had at least one computed tomography (CT) scan that could be used for morphometric analysis of the psoas muscle within 3 months of the index surgery (4) no missing data in the electronic medical records for calculation of the 6 PSSs	178 patients were excluded due to a lack of any abdominal CT scan within 3 months of the index surgery, and ight patients were excluded because their images were too blurry to be morphometrically analyzed due to artifacts from nearby spinal implants	CT, clinical features (psoas muscle area)	90-day, 1-year, and overall mortality rates, c-indexes, net benefit on DCA	Multivariate logistic regression analysis, Cox proportional-hazard regression analysis	-	-	-	PMA barely improved the discriminatory ability (c-index, .74; 95% confidence interval [CI], .67-.82 vs c-index, .74; 95% CI, .66-.81)	Brier score: .16-.23
Bakhsheshian et al, 2022	Patients receiving fusion for a malignant neoplasm of the spine => 3 different analyses each with their own inclusion criteria	-	Clinical features	Mortality, complications, length of stay (LOS), nonroutine discharges and costs	3 multivariable regression models (frailty, Elixhauser Comorbidity index (ECI), frailty & ECI)	-	NA (but validated binary frailty index was used to identify frail and non-frail groups)	-	Models using ECI alone (AUC = .636-.788) Models using frailty alone (AUC = .633-.752) frailty combined with ECI improved the prediction of increased LOS (AUC = .811)	-
Yoda et al, 2022	Patients who underwen thoracolumbar MRI scans within 60 days of the suspected time of fracture between 2008 and 2018		MRI	Fracture differentiation	ImageNet	100 epochs			STIR: .967, TIWI .984
Gui et al, 2021	De novo spinal metastases treated with SBRT at a single institution between 2009 and 2018 were identified retrospectively. Image-guided SBRT was delivered with a robotic or linear accelerator–based system in 5 or fewer fractions. The gross tumor volume was identified using CT, T1-weighted MRI, and T2-weighted MRI. The clinical target volume (CTV) was generally defined according to International spine research Consortium (ISRC) consensus guidelines.25 The planning target volume (PTV) was defined as the CTV plus a 1- to 2-mm uniform margin of expansion, minus the spinal cord planning risk volume. This was defined as the true spinal cord plus a 2-mm uniform expansion or the entire thecal sac if the target volume was inferior to the conus medullaris	Vertebral levels previously treated with surgical decompression or stabilization, cement augmentation, or radiation therapy were excluded. C1 and sacral metastases were excluded because of concern for different imaging characteristics and mechanisms of VCF compared with the remainder of the vertebral column. Cases in which the CTV was not defined according to the ISRC consensus guidelines or in which the tumor was limited to the paraspinal soft tissue were excluded	CT, MRI, clinical features	Prediction of vertebral compression fracture	RF	-	Leave one out cross validation	-	.878	-
Massaad et al, 2021	Consecutive patients, older than 18 years of age, who were surgically treated for spinal metastases between 2010 and 2019 at the Massachusetts General hospital. Included patients underwent surgery for decompression and stabilization of the spine for a diagnosis of epidural spinal cord compression, vertebral pathologic fracture, spinal instability, neurological deficit, and/or intractable pain	Patients were excluded if they were treated non-surgically	MSFTI parameters (clinical)	Frailty prediction	Lasso regression, RF gradient boosted decision tree	480	10-fold cross validation	48	.62	-
DiSilvestro et al, 2021	Spine tumor surgery patients	Missing quantitative or qualitative data	Clinical features	Postoperative 30-day mortality	Naïve bayes machine learning networks	1466	-	628	.898
Fan et al, 2020	Complete electronic medical records or clear histopathological confirmation of spinal metastases, PET/CT image quality meeting diagnostic requirements, and basic clinical information available	Primary spinal tumor, treated patients, tumor margins too difficult to delineate, incomplete clinical information, image quality not meeting diagnostic requirements, irregular spine, and inability to complete the follow-up	PET/CT	Detection of spinal metastases	Logistic regression, support vector machine, decision tree models	65	-	24	.902	87.5, 83.34, 75% accuracy
Gao et al, 2020	Clinical diagnosis records of SPECT nuclear medicine during the process of disease diagnosis		SPECT	Detection of bone metastases	K0means, C0v, and region growing	-	-	-	-	Tanimoto similarity coefficient: .777
Bongers et al, 2020	Patients of at least 18 years of age or older at the time of surgery, (2) surgical procedure for the resection of metastatic spine lesion performed between 2014 and 2016, (3) pathologic confirmation of primary tumor pathology		Clinical features	90-day, and one year mortality	SORG	732	-	200	.81 90-day, .84 1-year mortality	Brier score: .17 90-day, .16 1-year mortality
Karhade et al, 2020		-	Clinical features	90-day, 1-year, and overall mortality rates, c-indexes, net benefit on DCA	SORG	732	-	176	90-day = .81, 1 year = .78	.246
Kowalchuk et al, 2020	All patients treated with SBRT for spinal metastases from August 2007 to November 2017 at one institution were tabulated and considered for retrospective review	Patients with benign tumors or primary spinal cord cancers or who had a lack of follow up were excluded	Clinical features	Pain free status and local failure	Decision tree analysis	100	10-fold cross validation	42	90%	-
He et al, 2020	Patients operated for spinal metastatic disease between 2010 and 2018	Patients operated on exclusively with percutaneous vertebroplasty or percutaneous kyphoplasty	Clinical features	Overall survival	Shiny model (Cox)	185	Time-dependent AUC plot and calibration curve	80	AUC values both in training and validation samples were continuously greater than .750	-
Karhade et al, 2019	CT examinations associated with evident vertebral metastasis were searched in radiological records from June 2011 to September 2016 at our institute. CT images covering at least the chest and upper abdomen, past CT images in the same scan range, and CT images with and without contrast enhancement were included	Cases with vertebral metastases on previous CT images were excluded. Cases with >11 metastases were also excluded because it was difficult to evaluate individual metastases, since they might be fused or diffuse	Clinical features	90 day and 1 year mortality risk	Penalized logistic regression, stochastic gradient boosting, NN, SVM	586	10-fold cross validation	146	SBG: .83	Brier score: .14
Karhade et al, 2019	(1) primary current Procedural Terminology code for excision, osteotomy, decompression, fusion, or fixation (2) at the cervical, thoracic, or lumbosacral levels, (3) International classification of diseases diagnosis of secondary malignant neoplasm of bone, meninges, or spinal cord or diagnosis of pathological fracture, (4) confirmed comorbidity of disseminated cancer, (5) surgical subspecialty neurosurgery or orthopedics, (6) general anesthesia, (7) inpatient operation, and (8) year of operation between 2009 and 2016 (9) American Society of Anesthesiologist classification indicating systemic disease (II-V)	-	Clinical features	30-day mortality	Neural network, support vector machine, bayes point machine, and decision tree	1432	Validation method	358		Brier score: .701\|C-statistic: .769
Chmelik et al, 2018	-	-	CT, clinical features	Segmentation and classification of lytic and sclerotic metastatic lesions	Convolutional neural network (CNN)	19	Random validation and leave-one-out cross-validation	12	.80 AUC for lytic tissue\|.78 AUC fos Sclerotic tissue	-
Wei et al, 2018	1 or 2 spinal sites of disease with spinal cord compression, and no more than 2 vertebral bodies involved in tumor at that site. All patients underwent some form of spinal surgery. For control group, the senior investigator also determined that the patient would be a candidate for a minimally invasive surgical option	Patients who were managed non-surgically	Clinical features	Survival index	Cox proportional hazards model, generalized linear model	-	-	-	Positive predictive value of 92%	-
Wang et al, 2017	NA	NA	Sagittal spine MRI	Tumor delineation	Siamese neural network	-	10-fold cross validation	NA		.207 false positives/case when using aggregation approach
Paulino Pereira et al, 2016	Patients who were ≥18 years of age and who had undergone a surgical procedure for spine metastases between January 2002 and January 2014 at two tertiary care centers were included	Patients treated by kyphoplasty or vertebroplasty only, radiosurgery, or revision procedures	Clinical features	Prognosis	Multivariate Cox proportional hazards model	519	Fivefold cross validation	130	.70 for 30 days, .69 for 90 days, and .73 for 365	-

