09_이준호_43권6호( ).hwp

Size: px

Start display at page:

Download "09_이준호_43권6호( ).hwp"

슬옹 맹
2 years ago
Views:

1 방사선기술과학제43권제6호 Journal of Radiological Science and Technology, 43(6), ISSN (On-line) ISSN (Print) < 원저 > 폐경여성에서트리기반머신러닝모델로부터골다공증예측 동남보건대학교방사선과 Predictive of Osteoporosis by Tree-based Machine Learning Model in Post-menopause Woman Lee In-Ja Lee Junho Department of Radiological Technology, Dongnam Health university Abstract In this study, the prevalence of osteoporosis was predicted based on 10 independent variables such as age, weight, and alcohol consumption and 4 tree-based machine-learning models, and the performance of each model was compared. Also the model with the highest performance was used to check the performance by clearing the independent variable, and Area Under Curve(ACU) was utilized to evaluate the performance of the model. The ACU for each model was Decision tree 0.663, Random forest 0.704, GBM 0.702, and XGBoost and the importance of the variable was shown in the order of age, weight, and family history. As a result of using XGBoost, the highest performance model and clearing independent variables, the ACU shows the best performance of with 7 independent variables. This data suggests that this method be applied to predict osteoporosis, but also other various diseases. In addition, it is expected to be used as basic data for big data research in the health care field. Key Words : Osteoporosis, Bone mineral density, Machine learning, XGBoost, Prediction model 중심단어 : 골다공증, 골밀도, 머신러닝, XGBoost, 예측모델 Ⅰ. 서론 최근고령화시대의진입으로만성질환인골다공증의환자수는꾸준히증가추세이다. 건강보험심사평가원보건의료빅데이터개방시스템에따르면골다공증연도별환자수추이는 2015 년 821,754 명에서 2019 년 1,079,548 명으로연평균 7.06% 증가하였으며, 연도별요양급여비용총액추이는 2015 년 92,892,132 천원에서 2019 년 165,662,889 천원으로연평균 15.56% 증가하였다. 골다공증은폐경여성에서흔히발생하는질환으로골양의감소및미세구조변화로인하여골의강도가감소되어골절이일어날가능성이높은상태를의미한다 [1]. 골다공증은골절이일어나기전까 지무증상인경우가많으며골다공증환자의경우작은손상에도골절이발생할가능성이높기때문에증상이없더라도조기진단및치료를수행해야병변으로인한사회적, 경제적부담을줄일수있다 [2]. 따라서골다공증예측및예방을위한연구는꾸준히진행되어왔다 [3,4]. 최근인공지능기술은당뇨병예측, 피부유형판단등의료분야에서다양하게적용되고있다 [5,6]. 인공지능기술의한분야인머신러닝 (Machine learning) 은데이터를기반으로패턴을학습하고결과를예측하는기법을통칭하며, 일반적으로지도학습 (Supervised learning), 비지도학습 (Un-supervised learning), 강화학습 (Reinforcement learning) 으로나눌수있다 [7]. 이중지도학습은훈련데 This work was supported by the research grant of the Dongnam Health University. Corresponding author: Junho Lee, Department of Radiological Technology, Dongnam Health University, 50, Cheoncheon-ro 74beon-gil, Jangan-gu, Suwon-si, Gyeonggi-do, 16328, Republic of Korea/ Tel: / twice@dongnam.ac.kr Received 27 November 2020; Revised 22 December 2020; Accepted 24 December 2020 Copyright c2020 by The Korean Journal of Radiological Science and Technology 방사선기술과학 2020 년제 43 권제 6 호 495

