Temporal Trends in Cervical Spine Curvature of South Korean Adults Assessed by Deep Learning System Segmentation, 2006-2018

This cross-sectional study uses a deep learning system to analyze the temporal trends in cervical spine curvature across sex and age groups in a population of South Korean adults.


Introduction
The internet and mobile devices are widely used in modern society, affecting daily life in the workplace and at home. In 2016, internet use on mobile devices outpaced that on desktop computers 1 ; thus, people more commonly access the internet while active than while stationary. The rapid increase in the use of mobile devices and computers has resulted in increased repetitive or prolonged postures, neck flexion, 2 and mechanical loading in the cervical spine. 3 Prolonged use of mobile devices is known to aggravate the degree and severity of neck and shoulder pain (NSP). 4,5 In 2010, NSP became the fourth leading cause of significant musculoskeletal disability worldwide, 6 with an annual prevalence rate exceeding 30%. 7 Corresponding to the increase in static postures in the modern sedentary lifestyle, the skeletal shape of humans has changed during the past decade. 8,9 The prolonged use of mobile devices leads to deviations from the neutral head postures and induces deeper neck flexion, 10,11 increasing concerns that use of mobile devices may have facilitated spinal misalignment and degeneration. Loss of cervical lordosis is a common disorder of the cervical spine that results in loss or reversal of the normal curvature. Previous cross-sectional studies have examined the specific associations between the degree of cervical lordosis and sex 12,13 and age. 14,15 However, studies have not reported trends over time because manual measurement of the cervical curvature is time-consuming and labor intensive.
To address these trends, we used artificial intelligence (AI) to automatically analyze large-scale radiographs to assess the prevalence of loss of lordotic cervical curvature. Since the introduction of the picture archiving and communication system during the past 2 decades, vast quantities of radiologic images stored in servers have allowed the field of radiology to benefit from big data analysis techniques. 16 Artificial intelligence-based deep learning has demonstrated promising diagnostic performance for various applications, including lesion detection in radiologic images, 17,18 classification in breast cancer, 19 and diabetic retinopathy. 20 This analysis for cervical curvature in South Korea represents the largest epidemiologic study, to our knowledge, to document cervical vertebral curvature associated with age and sex during 13 years. This study analyzes the temporal trends in cervical curvature across sex and age groups using an automated deep learning system (DLS).

Study Population
This cross-sectional study included lateral view radiographs of the cervical spine of adult patients  Table 1. The flowchart of the entire inclusion process is presented in Figure 1. The study was approved by the institutional review board of Severance Hospital, Yonsei University College of Medicine. The data were anonymized before analysis, and because of the retrospective nature of this study, the need for individual consent was waived. This study followed the Strengthening the Reporting of Observational Studies in

Development and Validation of the DLS
The DLS model for anterior vertical line segmentation on lateral radiographs was trained on the basis of U-Net architecture, a type of convolutional neural network that demonstrated a powerful performance in the segmentation of medical images. 22  For the quantitative validation of the proposed DLS, we performed a reliability test using pixelwise accuracy and mean intersection over the union metrics (a mean of the correct pixel classification divided by the total number of pixels of that class metrics). 23 In addition, the level of error was assessed using interoperator and intraoperator investigations on 390 reviewed radiographs (30 radiographs per year). To evaluate the consistency between the manually measured curvature and the DLS-predicted anterior vertebral line, we used the root-mean-squared error and the intraclass correlation coefficient.

Measurement of Cervical Spine Curvature on DLS Segmentation
The curvature (from convex to concave) of the segmented cervical vertebral line of the DLS model was evaluated using polynomial regression, following the methods described by Tecco and Festa 13 and Ferrario et al. 24 The x-axis was superimposed on the DLS segmented vertebral line (eFigure 2 in the Supplement). The midpoint of this line was set as origin. The y-axis was drawn perpendicular to the x-axis, intersecting the origin. The digitized coordinates of the anterior margin of the vertebral bodies are interpolated by a second-order polynomial function (y = a + bx + cx 2 ) to determine the convexity (or concavity) of the cervical curvature. The convexity (or concavity) of the cervical column is expressed as c × 10 3 ; the larger the absolute value, the more convex (or concave) the curve.
Conventionally, the curve is lordotic when the sign of coefficient c is positive, kyphotic when it is negative, and straight when the value is zero. Python, version 3.5.6 (Python Software Foundation) was used for digitizing the points, calculating the cartesian coordinates, and interpolating the curve. The kyphotic, straight, and lordotic groups of the cervical curves were determined by converting the obtained coefficient to existing measurement methods validated by previous studies. [25][26][27] The four-line Cobb method and the Harrison posterior tangent method are widely acknowledged reliable methods. 25 The Harrison posterior tangent method was used because of a lower SEM compared with that of the Cobb method. An expert radiologist (Y.H.L.) measured the angle between the posterior wall of the vertebral bodies of C2 and C7. On the basis of the global angle cutoffs determined in previous studies (lordosis: <−4°, straight: −4°to +4°, and kyphosis: >+4°) 26,28 and converting the angles to the proposed cervical curvature coefficient, the patients were classified into lordotic, straight, and kyphotic groups (eFigure 3 in the Supplement). The corresponding cervical coefficient cutoffs to represent the lordotic, straight, and kyphotic groups are as follows: straight: −0.8 to 0.8, lordotic: more than 0.8, and kyphotic: less than −0.8.