Abbreviation:RF: Random Forest.

Several studies evaluated a specific cancer type as it metastasized to the spine. One study attempted to predict EGFR mutation presence based on spine metastasis in patients with NSCLC.⁴⁶ Another sought to identify post operative adverse events in patients with metastatic breast cancer.³⁶ A third study attempted to differentiate between lung cancer and MM on MRI.¹⁴

Of the studies that evaluated mortality or adverse events, Karhade et al³² reported an area under the curve (AUC) of .81 for predicting 6-week mortality using an Elastic-net penalized logistic regression. Fourman et al³⁶ used unsupervised cluster analysis to identify 5 groups that had distinct outcomes following open spine surgery for metastatic breast cancer. High functioning patients and those who were ER/PR positive had the best outcomes, while patients with structural instability had the worst outcomes. Bongers et al²⁹ used clinical features to develop an algorithm that could predict 90-day and 1 year mortality with AUC of .81 and .84 respectively, following surgical resection of metastatic tumor to the spine. Karhade et al³⁴ reported an AUC of .83 and a Brier score of .14 when predicting mortality risk at 90-days and one year. Karhade et al⁵⁰ then described an algorithm that could similarly predict 90-day and 1-year mortality rates with AUC of .81 and .78, respectively. Massaad et al³⁵ used a deep learning model to risk stratify patients into high and low risk based on imaging and clinical features focusing on age, sex, frailty, muscle radiodensity, and subcutaneous fat. They reported significant differences in survival between groups and suggested use of their model in preoperative planning. He et al³⁷ predicted overall survival with an AUC over .75 using clinical features such as primary tumor type, site of metastasis, Frankel grade, and number of surgical segments. Hu et al³¹ predicted 90-day, 1-year, and overall survival with a c-index of .74 using logistic regression analysis. Wei et al⁴¹ developed a survival index with a positive predictive value of 92% using Cox proportional hazards model. Bakhsheshian et al²⁸ were able to predict mortality, complications, and length of stay with an AUCs of .788, .723, and .811 respectively using multivariate regression models. Karhade et al predicted 30-day survival with a c-statistic of .769 and a brier score of .701.³⁴ DiSilvestro et al described an algorithm that predicted post operative mortality using clinical features with an AUC of .898.³⁰

The next most common focal area was tumor detection or differentiation. Zhao et al. Developed an algorithm that assigned a metastatic lesion to a vertebral level with an AUC of .947.¹⁶ Chen et al were able to differentiate metastatic disease of origin between MM and lung cancer with an AUC of .785.¹⁴ Fan et al used a support vector machine and Positron emission tomography/CT scans to detect spine metastases with an AUC of .902.¹⁵ Gao et al used single-photon emission computed tomography to detect bone metastases reporting a Tanimoto similarity coefficient of .777.²⁷ Netherton et al used CT data run through Unet++ and RF to determine the vertebral levels of metastatic lesions with an AUC of .82.²² Arrends et al reported a Dice Similarity Coefficient (DSC) of 97% and 95% for internal and external validation, respectively.¹⁷ Wang et al²⁴ demonstrated a 44.8% reduction in false positives for automated rates of detection of spinal metastases, from .375 to .207 using a Siamese Neural Network. Hoshai et al developed an algorithm that improved figure of merit (FOM) for physician and resident detection of vertebral metastases from .848 to .876 (P = .01) and .752 to .799 (P = .02), respectively.⁵¹ Chmelik et al¹⁸ classified metastatic lesions as lytic or sclerotic with AUCs of .80 and .78 respectively. Yoda et al²⁶ differentiated metastatic or osteoporotic origins of vertebral fractures using STIR and TIWI MRI with AUCs of .987 and .984 respectively.

