Current Applications of Machine Learning in Spine: From Clinical View

Abstract

Study Design:

Narrative review.

Objectives:

This review aims to present current applications of machine learning (ML) in spine domain to clinicians.

Methods:

We conducted a comprehensive PubMed search of peer-reviewed articles that were published between 2006 and 2020 using terms (spine, spinal, lumbar, cervical, thoracic, machine learning) to examine ML in spine. Then exclude research of other domain, case report, review or meta-analysis, and which without available abstract or full text.

Results:

Total 1738 articles were retrieved from database, and 292 studies were finally included. Key findings of current applications were compiled and summarized in this review. Main clinical applications of those techniques including image processing, diagnosis, decision supporting, operative assistance, rehabilitation, surgery outcomes, complications, hospitalization and cost.

Conclusions:

ML had achieved excellent performance and hold immense potential in spine. ML could help clinical staff to improve medical level, enhance work efficiency, and reduce adverse events. However more randomized controlled trials and improvement of interpretability are essential to clinicians accepting models’ assistance in real work.

Introduction

Due to popularity of soft computing approaches, artificial intelligence (AI) had made great impact on every aspect of daily life. AI techniques are revolutionizing medical domain by performing complex and huge computational tasks. Nowadays AI had made major progress in healthcare administration, clinical decision support, patient monitoring and healthcare interventions.^1,2 As the most promising branch of AI, machine learning (ML) automatically predict outputs based on features of inputs through algorithms.^3,4 ML have natural advantage of handling big data comparing to traditional statistical methods. They have more superior accuracy and repeatability than conventional models and even expert operators. ML could provide subtle information, which cannot detected by eye in desired image tasks. ⁵ In the era of big data, ML will dramatically improve diagnostic accuracy and prognosis.^3,6

Publications about applications of ML in spine significant increased recently. However, the development of those techniques for spine are still in infancy. Before clinicians adopt ML in the practical work, preclinical steps are raising their attention, establishing elementary cognition and getting involved in research. We aim to introduce current applications of ML in spine to clinicians. Then identify opportunities and utilization potentiality of future research.

Methods

We conducted a comprehensive PubMed search of peer-reviewed articles that were published between 2006 and 2020 using terms (spine, spinal, lumbar, cervical, thoracic, machine learning) to examine ML in spine. Then exclude research of other domain, case report, review or meta-analysis, and which without available abstract or full text.

Results

Total 1738 articles were retrieved from database, and 292 studies were finally included. Count of articles significantly increased in past 3 years, that indicates growing interest of researchers. Key findings of current applications were compiled and summarized in this review.

Machine Learning

In 1959, Arthur Samuel defined “machine learning” as giving computers the ability to learn without being explicitly programmed. ⁷ Simply put, ML is that algorithms get the ability of making decisions or predictions through learning data. Models present an analysis and generate desired outputs basing on inputs features. Main ML tasks contain classification, regression, clustering, and dimensionality reduction. ML process comprised by data collection, data preprocessing, feature engineering, model selection, training, model evaluation and optimization. Data collection is an important step, as quality and quantity of data directly affect outcome. Secondly, data preprocessing is to make data usable for computation, improve accuracy, and shorten calculation process. Feature engineering is creating features from raw data using domain knowledge. That largely determines the final algorithms performance, which can be broadly classified as feature extraction, feature construction and feature selection. Next, dataset is divided into training set, cross validation set, and testing set. Models are built and trained on training sets. Then models are scored in cross validation set for selecting better one. And various evaluation methods should be brought to assess models. For optimization, the most basic method is gradient descent algorithm, which minimize loss function through iterating.