Statistical Analysis
The means and SDs of the cervical curve coefficient were calculated by sex and age for 2006 to 2018.
The prevalence estimates of kyphotic and straight curvature were each calculated for both sex and all age groups for 2006 to 2018. The significance of the trend in each kyphotic and straight cervical curvature group was tested using multinomial logistic regression across the study cycle to

JAMA Network Open | Imaging
Trends in Cervical Spine Curvature Assessed by Deep Learning System Segmentation estimate the regression coefficients (β) and 95% CIs for the changes. The independent variables included in the model were year, sex, and age.
In addition, cases repeatedly measured in the same individuals were incorporated to analyze the change of cervical curvature over time (per year) by sex and age groups. We applied a linear mixed-effects model to analyze the trend of individual cervical coefficients. The year, sex, and age were considered fixed effects, and the random slope on year was added in the model. No data were missing. The P value for each term tests the null hypothesis that the coefficient is equal to 0, and 2-sided P < .05 was considered statistically significant. All statistical analyses were performed using R, version 3.6.2 (R Foundation for Statistical Computing).

Study Population
Automatic cervical spine segmentation was identified from lateral radiographs of 13

DLS Accuracy
The segmentation performance of the DLS was evaluated on the test set.

Trend of Loss of Normal Cervical Curve
Overall, the mean (SD) of the cervical curvature decreased from 1.  Figure 2).   Error bars indicate 95% CIs.

Discussion
In this cross-sectional study, among a cohort of patients who underwent cervical spine radiography, after the UK (82%), the Netherlands (79%), Sweden (79%), Germany (79%), and the US (77%). 34 A systematic review 10 reported that the head flexion angle was larger when individuals were text messaging compared with when they were browsing the web and watching videos and that NSP is related to the duration of the mobile device use. 5 In addition, a previous study 35 that found that neck-related musculoskeletal conditions were more prevalent for handheld device users by 17.3% to 67% than any other region of the spine suggests practical implications for the mobile device-centric society.
This study suggests that the decline in mean cervical curvature from 2006 to 2018 was definite for men and women and more definite in women. The prevalence of lordotic curvature was generally higher in women compared with men, whereas existing literature suggests lower cervical lordosis in women. 36 However, the decrease was more pronounced in women, and the difference in the mean cervical coefficients between the 2 groups provided evidence of narrowing in the later years of the study. This rapid decrease for women might be associated with the increase in mobile-centric environments in the past decade and higher levels of smartphone addiction and use time for

JAMA Network Open | Imaging
Trends in Cervical Spine Curvature Assessed by Deep Learning System Segmentation women. 37,38 Increasing evidence supports NSP more frequently being reported among women, 5,39 as well as pain related to the use of electronic devices. 40 Regarding the age groups, previous cross-sectional studies 36,41 suggested a higher rate of loss in cervical lordosis in younger adults compared with those studies 12,13 that reported the lack of any significant association between age and lordotic cervical curvature. This study confirms that younger generations had the lowest mean cervical curvature during the study. From 2006 to 2018, the mean of the cervical curvature and trends of individuals' cervical curves for age groups of 18 to 29 and 30 to 39 years have shown the steepest decreases among all age groups. In 2017, internet use rates by age in South Korea were more than 99.7% for those 10 to 40 years of age, 98.7% for those in their 50s, 82.5% for those in their 60s, and 31.8% for those 70 years and older. 42 This finding could be associated with the progressive increase in the prevalence of abnormal cervical curves in younger age groups. The concerning prevalence of abnormal curvature in the 18-to 29-year age group in later years (60.0%-62.2% in 2017-2018) indicates that the loss of cervical lordosis is becoming more frequent, thereby providing much needed evidence to support the theory of "text neck" (stress injury and pain in the neck resulting from excessive use of handheld devices) as a worsening epidemic of the modern era. 39,43 Considering the increasing neck pain in younger people, the burden of abnormal cervical curvature might increase in the future, potentially reducing or reversing the progress achieved in decreasing chronic morbidity during the next several decades. This analysis suggests that the critical age for straight and kyphotic cervical curves was in the age groups younger than 40 years. Furthermore, the steep increases in the rates of abnormal curvature were observed for not only young adults but also middle-aged adults 40 to 69 years of age, especially with straight cervical curvature. Even if public health care policy makers are aware of these risks, in-depth evidence that fully analyzes population-based, long-term observations was not available. Public health information on work-related upper-limb disorders is prevalent 44,45 ; however, community-level regulatory interventions or social environments for facilitating activities to avoid dysfunctional movements of the cervical spine are largely absent. The active promotion of interventions, particularly government recommendations on mobile phone use 46 and modifiable risk factors not limited to children but also for young adults, is necessary to encourage awareness.

Strengths and Limitations
This study has strengths, including the use of a fully automated DLS pipeline to investigate a largescale population in South Korea. To the best of our knowledge, this is the largest extensive study in the published literature, spanning 13 years. The proposed DLS offers future advantages to assist routine clinical evaluation, 47 surgical interventions, and pain management. 48 In addition, consistency of measurement, near-instantaneous assessment, and cost-effectiveness of DLS will facilitate larger-scale, population-based epidemiologic studies and clinical trials.
This study has limitations. First, inferences related to the cause and effect of environmental and social factors to the cervical curvature cannot be made. Additional studies on physical, psychosocial, and personal risk factors, including surveys on time spent using the internet or devices, are required to assess the direct relationship with the cervical curvature. Second, this population-based study evaluated incidences from a South Korean cohort within a single tertiary center. Third, the data only provide information on patients who visited the hospital and underwent standing cervical radiography because of neck pain, upper arm pain, osteoporosis, and degenerative changes. Thus, this trend analysis is based on a clinical cohort and does not reflect community settings.