2 이터에정답을포함시켜학습시키는방법으로분류와회귀에사용된다. 트리기반모델은다양한지도학습에사용되고있으며, 대표적인트리기반알고리즘은결정트리 (Decision tree), 랜덤포레스트 (Random forest) 등이있고최근높은정확도로인하여각광받고있는 XGBoost(Extra gradient boost) 알고리즘도트리기반모델이다 [8,9]. 본연구에서는폐경여성에서트리기반머신러닝모델로부터골다공증예측에대해알아보고자하였다. Ⅱ. 대상및방법 예측하였다. 따라서선행연구및도메인지식을이용하여골다공증과연관성이높은변수를파악하였다 [10-14]. Table 1과같이간단히파악할수있는나이, 류마티스성관절염유병여부, 성인여성고위험음주빈도, 하루평균수면시간, 현재흡연여부, 1주일간근력운동일수, 신장, 체중, 허리둘레, 부모중골다공증진단 / 허리휨 / 경미한외상골절경험여부 ( 가족력등 ) 등 10개의독립변수를선정하였고 T-score 기준으로산출된골다공증유병여부를종속변수로선정하였으며골다공증유병여부에따라비골다공증군 (Non-osoteoporosis group) 과골다공증군 (Osoteoporosis group) 으로분류하였다. 1. 데이터수집국민건강영양조사 (The Korea national health and nutrition examination survey, KNHANES) 는국민건강증진법제16 조에근거하여국민의건강및영양상태를파악하기위해실시되고있으며, 1998 년제1기를시작으로 2018 년제7 기원시자료까지공개되어있다. 본연구에서는골다공증검사가실시된제5기 ( ) 원시자료를사용하였으며골다공증검사를진행한폐경여성 2,135 명의데이터를수집하였으며, 이중결측치를제외한 1,995 명을연구대상으로선정하였다. 2. 데이터선정및연구대상본연구는트리기반머신러닝모델을이용하여생활습관및현재신체지표를이용하여골다공증유병여부유무를 3. 데이터전처리및통계적분석수집된데이터는머신러닝모델에적용하기위해분석에적합한형태로변형해야한다. 이를데이터전처리라하며모델의성능을높이기위하여초기데이터는가공하는과정을거쳐야한다 [15]. 수집된데이터에서결측치가속한행및각변수내용에서모름으로응답한것과수면시간에서 0, 1, 20시간등비정상적이라생각되는행은삭제하였다. 나이, 신장, 체중등연속형변수는 StandardScaler 를사용하여데이터를처리하였다. 데이터의유의성검정을위하여통계적분석을실시하였다. 연속형변수는독립표본 t검정을이용하여평균과표준편차를구하였고, 범주형변수는카이제곱검정을이용하여빈도 (%) 를구하였다. p-value는 0.05 미만인경우에통계적으로유의하다고판단하였다. 통계분석은 SPSS 23.0 software (SPSS Inc., Chicago, IL, USA) 를이용하였다. 데이터분류는데이터를훈련데이터와 Table 1. Variable description Name Description Unit Type age age - Continuous DM3_lt Rheumatoid arthritis disease status - Categorical BD2_32 High-risk drinking frequency - Categorical BP8 Average Sleep Time per Day Hour Continuous BS3_1 Current smoking status - Categorical BE5_1 Day of muscle exercise for one week - Categorical HE_ht Height cm Continuous HE_wt Weight kg Continuous HE_wc Waist circumference cm Continuous DX_Q_hsty DX_OST Survey: Diagnosis of osteoporosis among parents / Waist bending / Minor trauma fracture experience Osteoporosis disease status : T-score criterion of the total femur, femoral neck, lumbar spine - Categorical - Categorical 496 Journal of Radiological Science and Technology 43(6), 2020