ML was classified into supervised learning, unsupervised learning and reinforcement learning according to different forms. Supervised learning is that a learner describes the input-output relationship based on labeled input variables with a grounded truth. ⁸ A model analyzed the training data to synthesize the pattern between independent variables and dependent variables. Then the testing dataset is to be predicted. In contrast, the unsupervised learner describes relationship of input-output basing on unlabeled inputs. ⁹ Algorithm analyses input data features to identify clusters of data. ¹⁰ Reinforcement learning is that the learner constantly interacts with environment to find the best strategy through trial and error for maximizing rewards. ¹¹

Three common ML models are decision tree learning, support vector machines (SVMs), and artificial neural networks (ANNs). Classification and regression decision tree (CART) implements a classification or a regression task, which is more visible and easier to understand than other modalities. The tree comprises internal nodes (conditions), branches (decisions), and leaves (ends), that is not computationally intensive and therefore suitable for big data^5,12 (Figure 1). SVMs accomplishes classification tasks by creating a maximum margin hyperplane between 2 outcomes, or regression tasks by plotting a best-fit plane ¹³ (Figure 2). ANNs is a deep machine learner inspired by how neurons connected and interacted in brain. ¹⁰ Constructing in Hebbian learning, single neurons comprises an entire network, where information flows from the inputs layers to the outputs layers through multiple hidden layers, operating by weights ⁵ (Figure 3).

Figure 1.

Schematic representation of decision tree learning algorithm.

Figure 2.

The simple support vector machine as a binary classification.

Figure 3.

The deep artificial neural network comprised by input layer, output layer, and multiple hidden layers.

Underfitting and overfitting are 2 classical problems in ML. If a model is too simple to fully learn features of training data, it will present low accuracy both on the training set and test set, which is called underfitting. Overfitting is that a model performs well on training set, but poor on test data, since adjusting too closely to former. The model memorizes specific training observations, but cannot actually extract generalizable relationships between variables. ¹⁴ Common solutions of overfitting are increasing sample, regularization, and dropout technique.

Image Processing and Diagnosis

Image processing

Intention of developing ML is to reduce manual labor and save time for tasks needed judgment.^15,16 Localization, segmentation, grading, calculating parameters of image are time-consuming works, which should be accurately performed by algorithm. Several models accurately detected and quantitative measurement, comparable with human-generated segmentations. Rak et al proposed a novel approach to automatically segment vertebrae in three-dimensional MRI. ¹⁷ That CNN combined with a graph cut formulation based on encoding swaps, which segments multiple vertebrae without risking ambiguous segmentations of adjacent vertebrae. ¹⁷ FU-Net combined traditional region-based level set with CNN to accurate segment vertebrae, and outperformed other techniques. ¹⁸ Overall, deep learning pipelines for segmenting vertebrae and intervertebral discs (IVDs) on MRI provided high accuracy and eliminated laborious manual labeling.^19,20

Paugam et al used supervised neural networks to perform single-class or multi-class segmenting the grey and white matter of spinal cord on MRI, which showed good results on small amounts of data. ²¹ In addition, convolutional neural networks (CNNs) model showed state-of-the-art performance for automated lesion segmentation after spinal cord injury (SCI). ²² Volumes extracted from lesion segmentation were significantly associated with patient motor scores, that will potentially help to advance modernized SC MR image analysis for both research and clinical application. ²² Likewise, ML algorithms were trained to segment lumbar spinal canal areas on axial views of lumbar MRI. ²³ And deep multiscale multitask learning network (DMML-Net) achieved highly accurate localization and grading of neural foramina, surrounding vertebra and discs for the pathogenesis-based diagnosis of lumbar neural foraminal stenosis. ²⁴

Besides segmentation, models were also developed for automatically identifying organ and lesion. Jimenez-Pastor et al introduced a 2-stage decision forests and morphological image processing technique to automatically detect and identify vertebral bodies from arbitrary field-of-view body CT scans. ²⁵ Vertebrae automatic detection network achieved satisfying accuracy and precision on CT or unannotated MRI.^26,27 An automated framework could detect myelopathic areas, combining diffusion tensor imaging (DTI) metrics with SVM. ²⁸ Another random forest (RF) model ²⁹ had potential benefit of diagnosing and locating the level of cervical injury in CM, using time-frequency components of somatosensory evoked potentials.