Several studies evaluated prognosis, attempting to predict post-operative outcomes. Gui et al⁴³ used clinical features to predict vertebral compression fracture rates, AUC = .878, using an RF model. Elsamadicy et al⁴² predicted cause of 30-day readmission, AUC = .6, using clinical features, primarily frailty status, input into an RF model. Yen et al⁴⁵ evaluated overall prognosis using a stochastic gradient boosting algorithm and MRI data combined with laboratory values and other clinical features with an AUC of .78 and .76 for 90-day and 1-year prognosis, respectively. They also reported Brier scores of .16 and .78 for 90-day and 1-year prognosis, respectively. Kowalchuk et al evaluated pain free status and local failure using decision tree analysis with an AUC of .90.⁵² Massaad et al used LASSO regression and an RF gradient boosted decision tree to predict frailty with an AUC of .62.⁴⁴ Paulino Pereira et al⁴⁸ predicted post operative prognosis at 30-days, 90-days, and 1-year with AUCs of .70, .69, and .73 respectively using Multivariate COX proportional hazards model. (Table 2).

ML in metastatic spine disease had the most evidence of all types of spine cancer. ML tools displayed mixed results with overall utility in determining origin of tumor, survival, and surgical outcomes using both imaging and patient demographic factors.

Primary Bone Tumors

Four (8.7%) studies reported on the use of ML methods to predict survival and tumor differentiation. The details of these studies are included in Table 3. Three (75%) of these studies addressed either limited or overall survival, and the other one (25%) addressed tumor differentiation. Input for the ML models was clinical features, or CT images. Clinical features were varied for each study but some of the inputs included demographic data (sex, race, age, marriage status), tumor characteristics (location, type, number), and operative details. Tumors reported include Ewing sarcoma, chondrosarcoma, osteosarcoma, sacral chordoma, or sacral giant cell tumor. Two (50%) of the studies had tumors that encompassed all areas of the spine, one study (25%) focused on the pelvis, and vertebral column, while one (25%) focused only on the sacrum. From the 4 studies, there was a combination of RF, NN, Univariate and Cox Regression (UC), Best Subset Regression (BSR), LASSO, SVM, and Generalized Linear Model (GLM) used, with some of the studies testing to see which algorithm would perform the best for the desired measurement/outcome. The performance of the ML models was assessed using AUC, or other validation metrics such as Brier score, or dice score, which are defined in Table 1.

Table 3.

Summary of Machine Learning Studies Focused on Primary Site Malignancy in Spine Oncology.

Authors	Inclusion Criteria	Exclusion Criteria	Disease (Cancer)	Input Features	Output	Machine Learning Model	Size of Training Set	Validation Method	Size Test Set	Prediction Performance (CA/AUC)
Fan et al, 2022	Patients diagnosed with Ewing sarcoma between 1975 and 2016 with primary site of vertebral column or pelvis	Ewing sarcoma primary outside of pelvis or vertebral column	Ewing sarcoma	Clinical features	Survival	Ensemble model (Cox, random survival forest, CoxBoost, DeepCox)	666	5-fold cross validation	75	AUC of .705/0.747/0.851 for predicting 3-/5-/10-year survival, while .734/0.779/0.830 for predicting 3-/5-/10-year survival
Li et al, 2022	1) Patients with; (2) the post-2010 SEER database incorporated information on metastatic sites and therefore included patients diagnosed between 2010 and 2016; (3) chondrosarcoma was the first and only primary malignancy; and (4) complete clinical information, including patient’s age, gender, race, primary site, tumor size, tumor TNM stage and grading, metastatic site, surgery, and whether radiotherapy and chemotherapy were administered	MR images of intradural or intramedullary nervous system tumors and ones acquired from other hospitals	Osteosarcoma	Clinical features	Overall survival	Univariate Cox, BSR, LASSO	1209	Cross validation		.836
Ryu et al, 2020	Patients from SEER database with chondrosarcoma diagnosed from 1973 to 2014 were selected	Cases only with an autopsy/death certificate were excluded due to the unknown survival periods	Chondrosarcoma	Clinical features	Survival	Risk Estimate Distance survival neural network	870	5-fold cross validation	218	.84
Yin et al, 2019	had a sacrum tumour mentioned in preoperative CT and CTE reports; CT and CTE images were complete; pathology reports confirmed SC or SGCT	Images with poor quality including obvious artefacts; patients without enhanced images	Sacral chordomas, sacral giant tell tumors	CT	Differentiated SCs and SGCTs	SVM, generalized linear models, and RF	66		29	.984 AUC(LASSO + GLM)

Most of the articles that discussed primary spine lesions were investigating overall survival using clinical features, with each article looking at a different primary tumor type. Fan et al¹⁵ reported that an ensemble model combining CoxBoost and Cox with feature selection outperforming the traditional TNM system for survival rates of Ewing sarcomas with an AUC of .851 for 10-year survival. Li et al³⁸ developed an open access web platform using BSR, UC, and LASSO that predicts overall survival for chondrosarcoma patients, demonstrating high AUC’s > .8 for internal and external validation sets. Lastly, Ryu et al trained a neural network to predict survival of chondrosarcoma, and the average AUC for the 5 subgroups, created based on predicted percentage survival, using the test set was .84.⁴⁰

The last article in the primary section focused on tumor differentiation. Yin et al used multiple ML models, including LASSO, SVM, and GLM to determine which yielded the best results with 2 feature sets. They showed that LASSO + GLM had the highest accuracy of any combination of ML models, and that using CT-enhanced features yielded better differentiation of sacral chordomas and sacral giant cell tumors with an AUC of .984.⁵³ Overall, these studies show that various ML models can be used for different purposes in the field of spine oncology, with similar AUCs between models and desired output classes. (Table 3).

ML in primary bone (extradural) tumors of the spine demonstrates evidence in predicting tumor origin and overall survival, providing surgeons another tool in the management of oncological spine patients.