3 폐경여성에서트리기반머신러닝모델로부터골다공증예측 테스트데이터로랜덤하게분리하여교차검증을실시하는홀드아웃 (Hold-out method) 방법을사용하였다. 4. 분류알고리즘및평가골다공증유병여부를예측하기위하여트리기반머신러닝분류알고리즘인결정트리 (Decision tree), 랜덤포레스트 (Randon forest), GBM(Gradient boost machine), XGBoost (Extra gradient boost) 모델을사용하였다. 머신러닝분류알고리즘의평가는 Table 2와같이혼동행렬 (Confusion matrix) 을기반으로구성되며, 예측클래스와실제클래스를기준으로 Negative, Positive로분류하여 True positive(tp), False positive(fp), False negative(fn), True negative(tn) 형태로구성된다. 이를이용하여 ROC(Receiver operation characteristic) 곡선과 AUC(Area under curve), 정확도 (Accuracy), 정밀도 (Precision), 재현율 (Recall), F1 스코어 (F1_score) 를산출하여모델을평가할수있다. 본연구에서는이진분류모델의성능평가와비교를위하여의학진 단, 생물학, 신용평가등의분야에서예측및분류모델의성능평가지표로주로사용되는 AUC 를이용하여모델의성능을평가하였다 [16,17]. AUC 는 ROC 곡선의아래면적을의미하고 1에가까울수록좋은모형이라고판단한다. 평가결과가장높은성능의알고리즘에서독립변수를줄여가며예측결과를파악하여모델의최적화를수행하였다. 분석도구는 Python(v3.7.3), Jupyter notebook 을사용하였다. Ⅲ. 결과 1. 데이터유의성평가데이터유의성평가결과 Table 3과같이나이, 고위험음주빈도, 1주일간근력운동일수, 신장, 체중, 허리둘레, 가족력등은통계적으로유의하게나타났으며 (p<0.05), 류마티스성관절염유병여부, 하루평균수면시간, 현재흡연여부는유의하게차이가나타나지는않았다 (p>0.05). Table 2. Confusion matrix of classification model Actual class Positive (Osteoporosis) Negative (Non-osteoporosis) Positive (Osteoporosis) True positive(tp) (Osteoporosis prediction success) False positive(fp) (Non-Osteoporosis prediction failure) Predicted class Negative (Non-osteoporosis) False negative(fn) (Osteoporosis prediction failure) True negative(tn) (Non-osteoporosis prediction success) Table 3. Baseline characteristics of study population Variables Osoteoporosis group(n=685) Non-osoteoporosis group(n=1,310) p-value age ± ±8.264 < DM3_lt No 643(93.869%) 1245(95.038%) Yes 42(6.131%) 65(4.962%) BD2_32 < None 190(27.737%) 393(30.000%) Less than once a month 47(6.861%) 155(11.832%) Once a month 14(2.044%) 80(6.107%) Once a week 9(1.314%) 55(4.198%) Almost everyday 4(0.584%) 15(1.145%) Not applicable 421(61.460%) 612(46.718%) BP ± ± BS3_ Yes 34(4.964%) 43(3.282%) 방사선기술과학 2020 년제 43 권제 6 호 497

4 Table 3. Baseline characteristics of study population (Cont.) Variables Osoteoporosis group(n=685) Non-osoteoporosis group(n=1,310) Sometimes 2(0.292%) 5(0.382%) p-value Not currently 32(4.672%) 40(3.053%) Not applicable 617(90.073%) 1222(93.282%) BE5_1 < None 627(91.533%) 1073(81.908%) 1 day 10(1.460%) 48(3.664%) 2 days 9(1.314%) 53(4.046%) 3 days 14(2.044%) 51(3.893%) 4 days 7(1.022%) 29(2.214%) More than 5 days 18(2.628%) 56(4.275%) HE_ht ± ±5.306 < HE_wt ± ±8.326 < HE_wc ± ±9.257 < DX_Q_hsty < Yes 90(13.139%) 254(19.389%) No 595(86.861%) 1056(80.611%) * age: age, DM3_lt: Rheumatoid arthritis disease status, BD2_32: High-risk drinking frequency, BP8: Average Sleep Time per Day, BS3_1: Current smoking status, BE5_1: Day of muscle exercise for one week, HE_ht: Height, HE_wt: Weight, HE_wc: Waist circumference, DX_Q_hsty: Survey: Diagnosis of osteoporosis among parents / Waist bending / Minor trauma fracture experience * Values are presented as number (%) or mean±standard deviation. 2. 데이터셋분류폐경여성의골다공증유병여부를예측하기위하여 1,995 명의원천데이터를사용하였다. 이중비골다공증그룹은 1,310 명이며, 골다공증그룹은 685 명이다. 머신러닝분류모델에사용하기위하여원천데이터를홀드아웃방법을이용하여훈련데이터셋과테스트데이터셋으로랜덤하게나누었으며 7:3 의비율로나누어각각 1,396명, 599명의데이터를사용하였다. 3. 예측모델평가골다공증유병여부를예측하기위하여훈련데이터를활용하여 Decision tree, Random forest, GBM, XGBoost 분 류모델을구축하였다. 생성한모델에테스트데이터를적용한결과는 Table 4와같이혼동행렬로나타난다. 예측력평가결과각모델의 AUC 는 Fig. 1과같이 Decision tree 0.663, Random forest 0.704, GBM 0.702, XGBoost 0.710으로나타났다. Table 5는생성된모델들에대한이진분류성능평가를나타낸다. 4. 변수중요도평가및최적화각모델의 10개독립변수가골다공증예측에영향을미친정도는 Fig. 2와같다. 결정트리의경우나이 0.304, 체중 0.231, 신장 0.170, 허리둘레 0.136, 하루평균수면시간 0.068, 고위험음주빈도 0.035, 1주일간근력운동일수 0.022, 가족력등 0.020, 류마티스성관절염유병여부 Table 4. Confusion matrix of test data-set True Predicted Osteoporosis Decision tree Random forest Gradient boost Extra gradient boost Non-osteopo rosis Osteoporosis Non-osteopo rosis Osteoporosis Non-osteopo rosis Osteoporosis Non-osteopo rosis Osteoporosis Non-osteoporosis Journal of Radiological Science and Technology 43(6), 2020