Algorithms had automatic measured parameters of spine, comparable to manual measurements. There were fully automatic algorithms to evaluate of lumbar lordosis and Cobb angle for scoliosis.^30,31 A CNN model was able to determine the spine shape and calculate posture parameters in biplanar radiographs, which could act on intervention in deformities or degeneration. ³² Another novel cascade amplifier regression network robustly auto-mated quantitative measure spine on T1-weighted MR. ³³

Diagnosis

In the outpatient primary care setting, a general practitioner must determine to refer a patient to spinal subspecialist for consultation when discover specific radiographic findings, however large proportion of referrals ultimately fail to meet criteria for surgical intervention.^34,35

ML could provide auxiliary information to optimize the referral process and save time of patients and surgeons. ³⁴ Deep learning had been reported to successfully apply for diagnosis of liver masses, parkinsonian disorders, hip fractures and estimating bone age.^36-40 Combining multiple data, ML can provide accurate diagnostic predictions and risk warnings. With increasing applications of CNNs to radiological imaging, AI was expected to gradually change clinical practice. ⁴¹

ML obtained fine diagnostic accuracy in disease prediction and differentiation, mainly running through radiological imaging. A deep neural network (DNN) had predicted cervical myelopathy (CM) based on MRI. ⁴² CNNs had well differentiated spinal schwannomas and meningiomas as well as tuberculous and pyogenic spondylitis using MRI.^43,44 Likewise, ML methods had been developed for preoperative differentiation of cancer based on 3D CT, enhanced CT, and enhanced MRI.^15,45-47

Similarly, clinical and radiographic data-driven ML models showed excellent prediction of radiographic progression in axial spondyloarthritis (axSpA).^48,49

ML models could obtain features that cannot be detected by naked eye. Texture features of IVDs and endplate zones are different between people with and without low back pain (LBP), that may not be consistently discovered by clinicians. ⁵⁰ RF and linear mixed-effect model analysis identified texture features from lumbar MRI to assist surgeons recognizing people with LBP. ⁵⁰ Intermittent claudication may cause by multiple diseases, like lumber spinal canal stenosis (LSS) and peripheral arterial disease. Texture analysis with ML offered highly reproducible quantitative parameters that increase accuracy for severe LSS detection. A decision tree classifier revealed higher performances for LSS grading compared to qualitative assessments using the reference cut-off cross-sectional area (CSA). ⁵¹

ML also added extra information from other disease examination or routine screening. These features may overlook by specialist who just focus on their domain. Additional health assessment will enhance healthcare, save payment and decrease examinational radiation. Several tools tended to be excellent approaches for distinguishing osteoporosis, such as metabolites, low-frequency guided waves, ocal classification of CT textures, segments of DXA images, and spine radiographs.^52-61 Currently a few deep learning models predicted high-risk osteoporosis populations using spine X-ray. ⁶² The best method employed VGGnet for feature extraction and RF training. ⁶² Yasaka et al ⁶³ operated a CNNs with lumbar vertebrae CT to output BMD, which significantly correlated with DXA’s outcome and perform better than vertebrae CT values. Models could be applied to predict osteoporosis through abdominal CT or chest CT that widely used in daily clinical practice.^63,64 Likewise, SVM classifiers could use dual-energy X-ray absorptiometry (DEXA) studies to identify lumbar spine fractures routinely. ⁶⁵ Jamaludin et al developed an automatic ML model to accurately identify and quantify scoliosis from total body DXA. ⁶⁶

Besides intuitive radiomics, ML combining various auxiliary examination can provide accurate diagnosis. LBP has physiological relationships with abnormal muscle activation. ⁶⁷ Clinically interpretable models produced good to excellent predictive capability to LBP using functional kinematic and electromyography (EMG) variables with FDboost. ⁶⁷ Another SVM analyzed two-dimensional walking motion, and classified the underlying disease of the intermittent claudication with 79.7% accuracy. ⁶⁸ Also SVM had applied to distinguish chronic lumbar radiculopathy on electroencephalography (EEG). ⁶⁹ In addition, ML may offer the initial suspected diagnosis based on a simple test which is more available in regions with limited medical resources. ⁷⁰ Objective functional impairment (OFI) could provide hints for the suspected cause of back and leg pain, for which Staartjes et al carried out five-repetition sit-to-stand test combining with a ML algorithm. ⁷¹ Tan et al ⁷² developed and validated a natural language processing (NLP) system to identify 26 findings related to LBP from x-ray and MR radiology reports. Furthermore, deep learning algorithms were devised for automatic scoliosis screening and classifications, using unclothed back images. ⁷³