Intradural

Five (10.9%) studies reported on the use of ML methods to predict survival and tumor differentiation. The details of these studies are included in Table 4. Three (60%) of these studies addressed tumor detection, or segmentation, one (20%) of these studies addressed overall survival, and the other one (20%) addressed tumor differentiation. Input for the ML models was clinical features, or MRI (both T1W and T2W). Clinical features were varied for each study but some of the inputs included demographic data (sex, race, age, marriage status), tumor characteristics (location, type, number), and operative details. Tumors reported include schwannoma, astrocytoma, ependymoma, hemangioma, and meningiomas. All 4 of the studies had tumors that encompassed all areas of the spine. From the 4 studies, there was a combination of Random Forest (RF), You Only Look Once version 3, NN, and 3d U-net algorithms.

Table 4.

Summary of Machine Learning Studies Focused on Intradural Malignancy in Spine Oncology.

Authors	Inclusion Criteria	Exclusion Criteria	Input Features	Output	Machine Learning Model	Size of Training Set	Validation Method	Size Test Set	Prediction Performance (CA/AUC)	Other Validation Metrics
Ryu et al	Histopathologically confirmed spinal ependymoma between 1973 and 2014	Unknown survival periods, with unknown treatment modality, or who had undergone neoadjuvant RT	Clinical features	Overall survival	Random forest (RF)	2282	Stepwise logistic regression model	540	.74 (5-year OS),/0.76 (7-year OS),/0.81 (10-year OS)	-
Ito et al	Patients with intradural extramedullary spinaltumors that had been histologically diagnosed as schwannomas and who had undergone tumor resection at our hospital for the period between April 2015 and March 2019	-	MRI (T1WI, T2WI)	Detection of Schwannoma	You only look once v3	-	Five-fold cross validation	-	80.3%, 91.0%, 93.5% with T1W1, T2W1, and both, respectively	-
Lemay et al	Preoperative examination between October 2012 to September 2018, with heterogeneous vertebral coverage (cervical, thoracic, and lumbar), from patients diagnosed with spinal cord tumor	-	MRI (T1WI, T2WI)	Tumor segmentation	3d U-net	218	-	73	DICE score 76.7 % for tumor, edema, and cavity	-
Maki et al. 2020	Intradural extramedullary spinal tumors who had undergone tumor resection between February 2008 and December 2018	Other intradural extramedullary tumors such as myxopapillary ependymomas, neurofibromas, malignant nerve sheath tumors, and metastasis. Also, dumbbell-type tumors	MRI	Differentiated Schwannoma and meningioma	CNN	-	5-fold cross validation		.876 AUC (T2WI MRI)	-
Zhuo et al. 2022	January 2012 to December 2018 and identified 494 patients (119 patients with astrocytoma, 131 with ependymoma, 101 with MS, and 143 with NMOSD)	Insufficient image quality	MRI	Astrocytoma and ependymoma	MultiResUNet	490	-	157	Dice scores of .77 and .80 for model segmentation of astrocytoma and ependymoma respectively, against manual labeling

The largest subset of ML use with intradural spine tumors was investigating tumor detection, and segmentation. Of those that focused on detecting tumors, both investigators used magnetic resonance imaging (MRI) as the input for the ML models, and focused on less common tumors like schwannomas, astrocytoma, ependymomas, and hemangioblastomas. Ito et al¹⁹ used YOLO version 3 to detect Schwannomas on MRIs, and the results using T2W images compared favorably to those of 2 physicians with the accuracy being 91%, 90.2%, and 89.3%, respectively. For segmentation, Lenmay et al showed that the 3d U-Net model was able to better segment spine lesions including astrocytoma’s, ependymomas, and hemangioblastomas when including tumor, cavity, and edema as compared to tumor only with Dice scores of 76.7% and 61.8%, respectively.²⁰ Zhuo et al developed a deep learning framework for segmentation and classification of spinal cord lesions included astrocytoma and ependymoma with AUC .99.⁵⁴

The other articles focused on overall survival, or tumor differentiation. Ryu et al³⁹ found that a RF model investigating survival patterns using patient and tumor characteristics of ependymomas performed better than a stepwise logistic regression model with 5-year and 10-year survival, with the AUC being .74, and .81, respectively. The last article for the intradural section focused on tumor differentiation. Maki et al²¹ used a Convolutional NN (CNN) to differentiate schwannomas from hemangiomas, with the accuracy of the CNN being comparable to 2 trained radiologists both using T1W 81%, 82%, and 81%, respectively, and for T2W MRI being 80%, 69%, and 73%, respectively. Overall, these studies showed that ML is a helpful tool in tumor detection, differentiation, segmentation, and predicting survival of intradural spine tumors (Table 4).

With only 5 studies, ML for intradural tumors lacks robust volume of evidence however each manuscript demonstrated significant results where ML techniques assisted in tumor detection, segmentation, and survival.

Mixed Studies

Eight (17.8%) studies reported on the use of ML methods to predict tumor type, malignancy, post-operative/discharge features, tumor detection, predict 5-year survival, and fracture type. The details of these studies are included in Table 5. Two of the studies addressed malignancy status, 2 addressed post-operative/discharge features, one addressed fracture type, one addressed tumor differentiation, one addressed 5-year survival, and the last one looked at tumor detection. The input features for these studies included clinical features, consisting of demographic data, comorbidities, pre-operative laboratory values, operative variables, tumor variables, and MRIs. Tumors were present in all areas of the spine, including sacral chordomas, sacral giant cell tumors, and metastatic tumors from lung, breast, and prostate. The ML models tested from these 6 studies were deep artificial NN, RF, ResNet50, SVM, LASSO, XGBoost, LightGBM, Faster R-CNN, BPM, Boosted decision tree and CatBoost. The performance of the models was assessed using AUC, or other validation metrics such as Brier score, or dice score, which are defined in Table 1, with some of the studies testing to see which algorithm would perform the best for the desired measurement/outcome.