폐경여성에서트리기반머신러닝모델로부터골다공증예측 Table 5. Result of prediction model evaluation Model Area under curve Accuracy Precision Recall F1 score Decision tree 0.663 0.705 0.782 0.780 0.781 Random forest 0.704 0.

704) Gradient boost(area under curve = 0.702) Extra gradient boost(area under curve = 0.710) Fig. 1. Receiver operation characteristic curve by classification model 0.007, 현재흡연여부 0.

061, 가족력등 0.018, 1주일간근력운동일수 0.017, 하루평균수면시간 0.013, 현재흡연여부 0.008, 고위험음주빈도 0.007, 류마티스성관절염유병여부 0.002 로나타났으며 XGBoost 는나이 0.370, 체중 0.156, 가족력등 0.098, 1주일간근력운동일수 0.084, 신장 0.078, 허리둘레 0.060, 하루평균수면시간 0.

5 폐경여성에서트리기반머신러닝모델로부터골다공증예측 Table 5. Result of prediction model evaluation Model Area under curve Accuracy Precision Recall F1 score Decision tree Random forest Gradient boost Extra gradient boost Decision tree(area under curve = 0.663) Random forest(area under curve = 0.704) Gradient boost(area under curve = 0.702) Extra gradient boost(area under curve = 0.710) Fig. 1. Receiver operation characteristic curve by classification model 0.007, 현재흡연여부 이었으며랜덤포레스트는나이 0.268, 신장 0.203, 체중 0.202, 허리둘레 0.156, 하루평균수면시간 0.071, 고위험음주빈도 0.036, 가족력등 0.021, 1주일간근력운동일수 0.021, 현재흡연여부 0.012, 류마티스성관절염유병여부 0.010으로나타났다. GBM 은나이 0.500, 체중 0.265, 신장 0.110, 허리둘레 0.061, 가족력등 0.018, 1주일간근력운동일수 0.017, 하루평균수면시간 0.013, 현재흡연여부 0.008, 고위험음주빈도 0.007, 류마티스성관절염유병여부 로나타났으며 XGBoost 는나이 0.370, 체중 0.156, 가족력등 0.098, 1주일간근력운동일수 0.084, 신장 0.078, 허리둘레 0.060, 하루평균수면시간 0.056, 현재흡연여부 0.049, 고위험음주빈도 0.048, 류마티스성관절염유병여부 으로나타났다. 모델의성능이가장좋게평가된 XGBoost 모델에서 10개독립변수를하나씩축소하여평가한결과 Fig. 3과같이 7 개의독립변수를사용하였을때 AUC 가가장높았으며이때 AUC는 0.750, 사용변수는나이, 체중, 가족력등, 1주일간근력운동일수, 신장, 허리둘레, 하루평균수면시간이다. 방사선기술과학 2020 년제 43 권제 6 호 499

Area under curve according to feature reflection Ⅳ.