Treatment

ML models had been put into treatment decision-making process, such as surgery level, instrumented vertebra and correction range. SpineNet automatically generated MRI grading to predict the surgical level of single-level decompression. ⁷⁴ It calculated an aggregate score of grading for the following: central canal stenosis, disc narrowing, disc degeneration, spondylolisthesis, upper/lower endplate morphologic changes, and upper/lower marrow changes. For adult spinal deformity correction, an ANN successfully mimicked lead surgeons’ decision making in selection of the upper instrumented vertebra. ⁷⁵ In another literature, Shen et al ⁷⁶ presented a new classifier to assist surgeons optimizing surgical plan and therapeutic management. The method classified adolescent idiopathic scoliosis (AIS) using a fuzzy clustering algorithm and 3D parameters describing characteristics of deformity. ⁷⁶ More recently, a decision tree identified optimal range of spontaneous lumbar Cobb correction in Lenke 1 AIS. ⁷⁷ Also, a semi-automatic method had been built to classify scoliosis severity and treatment group with 3D markerless surface topography scans. ⁷⁸ And approaches were proposed to predict curve shape variation, curve types and progression in AIS.^79-81

Recently scholars built models to automatically provide measurements of implants, placement trajectories, pedicle screw planning and surgical navigation.^82-84 An automated pedicle detector provided robust quasi-automated pedicle localization by calculating vertebral axial rotation values in frontal radiographs of scoliosis with minimal user intervention. ⁸⁵ That was useful to vertebral rotation estimation and 3D spine reconstructions. Another intraoperative 3D pedicle screws navigation system’s accuracy was 86.1%. ⁸³ von Atzigen et al proposed a purely ML marker-less surgical navigation to bending rod implants. ⁸⁶ This method required significantly less time than the marker-based benchmark navigation approach needed to contact with anatomy and achieved better or comparable accuracy. ⁸⁶ Other pullout strength models had predicted combination of density, insertion depth, and insertion angle for the chosen range, for understanding pedicle screw pullout and pre-surgical planning.^87,88 Besides, deep learning-based framework could automatically adjust the C-arm pose to a desired standard projection from the first X-ray, and localize needle target for epidural needle placement in ultrasound images.^89,90

AI combining with virtual reality simulation provide safer training and objective assessment of surgical skills, leading to improved patient care. ANNs will gain insight into the importance of virtual reality surgical simulators for surgical training. Mirchi et al performed ANNs to distinguish safety metrics in a virtual reality-simulated anterior cervical discectomy scenario. ⁹¹ An ML tool assessed surgical expertise in a virtual reality spine procedure, and potentially apply to ensuring surgeons’ technical competency in future. ⁹² Additionally, identifying patterns of surgical practice will be an important step to understand surgical processes. Researchers built a framework to automatically identify practical patterns for discriminating different experience surgeon from surgery recordings. ⁹³

In rehabilitation of SCI patients, ML framework could support researchers and clinicians for selection of epidural stimulation parameters. ⁹⁴ During electrically evoked contractions, SVM increased safety by adapting the functional electrical stimulation parameters in motor complete SCI individuals. ⁹⁵ Another three-steps ML model provided monitor of tenodesis grasp based on egocentric video at home, that implied remote cSCI therapeutic guidance. ⁹⁶

Prognosis

Clinical outcomes

Preoperative prediction for clinical outcomes could enhance informed patient consent, reduce drug consumption, promote recovery and personalize shared decision-making.^97,98 ML algorithms indicated that lower preoperative PROMIS scores, fewer comorbidities, and certain sociodemographic factors increased the likelihood of achieving minimal clinically important difference (MCID), which helps surgeons to determine the appropriateness and timing of surgery. ⁹⁹ Algorithms had been explored for analyzing of multiple data(patient demographics, clinical presentation, DTI maps) to determine the prognosis in CM.^{100, 101} A few models were designed for predicting outcome like, VAS, ODI, mJOA, and invasiveness score based on preoperative factors in lumbar disc herniation (LDH) or LBP patients. ¹⁴ , ^102-105 Similarly, ML meaningful predicted survival outcomes of spinopelvic chondrosarcoma, ependymoma, malignant peripheral nerve sheath tumor and spinal metastases patients.^106-110 In spinal metastatic disease, SORG algorithms had been externally validated for survival prediction.^{111, 112}