Table 5.

Summary of Machine Learning Studies Focused on Both Primary and Metastatic Site Malignancy in Spine Oncology.

Authors	Inclusion Criteria	Exclusion Criteria	Input Features	Output	Machine Learning Model	Size of Training Set	Validation Method	Size Test Set	Prediction Performance (CA/AUC)	Other Validation Metrics
Chianca et al, 2021	Spinal bone lesions between 2009 and 2018	-	T1WI and T2WI MRI	Benign or malignant	Deep artificial neural network	100	k-fold cross validation	35	.83 external testing cohort	-
Jin et al, 2022	Inpatient procedure code indicating either laminectomy or corpectomy for intradural lesion removal concurrent with a diagnosis code indicating spinal neoplasm	Patients with null values for all noncomorbidity predictors	Clinical features	Post-discharge features	LASSO	70%		30%	.786	.155 (using an integrated model)
Yin et al, 2019	Patients with sacral tumor mentioned in MRI reports, T2w FS and CE T1w images, available pathology reports	Images with low signal noise ratio (≤ 1.00), lack of CE T1w images, data acquired from 1.5 T MR	MRI	Type of cancer (SC, SGCT, SMT)	Random forest	83		37	.643 (T2W FS and CE T1W combined)	-
Yeh et al, 2022	Selected from the radiological reporting system in a period of 4 years, using the key words fracture, vertebral collapse, pathological, and metastasis	-	MRI (T1W and T2W)	Diagnosis of vertebral fracture	ResNet50		10-fold cross validation		92% (accuracy, compared to 98% by attending radiologist)	-
Gitto et al, 2022	Histoloigcally proven spine tumor	-	MRI (T2W)	Benign or malignant	Support vector machine		10-fold cross validation		.76 (8 features used for this model)
Karabacak et al, 2023	(1) Elective surgery, (2) inpatient operation, (3) current procedural terminology (CPT) codes for surgical resection of spinal tumors, (4) operation under general anesthesia, (5) surgical subspecialty neurosurgery or orthopedics	(1) Emergency surgery, (2) patients with preoperative ventilator dependence, (3) patients with any unclean wounds, (4) patients with sepsis/shock/systemic inflammatory response syndrome 48 h before surgery, (5) patients with ASA physical status classification score of 4 and 5 or non-assigned	Clinical features	Prolonged LOS, nonhome discharges, and postoperative complications	XGBoost, LightGBM, CatBoost, and random forest	1844	-	615	.743, Random forest	-
Karhade et al	Diagnosis of chordoma, primary site of vertebral column, pelvis, sacrum or coccyx, age greater than or equal to 18 years, year of diagnosis b/w 1995-2010, documentation of tumor size and dx confirmed microscopically	-	Clinical features	5-Year survival	Boosted decision tree, support vector machine, bayes point machine		10-fold cross validation			.16 (bayes point machine) brier score
Ouyang et al	-	-	MRI (T1W and T2W)	Detecting tumors	Faster R-CNN	12179		6635	-	-

Two studies used MRIs to predict malignancy status of spine lesions. Chianca et al⁵⁵ used a deep artificial NN to predict malignancy, finding that using features extracted from 3D Slicer heterogeneity CAD and a 2-label classification, yielded the highest accuracy of 86%, compared to PyRadiomics feature extraction and 3-label classification. Gitto et al⁵⁶ used a SVM and found that the best performance using the lowest number of features had an AUC of .78, with only 8 features used for tumor classification. 2 studies used clinical features to predict nonhome discharge and 90-day post-discharge readmission. Jin et al⁵⁷ found an integrated model with LASSO that combined patient-specific and tumor-specific features had an AUC of .786 for nonhome discharge, and .693 for 90-day readmission. Meanwhile, Karabacak et al showed that the RF model performed the best of the ML models with an AUC of .743, and the most accurately predicted outcome was length of stay with an AUC of .745.⁵⁸ In terms of tumor differentiation, Yin et al²⁵ reported that using RF, features extracted from joint T2W fat saturated and contrast enhanced T1W images outperformed those separate imaging modalities with an AUC of .773 in differentiating SC, SGCT, and sacral metastatic tumors. Yeh et al⁵⁹ found that a ResNet50 algorithm was able to identify abnormal vertebrae with 92% accuracy, and it outperformed first-year residents reading the same images. For survival, Karhade et al³³ created an open access web application using a Bayes Point Machine that took clinical and patient features, and predicted survival after chordoma diagnosis, with a C-statistic and Brier score of .8, and .16, respectively. Lastly, Ouyang et al²³ used a version of a NN, denoted as faster R-CNN, to detect spine lesions, with the model being assessed via the Turing test. The study showed that human respondents could not determine if an image detection came from AI or from human detection, as shown by the misclassification rate being above 30% for all 6 respondents, which means that the ML model passed the Turing test. Overall, ML models have performed well on tumor differentiation, tumor detection, predicting post-operative discharge complications, 5-year survival, determining malignancy status, and performing as effective as early-stage trainees in diagnosing vertebral fracture type. (Table 5).

Studies that evaluated mixed groups of spine tumors provided the ability to compare different pathologies, utilizing imaging to predict malignancy and 5-year mortality.

Discussion

Within neurosurgery, the appropriate management of spine tumors is dependent on the evaluation of various preoperative variables, including imaging studies, systemic disease burden, neurologic symptoms, and tumor biology. Patient centered care plans are critical in managing diverse patient populations with a constellation of comorbidities ML algorithms can be trained to analyze large patient datasets consisting of demographic data, clinical presentation, imaging findings, and treatment details, to identify factors associated with clinical outcomes.

Literature published on the use of ML in spine oncology has been increasing in prevalence over the last 30 years (Figure 2). Most of these studies focus on the supplementary role of ML in guiding physicians, surgeons, and other health care providers in treating patients with primary or metastatic spine tumors. Tumor detection, and differentiation using advanced imaging modalities such as MRI, CT, PET/CT, or SPECT may impact treatment pathways through detection of cancer earlier in presentation and more accurate diagnosis leading to an individualized treatment. With an aging population and an increased prevalence of metastatic cancer to the spine, these ML tools may also increase processing speed of providers that read these advanced images, helping to reduce the burden on health care systems.