6 Decision tree Random forest Gradient boost Fig. 2. Result of feature importance evaluation Extra gradient boost Fig. 3. Area under curve according to feature reflection Ⅳ. 고찰 골다공증은골형성의감소및골파괴의증가로인하여골량과밀도가낮아져일상적인충격에도골절위험이높아지는질환이다 [18,19]. 또한, 수년간자각증상없이진행되 므로건강검진중에발견되는경우가많아조기검진으로인한예방이필요하며이를위해국민건강보험공단일반검진을통하여만 54세여성과만 66세여성에한하여골밀도검사를진행하고있다 [20]. 골다공증은나이의증가및생활습관영향에의해골소실속도가다르고증상이없기 500 Journal of Radiological Science and Technology 43(6), 2020

7 폐경여성에서트리기반머신러닝모델로부터골다공증예측 때문에더욱주의깊게관찰할필요가있다. 특히여성의경우골밀도의감소는폐경을전후로가장중요한시점을맞이하게되며폐경이후에스트로겐결핍에의한골흡수의증가와장에서칼슘의흡수가감소되어골밀도감소가진행된다 [21]. 따라서폐경이후골밀도감소의예측및골다공증예방에관한연구가꾸준히진행되어왔다 [22,23]. 머신러닝기술은다양한분야의분류및예측에활용되고있으며, 질병예측으로범위를넓혀가고있다 [24-26]. 본연구는트리기반머신러닝모델 4종의골다공증예측성능을비교하였다. 트리기반모델의기본이되는결정트리는결과에이르기위해예, 아니오를반복하며학습을이어간다. 랜덤포레스트는다수의결정트리를이용하여예측을진행하며, GBM 과 XGBoost 는이전트리의오차를보완하는방식으로학습하여예측에활용되며 XGB 는랜덤포레스트와 GBM보다학습속도가빠르며과적합이잘일어나지않고다른알고리즘과연계활용성이좋은장점이있다 [27]. 트리기반모델학습결과 4개모델전부나이가골다공증유병여부를예측하는데영향력이가장큰변수였으며, 모델별 AUC 는 Decision tree 0.663, Random forest 0.704, GBM 0.702, XGBoost 로 Decision tree 를제외하고나머지는비슷하였다. 가장좋은성능을나타낸 XGBoost 를이용하여변수를순서대로 5개까지줄여학습시킨결과 7개의변수를사용했을때 AUC 는 으로가장좋은성능을나타냈다. 머신러닝을이용한질병예측을다양한질병에적용한다면질병의사전예측및예방에기여할수있을것으로생각된다. 본연구에서는 10개의독립변수를사용하여예측하였다. 머신러닝은빅데이터를활용하여유용한정보의추출및예측이가능한만큼더많은변수들을사용한다면모델성능향상및일정수준의예측성능을유지할수있는최소변수를찾는데유용하게사용될수있을것이다. 또한, 본연구에서는기본모델을사용하여모델간의성능을비교하였으나머신러닝모델은사용자가직접세팅가능한하이퍼파라미터 (Hyper parameter) 조절을통하여성능을향상시킬수있다. 하이퍼파라미터는정해진값이없기때문에데이터의종류와회귀, 분류등예측결과의종류에따라적절히조정해야하며, 경험법칙 (Rule of thumb) 또는 GridSearchCV 등을이용하면모델의성능을향상시킬수있을것으로생각된다. Ⅴ. 결론 본연구에서는트리기반의머신러닝분류알고리즘인결정 트리 (Decision tree), 랜덤포레스트 (Randon forest), GBM (Gradient boost machine), XGBoost(Extra gradient boost) 모델을사용하여폐경이후여성의골다공증유병여부를예측하였다. 신장, 체중, 가족력등 10개의독립변수를이용하여예측을진행하였으며이용한변수중골다공증유병여부를예측하는데영향력가장큰변수는나이로나타났다. 또한, 예측결과 XGBoost 의 AUC 가 0.710으로가장높았으며변수를줄여가며평가한결과 7개변수를사용하였을때 AUC는 으로모델의성능이가장높았다. 이러한연구의결과는다양한질병의예측에적용될수있을것으로생각되며, 향후보건의료분야의빅데이터연구에기초자료로활용될것으로기대된다. REFERENCES [1] Park H, So J. The validational study of OSTA (Osteoporosis Self Assessment Tool for Asian) for prediction of osteoporosis in korean post- and perimenopausal women. Obstetrics & Gynecology Science. 2003;46(2): [2] Kang B, Kwon S, Kim D, Kim E, Kim I, Kim J, et al. Manual of bone densitometry. Seoul: Cheongwoon; [3] Yoo J, Lee B. Prediction model of osteoporosis using nutritional components based on association. The Journal of Convergence on Culture Technology. 2020;6(3): [4] Lee K, Yoon C, Lee J. Comparison of body weight and body mass index as predictors for osteoporosis among postmenopausal Korean women. Journal of the Korean Academy of Family Medicine. 2005; 26(10): [5] Kim K. A study on model of skin type judgment tool using machine learning technique. A Treatise on The Plastic Media. 2018;21(4): [6] Jang J, Lee M, Lee T. Development of T2DM prediction model using RNN. Journal of Digital Convergence. 2019;17(8): [7] Kwon C. Python machine learning complete guide. Gyeonggi: Wikibooks; [8] Géron A. Hands-on machine learning with scikit-learn & tensorflow. Seoul: Hanbit Media Inc; [9] Han J, Ko D, Choe H. Predicting and analyzing fac- 방사선기술과학 2020 년제 43 권제 6 호 501