Complications

Adverse events (AEs) following spine surgery negatively impact patients, surgeons, and the health care system. Therefore, it is critical to build predictive models for investigating factors associating with AEs and develop risk stratification strategies. ¹¹³ Researchers presented a set of predictive models for postoperative AEs. They accounted for patient-, diagnosis-, and procedure-related factors which could contribute to timely intervene, patient-counseling, accurate risk adjustment, and quality metrics. Algorithms accurately predicted the risk of proximal junctional kyphosis and spinopelvic compensation for long level fusion surgery. ^{114 115} ANN and LR models achieved better performance than the ASA classification for predicting complications of ACDF. ¹¹⁶ ML were more accurate than ASA scores for complications risk factor analysis. ^{117 118} Also, Natural language processing (NLP) algorithms automatically detected incidental durotomies, intraoperative vascular injury, and postoperative wound infection in lumbar surgery.^119-121 As a result, ML algorithms could provide prediction of AEs, improve risk stratification and help guide the surgical decision-making process in spine surgery. ¹²²

Hospitalization and Cost

Hospital readmission and prolonged length of stay (LOS) will bear a great burden to the healthcare system. ML improved understanding of predicting readmission after spinal surgery.^{123, 124} Risk features includes returning to operating room, septic shock and superficial surgical site infection. ¹²³ A RF model suggested demographic features may contribute more readmission risk than perioperative variables. ¹²⁵ Similarly, models recommended that non-electively admitting and staying in ICU need additional attention to avoid unanticipated prolonged LOS following spine surgery ¹²⁶

Likewise, ML was employed to predict the medical costs of spinal fusion, and inform hospital strategy to increase the financial management efficiency.^{127, 128} Agglomerative hierarchical clustering was applied to identity factors associated with higher 2-year post-surgical costs, containing greater utilization of antidepressants, opioids, and behavioral health services. ¹²⁹

Discussion

Although ML models has got fine outcome currently, there is still more effort need to make for ML development in the practical spine work. Several limits hinder the application of ML. First, single kind data and small size sample cannot well represent complex disease. Then traditional models may have similarly or even better predictive ability than deep learning in the low order of magnitude data. But deep learning has superiority in processing big data and figuring out complex relationships between variables. ¹³⁰ More advanced algorithms should be leaded in this field. Finally, we discuss the outlook and challenge of ML.

• 1. Data

The clinical dataset is typical multimodal heterogeneous data including, clinical indicators, images, genetic data and biomarkers. Doctors need to cautiously offer a proposal in view of individual difference and various comorbidity. However, studies mainly based on one kind data like pharmaceutical, clinical or imaging data.^{74, 102, 131} For instance, predication from pharmaceutical data couldn’t achieve clinical causes of risk factors without complete clinical, symptomatic and imaging data. ¹³¹ In another article for postoperative endpoints, predictive tool did not consider features such as quality of life and objective functional impairment, anxiety and depression, severity of stenosis, comorbidities, or neurological deficits. ⁹⁷ In a word, based on single kind of data models was hard to be fully convinced by decision makers, and provide substantial assistance. It is essential that combining multimodal heterogeneous data to improve value of outcome. In the future, deep learning will establish a connection of radiomics, proteomics, and metabolomics. ¹³² Another common issue is small size of sample, which is expediently obtained to medical staff. With many variables, department level data may lead to overfitting and poor performance. Multicenter data is crucial to ML performance. Also, single center analysis needs external validation from multiple institutions to guard against institutional biases. ¹³³

• 2. Transfer learning:

In medical domain, lacking huge training data is a prominent problem. Main cause are expensively professional preprocessing tasks and privacy protection. Fortunately, we can build a model using annotated data that similar with target data, and then employ transfer learning methods. Transfer learning is that applying models learned from old field, in a new field through similarities of data, tasks, or models. ¹³⁴ The meaningful approach can utilize existing large image datasets to perform image analysis and obtain even faster convergence and higher accuracy than training from scratch.^{135, 136} This process identifies and integrates one or more types of domain knowledge relating to the designated task for improving deep learning models performance. ¹³⁷ Recently, transfer learning was employed in deep learning tasks, as pathology image analysis and COVID-19 screening.^{138, 139}