Several studies compared the performance of their model to a set of experienced physicians. Yoda et al²⁶ developed an algorithm to differentiate between osteoporotic vertebral fractures (OVF) and pathologic vertebral fractures (PVF) that was 96.4% accuracy with TIWI. Using the same data set, 3 independent spine surgeons with 14, 10, and 7 years of experience differentiated OVF and PVF with 91.1%, 88.4%, and 85.7% accuracy. Hoshiai et al⁵¹ found both experienced radiologists and radiology residents demonstrated increased figure of merit when reading CT scans while being assisted by a ML algorithm for lesion identification. These examples point to the efficacy of ML in tumor detection and differentiation with increasing speed and accuracy.

Current research using ML to predict mortality or adverse events focuses on 30-day, 90-day, or 1-year outcomes.^{29,30,32,34–36,50} A study by Li et al published results of 5-year mortality for patients with chordoma using support vector machine, bayes point machine, and neural networks reporting a Brier score of .16. Due to the high variety of pathologies, surgeons may reference this work to identify algorithms that provide data to support clinical decision making.

ML modeling and algorithmic frameworks could potentially be used for tackling other critical topics which the literature is unsure about. For instance, Gui et al demonstrated the effectiveness of a combined clinical and radiomic ML model in predicting fracture 1-year post-SBRT for metastatic spine disease, achieving high sensitivity and specificity.

Multi-institutional collaboration in developing ML application in spine oncology is an interesting strategy to increase its robustness and generalizability. By pooling diverse datasets and leveraging different sources' inputs, a collaborative effort in this regard will expedite the development of this field.

The studies cited in this review demonstrate the usefulness of ML in spine oncology. Continued research is indicated to improve ML ability to detect and differentiate tumors, predict mortality and adverse events, and aid in prognostication.

Limitations

Despite the comprehensive review, there are several limitations to our study. The articles included in this study are heterogenous in their input and output variables, machine learning algorithms, outcome measures, and differ in focuses. The prediction accuracy and AUC represent relative performances of the ML algorithms while the true reliability and generalizability is dependent upon the institution’s patient characteristics, training sets, algorithm used, input quality and size, and patient characteristics. All of these variables can affect the true accuracy and overall generalizability of ML models. Many of the studies included also use a binary outcome allowing the use of AUC measurements. If models utilized a continuous outcome, the results may have been different affecting the model’s clinical efficacy. Due to the variety in these measures, we qualitatively describe the patterns of the use of ML in spine oncology. The variety in measurements also makes quantitative analysis challenging. As with other systematic reviews, publication bias may also be a factor affecting our results. More successful models may have been more likely to be published, falsely inflating the success of ML in spine oncology. Despite this however, we provide a robust review of all the published algorithms utilized in spine oncology, and we provide valuable information regarding the application of ML and future directions of the field.

Future Directions

Current literature includes studies that have investigated the ability of machine learning to predict short- and mid-term survival after surgery for spine tumors.^33,50 However, there is a paucity of literature that evaluates postoperative survival for metastatic disease using machine learning.⁶⁰ In one study, Forsberg et al investigated one- and 6-month survival in metastatic bone disease using ML, demonstrating an AUC of .76.⁶⁰ There is significant clinical utility in expanding upon Forsberg et al's work given that metastatic spine disease occurs at a higher incidence than primary bone tumors. Future studies could investigate longer term survival in patients with metastatic bone disease and stratify cohorts by demographic and clinical factors (ex: age, race, and comorbidities). A multifactorial ML-based analysis applied in a personalized approach would be valuable to neurosurgeons in counseling patients about prognosis.

Compared to metastatic disease, few studies have utilized ML for primary bone tumors.^46,53,61,62 Expansion on optimizing ML models to improve accuracy, improve identification and standardization of radiomic features, and improve detection rates of spinal tumors would prove valuable in resource scarce hospitals that may lack comprehensive imaging modalities such as MRI. Thus, machine learning can support radiologists and spine surgeons in these areas as an adjunctive tool in screening for primary bone tumors based on imaging studies.

Conclusions

Machine Learning research in spine oncology aims to better understand patient outcomes, accelerate treatment planning, detect spine tumor, and differentiate between cancer types. Although these outcomes have been investigated in both primary and metastatic disease, literature trends demonstrate tumor differentiation to play a larger role in the study of metastatic disease. The utilization of machine learning in spine oncology predictive analytics is a promising translational development that has the potential to revolutionize personalized medicine and supplement the neurosurgeon’s armamentarium, also contributing to optimize the adjuvant treatment, including the refinement of radiotherapeutic isodose planning.

Footnotes

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

ORCID iD

Seth B Wilson

References

Hsu

Siu

Pradilla

Gokaslan

Jallo

Gallia

. Animal model of intramedullary spinal cord glioma using human glioblastoma multiforme neurospheres. J Neurosurg Spine. 2012;16:315-319.

Van den Brande

Cornips

Peeters

Ost

Billiet

Van de Kelft

. Epidemiology of spinal metastases, metastatic epidural spinal cord compression and pathologic vertebral compression fractures in patients with solid tumors: a systematic review. J Bone Oncol. 2022;35:100446.

Ciftdemir

Kaya

Selcuk

Yalniz

. Tumors of the spine. World J Orthoped. 2016;7:109-116.

Ottenhausen

Ntoulias

Bodhinayake

, et al. Intradural spinal tumors in adults-update on management and outcome. Neurosurg Rev. 2019;42:371-388.

Mossa-Basha

Gerszten

Myrehaug

, et al. Spinal metastasis: diagnosis, management and follow-up. Br J Radiol. 2019;92:20190211.

Davenport

Kalakota

. The potential for artificial intelligence in healthcare. Future Healthc J. 2019;6:94-98.

Mofatteh

. Neurosurgery and artificial intelligence. AIMS Neurosci. 2021;8:477-495.

Schmidhuber

. Deep learning in neural networks: an overview. Neural Network. 2015;61:85-117.

Habehh

Gohel

. Machine learning in healthcare. Curr Genom. 2021;22:291-300.

10.

ZainEldin

Gamel

El-Kenawy

ESM

, et al. Brain tumor detection and classification using deep learning and sine-cosine fitness grey wolf optimization. Bioengineering (Basel). 2022;10:18.

11.

Gupta

Tran

Luo

, et al. Machine-learning prediction of cancer survival: a retrospective study using electronic administrative records and a cancer registry. BMJ Open. 2014;4:e004007.

12.

Global Burden of Disease 2019 Cancer Collaboration Kocarnik

Compton

, et al. Cancer incidence, mortality, years of life lost, years lived with disability, and disability-adjusted life years for 29 cancer groups from 2010 to 2019: a systematic analysis for the global burden of disease study 2019. JAMA Oncol. 2022;8:420-444.

13.

Page

McKenzie

Bossuyt

, et al. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ. 2021;372:n71.

14.

Chen

Cao

Zhang

, et al. Differentiation between spinal multiple myeloma and metastases originated from lung using multi-view attention-guided network. Front Oncol. 2022;12:981769.

15.

Fan

Yang

Qin

, etal. Machine learning predict survivals of spinal and pelvic ewing’s sarcoma with the SEER database. Global Spine J 2024;14(4):1125-1136. doi:10.1177/21925682221134049

16.

Zhou

Wang

Feng

, et al. Emerging applications of deep learning in bone tumors: current advances and challenges. Front Oncol. 2022;12:908873.

17.

Arends

SRS

Savenije

MHF

Eppinga

WSC

van der Velden

van den Berg

CAT

Verhoeff

JJC

. Clinical utility of convolutional neural networks for treatment planning in radiotherapy for spinal metastases. Phys Imaging Radiat Oncol. 2022;21:42-47.

18.

Chmelik

Jakubicek

Walek

, et al. Deep convolutional neural network-based segmentation and classification of difficult to define metastatic spinal lesions in 3D CT data. Med Image Anal. 2018;49:76-88.

19.

Ito

Ando

Kobayashi

, et al. Automated detection of spinal schwannomas utilizing deep learning based on object detection from magnetic resonance imaging. Spine. 2021;46:95-100.

20.

Lemay

Gros

Zhuo

, et al. Automatic multiclass intramedullary spinal cord tumor segmentation on MRI with deep learning. Neuroimage Clin. 2021;31:102766.

21.

Maki

Furuya

Horikoshi

, et al. A deep convolutional neural network with performance comparable to radiologists for differentiating between spinal schwannoma and meningioma. Spine. 2020;45:694-700.

22.

Netherton

Nguyen

Cardenas

, et al. An automated treatment planning framework for spinal radiation therapy and vertebral-level second check. Int J Radiat Oncol Biol Phys. 2022;114:516-528.

23.

Ouyang

Meng

Liu

, et al. Evaluation of deep learning-based automated detection of primary spine tumors on MRI using the turing test. Front Oncol. 2022;12:814667.

24.

Wang

Fang

Lang

Yuan

Baldi

. A multi-resolution approach for spinal metastasis detection using deep Siamese neural networks. Comput Biol Med. 2017;84:137-146.

25.

Yin

Mao

Zhao

Chen

Hong

. A triple-classification radiomics model for the differentiation of primary chordoma, giant cell tumor, and metastatic tumor of sacrum based on T2-weighted and contrast-enhanced T1-weighted MRI. J Magn Reson Imag. 2019;49:752-759.

26.

Yoda

Maki

Furuya

, et al. Automated differentiation between osteoporotic vertebral fracture and malignant vertebral fracture on MRI using a deep convolutional neural network. Spine. 2022;47:E347-E352.

27.

Gao

RLQ

Lou

Man

Cao

Wang

. Automatic segmentation of metastasis lesions in SPECT bone scan images. In: International Conference on Intelligent Computing, Automation and Systems (ICICAS). Chongqing, China, 11-13 Dec: IEEE; 2020.

28.

Bakhsheshian

Shahrestani

Buser

, et al. The performance of frailty in predictive modeling of short-term outcomes in the surgical management of metastatic tumors to the spine. Spine J. 2022;22:605-615.

29.

Bongers

MER

Karhade

Villavieja

, et al.

Does the SORG algorithm generalize to a contemporary cohort of patients with spinal metastases on external validation?

Spine J. 2020;20:1646-1652.

30.

DiSilvestro

Veeramani

McDonald

, et al. Predicting postoperative mortality after metastatic intraspinal neoplasm excision: development of a machine-learning approach. World Neurosurg. 2021;146:e917-e924.

31.

Yen

Chen

, et al. Decreased psoas muscle area is a prognosticator for 90-day and 1-year survival in patients undergoing surgical treatment for spinal metastasis. Clin Nutr. 2022;41:620-629.

32.

Karhade

Fenn

Groot

, et al. Development and external validation of predictive algorithms for six-week mortality in spinal metastasis using 4,304 patients from five institutions. Spine J. 2022;22:2033-2041.

33.

Karhade

Thio

Ogink

, et al. Development of machine learning algorithms for prediction of 5-year spinal chordoma survival. World Neurosurg. 2018;119:e842-e847.

34.

Karhade

Thio

Ogink

, et al. Predicting 90-day and 1-year mortality in spinal metastatic disease: development and internal validation. Neurosurgery. 2019;85:E671-E681.

35.

Massaad

Bridge

Kiapour

, et al. Evaluating frailty, mortality, and complications associated with metastatic spine tumor surgery using machine learning-derived body composition analysis. J Neurosurg Spine. 2022;37:263-273.

36.

Fourman

Siraj

Duvall

, et al.