Unsupervised domain adaptation (UDA) is a type of transfer learning with restrictive conditions. It is necessary that source domain and target domain have same label space, feature space, and conditional probability distribution. With labels only from the source domain, UDA could promote neural networks performance on target domain. That especially suits for medical field, as quite labor-intensive annotating and commonly lack of annotated datasets. ¹³⁷ The effectiveness of UDA had been proved in chest X-ray segmentation. ¹⁴⁰

In addition, treating issues as many single tasks (diagnosis or prognosis) will ignore the correlation information of tasks. Multi-task learning (MTL), another domain of transfer learning, is a suitable solution to this problem. The approach is learning multiple related tasks together and sharing information, they learned. Main goal of MTL is to improve generalization ability by using specific domain’s information in multiple related tasks training. ¹⁴¹ Extra information can enhance performance of the current task, such as generalization accuracy, learning speed, and comprehensibility. Pan et al presented a multi-task disease progression model based on hierarchical attention mechanism, that focus on different features and medical records. ¹⁴² In the future, it is significant that a comprehensive deep learning model simultaneously work out spine diagnosis and prognosis tasks.

For privacy protection, the federated transfer learning (FTL) is a method that different participants first respectively train models on their own data. Then encrypt models and data for avoiding disclosure of participants’ privacy. On this basis, these models are joint trained to get the final optimal model, which will be returned to each participant. ¹⁴³ The FTL framework could improve original ML algorithms, that has been verified on public datasets, such as NUS-WIDE and Default-Credit datasets. ¹⁴⁴ Also, privacy-preserving FTL had been proposed to extract common discriminative information from multi-subject EEG data for EEG classification. ¹⁴⁵

• 3. Disease progression models:

Spinal degenerative diseases, main component of spinal diseases, are typical chronic diseases. Models were necessary to gather data at different time points from time series database, obtain pattern of disease progression and make accurate prediction, in order to timely intervention and avoiding progression or deterioration of diseases. Disease progression models (DPMs) is employed to characterize the course of disease progression from longitudinal health records, for early detection and precision care at the appropriate time. ¹⁴⁶ DPMs had been applied in progression of Alzheimer’s disease (AD), and breast cancer lung metastasis.^147-149 For instance, Hidden Markov models (HMMs) could infer discrete latent states and transitions between inferred states from time-varying multivariate data. ¹⁴⁶ Kwon et al used HMMs and its variants to discover disease states and make inferences of health states for chronic patients. ¹⁴⁶ For doctors understanding patient status and progression patterns, they developed DPVis incorporating HMMs to seamlessly integrates HMMs’ parameters and outcomes into interpretable and interactive visualizations. ¹⁴⁶ Oxtoby et al used 2 generative data-driven DPMs to extract patterns of observable biomarker changes in dominantly-inherited AD. ¹⁴⁹ Their models reveal probabilistic sequences of biomarker abnormality and estimates of biomarker trajectories through a non-parametric differential equation model from a cross-section of short-term longitudinal data.

• 4. Incremental learning:

AI has a main criticism of catastrophic forgetting, that is when moving from 1 task to another, models perform well on the new task, but underfit the old one. It is due to original data memory is overwritten by new one. We hope models get ability of gradually sustained accumulating knowledge like human. Incremental learning is constantly learning knowledge from new data and preserving most of previous knowledge. ¹⁵⁰ The learning could avoid historical data occupying storage space. On the other hand, it makes full use of previous results, and thus significantly saves time for new training. This learning is mainly applied for big data, especial suitable to medical field. Common methods of incremental learning including feature extraction, fine-tuning, joint training, and knowledge distillation. ¹⁵¹ Recently incremental learning has been widely employed in domains, such as muscle activity and kinematics, nonconvulsive epileptic seizure, atrial fibrillation, and segmentation anatomical structures.^152-155

• 5. Outlook and challenge:

In the past 2 decades, ML field had been driven by huge development of computing power. Currently, the photonic quantum computer, Jiuzhang, is faster than using the state-of-the-art simulation strategy and supercomputers by a factor of ∼10¹⁴ on boson sampling task. ¹⁵⁶ Although, it cannot be used for other calculation, and still far from practical application at present. Someday in the future, quantum computer will produce computing forces beyond classical computers, process big data with less time and energy dissipation, and extend the scope of ML.