Can we use artificial intelligence cluster analysis to identify patients with metastatic breast cancer to the spine at highest risk of postoperative adverse events?

World Neurosurg. 2023;174:e26-e34.

37.

Jiao

Yang

. A novel prediction tool for overall survival of patients living with spinal metastatic disease. World Neurosurg. 2020;144:e824-e836.

38.

Wang

, et al. Dynamic predictive models with visualized machine learning for assessing chondrosarcoma overall survival. Front Oncol. 2022;12:880305.

39.

Ryu

Lee

Kim

Eoh

. Predicting survival of patients with spinal ependymoma using machine learning algorithms with the SEER database. World Neurosurg 2018;28:S1878-8750(18)32914-0. doi:10.1016/j.wneu.2018.12.091

40.

Ryu

Seo

Lee

. Novel prognostication of patients with spinal and pelvic chondrosarcoma using deep survival neural networks. BMC Med Inf Decis Making. 2020;20:3.

41.

Wei

Nistal

Sobotka

Martini

Hawks

Jenkins

. New predictive index for survival in symptomatic spinal metastases. World Neurosurg. 2019;123:e133-e140.

42.

Elsamadicy

Koo

Reeves

, etal. Utilization of machine learning to model important features of 30-day readmissions following surgery for metastatic spinal column tumors: the influence of frailty. Global Spine J 2024;14(4):1227-1237. doi:10.1177/21925682221138053

43.

Gui

Chen

Sheikh

, etal. Radiomic modeling to predict risk of vertebral compression fracture after stereotactic body radiation therapy for spinal metastases. J Neurosurg Spine 2021;36(2):294-302. doi:10.3171/2021.3.SPINE201534

44.

Massaad

Williams

Hadzipasic

, et al. Performance assessment of the metastatic spinal tumor frailty index using machine learning algorithms: limitations and future directions. Neurosurg Focus. 2021;50:E5.

45.

Yen

Zijlstra

, et al. Prognostic significance of lab data and performance comparison by validating survival prediction models for patients with spinal metastases after radiotherapy. Radiother Oncol. 2022;175:159-166.

46.

Lang

Zhang

, et al. Differentiation of spinal metastases originated from lung and other cancers using radiomics and deep learning based on DCE-MRI. Magn Reson Imaging. 2019;64:4-12.

47.

Liu

Wang

Guo

, et al. A preliminary study using spinal MRI-based radiomics to predict high-risk cytogenetic abnormalities in multiple myeloma. Radiol Med. 2021;126:1226-1235.

48.

Paulino Pereira

Janssen

van Dijk

, et al. Development of a prognostic survival algorithm for patients with metastatic spine disease. J Bone Joint Surg Am. 2016;98:1767-1776.

49.

Ren

Yang

Lai

, et al. MRI-based radiomics analysis for predicting the EGFR mutation based on thoracic spinal metastases in lung adenocarcinoma patients. Med Phys. 2021;48:5142-5151.

50.

Karhade

Thio

Ogink

, et al. Development of machine learning algorithms for prediction of 30-day mortality after surgery for spinal metastasis. Neurosurgery. 2019;85:E83-E91.

51.

Hoshiai

Hanaoka

Masumoto

, et al. Effectiveness of temporal subtraction computed tomography images using deep learning in detecting vertebral bone metastases. Eur J Radiol. 2022;154:110445.

52.

Kowalchuk

Waters

Richardson

, et al. Stereotactic body radiation therapy for spinal metastases: a novel local control stratification by spinal region. J Neurosurg Spine. 2020;34:267-276.

53.

Yin

Mao

Zhao

, et al. Comparison of radiomics machine-learning classifiers and feature selection for differentiation of sacral chordoma and sacral giant cell tumour based on 3D computed tomography features. Eur Radiol. 2019;29:1841-1847.

54.

Zhuo

Zhang

Duan

, et al. Automated classification of intramedullary spinal cord tumors and inflammatory demyelinating lesions using deep learning. Radiol Artif Intell. 2022;4(6):e210292.

55.

Chianca

Cuocolo

Gitto

, et al. Radiomic machine learning classifiers in spine bone tumors: a multi-software, multi-scanner study. Eur J Radiol. 2021;137:109586.

56.

Gitto

Bologna

Corino

VDA

, et al. Diffusion-weighted MRI radiomics of spine bone tumors: feature stability and machine learning-based classification performance. Radiol Med. 2022;127:518-525.

57.

Jin

Feng

, et al. Prediction of discharge status and readmissions after resection of intradural spinal tumors. Neurospine. 2022;19:133-145.

58.

Karabacak

Margetis

. A machine learning-based online prediction tool for predicting short-term postoperative outcomes following spinal tumor resections. Cancers. 2023;15:812.

59.

Yeh

Zhang

Chen

, et al. A deep learning-based method for the diagnosis of vertebral fractures on spine MRI: retrospective training and validation of ResNet. Eur Spine J. 2022;31:2022-2030.

60.

Forsberg

Wedin

Boland

Healey

. Can we estimate short- and intermediate-term survival in patients undergoing surgery for metastatic bone disease? Clin Orthop Relat Res. 2017;475:1252-1261.

61.

Mehta

Sebro

. Random forest classifiers aid in the detection of incidental osteoblastic osseous metastases in DEXA studies. Int J Comput Assist Radiol Surg. 2019;14:903-909.

62.

Pan

Bao

, et al. Deep learning-based classification of primary bone tumors on radiographs: a preliminary study. EBioMedicine. 2020;62:103121.

Machine Learning in Spine Oncology: A Narrative Review

Abstract

Study Design

Objective

Methods

Results

Conclusion

Keywords

Introduction

Background on Machine Learning

Methods

Search Strategy

Inclusion and Exclusion Criteria

Data Extraction

Results

Metastatic Spine Disease

Primary Bone Tumors

Intradural

Mixed Studies

Discussion

Limitations

Future Directions

Conclusions

Footnotes

Declaration of Conflicting Interests

Funding

ORCID iD

References