Although achievement of universal health coverage (UHC) was policy priority and ultimate goal for both countries and global institutions, estimated by current projections about 3.1 billion people will still lacking UHC effective coverage in 2023. ¹⁵⁷ Several determinants hindered UHC progress, like wealth inequality, race, gender, caste discrimination, air pollution, lack of water and sanitation facilities. ¹⁵⁸ There are a large number of disadvantaged groups hard to achieve good-quality healthcare and afford the corresponding cost. In the COVID-19 pandemic era, the global economy is entering a major crisis. COVID-19 pandemic decreases limited medical resources and capacity of government financial support. Meanwhile, home quarantine leads to tremendous sicken people cannot getting treatment for common disease. Patients and accompanying members pay more time and expenditure on virus screening in questing outpatient service. In this special period, illness, losing job and financial hardship are making matters worse especially for impoverished people. That lead to poor members even harder getting good-quality healthcare than at ordinary times. ML may be powerful approaches to enhance poor peoples’ healthcare, save government medical financial support, and promote the UHC progress. A study introduced clinical decision support system (CDSS) support self-referral for patients with LBP and further referral by healthcare professionals, that has the potential to decrease the current long waiting lines in healthcare. ¹⁵⁹ Popularization of ML will be a solution for poor people achieving daily healthcare and confronting major health event. If they could use an app to get a preliminary screening of disease, they will get the correct registration of outpatient and even the initial diagnosis at first time. These mobile applications could save time, avoid wrong registration, decrease risk of delaying disease, and relieve pressure of limited medical resources. Indeed, they could be totally free to poor members, which will make significant impact on UHC. With lowest cost, disadvantaged groups could achieve medical service approaching expert level. In someday, patients could achieve meaningful advice from ML, including what disease may have, where to get medical service, and which suitable expert to perform the operation. These models may remove disequilibrium of medical resources. They could provide contribution to rational allocation of government medical expenditures, reducing the risk of dissemination disease, saving insurance expenses of countries and companies.

As improvement of external performance and interpretability, CNNs are anticipated to help radiologists maximize the value of extracted image, achieve diagnostic excellence, enhance interrater reliability, and improve workflow for more timely recommendations. ⁴¹ ML could also assist surgeons to decide preoperative patient-specific planning and implant, for enhancing patient healthcare. ¹⁶⁰ The computer assisted navigation systems will be an essential tool to reduce adverse event, save operative time, and surmount the technical barrier. ML tools could bring the state-of-the-art technology to the remote region. An ideal single platform should affords central control over functions, like data pre-processing, data governance, regulatory requirements, and operational interaction with existing electronic health record systems. ¹⁶ ML has the capacity to primarily generate structured data from raw electronic health records, thus it could have a strong impact on hospital analytics, epidemiological studies, and systematic reviews. ¹⁶¹

Although AI had tremendous development in healthcare, only few applications had been actually implemented evaluated at the frontlines of clinical practice. ¹⁶² The rigorous and comprehensive evaluation of clinical AI needs to be improved by more robust randomized controlled trials in the future. ¹⁶² There were several challenges need to overcome in the further practical application of ML. Black box is that, most AI technologies operate largely by an opaque logic, which are hardly understandable to users. This ethical challenge limits marketing approval of AI in medicine. For instance, who is responsible for misdiagnosis among doctors, system and AI manufacturers. ¹⁶³ The local interpretable model-agnostic explanations is a mean to improve the interpretability of model. ¹⁶⁴ Another important hurdle is that the lack of standardized regulations related to the application of AI in medicine. ¹⁶⁵ Besides, reward hacking means that machines find ways to achieve outcomes that circumvent rules or cheat the system. ¹⁶⁶

Conclusions

ML had achieved excellent performance and hold immense potential in spine. ML could help clinical staff to improve medical level, enhance work efficiency, and reduce adverse events. However more randomized controlled trials and improvement of interpretability are essential to clinicians accepting models’ assistance in real work.