8 tors affecting financial stress of household using machine learning: Application of XGBoost. Journal of Consumer Studies. 2019;30(2): [10] Lee S, Jang S, Jung D, Lee J. Reconsideration of the mechanical loading hypothesis: Is obesity protective against osteoporosis? Journal of the Korean Official Statistics. 2014;19(2):1-29. [11] Baek K, Kang M. Official positions of the international society for clinical densitometry. Endocrinology and Metabolism. 2005;20(1):1-7. [12] Kim D. New guidelines for the diagnosis and fracture risk assessment of osteoporosis. Journal of Bone Metabolism. 2008;15(1):1-8. [13] Kim K. Factors associated with the bone mineral density in Korean adults: Data from the Korean National Health and Nutrition Examination Survey (KNHANES) V. Journal of Agricultural Medicine & Community Health. 2014;39(4): [14] Choi J, Han S, Shin A, Shin C, Park S, Cho S, et al. Prevalence and risk factors of osteoporosis and osteopenia in Korean women: Cross-sectional study. Journal of Menopausal Medicine. 2008;14(1): [15] Lee H, Lee S, Cho J, Cho I. Analysis of feature importance of ship s berthing velocity using classification algorithms of machine learning. Journal of the Korean Society of Marine Environment & Safety. 2020;26(2): [16] Lee B. Prediction model of hypercholesterolemia using body fat mass based on machine learning. The Journal of Convergence on Culture Technology. 2019;5(4): [17] Hong C, Won C. Parameter estimation for the imbalanced credit scoring data using AUC maximization. The Korean Journal of Applied Statistics. 2016; 29(2): [18] Park J, Choi M, Lee S, Choi Y, Park Y. The association between bone mineral density, bone turnover markers, and nutrient intake in pre- and postmenopausal women. Journal of Nutrition and Health. 2011;44(1): [19] Kim J, Yang Y, Lee M. The influence of osteoporosis knowledge, health belief and self efficacy on osteoporosis prevention behavior osteoporosis in middle-aged men. Korean Public Health Research. 2020;46(2): [20] Lee H, Rho J. Study on the osteoporosis knowledge, concern about osteoporosis factors, and health behavior to prevent osteoporosis of women in Jeonbuk area. Journal of Nutrition and Health. 2018;51(6); [21] Kim T, Lee H, Chung S, Park H. Differentiation in the management of osteoporosis between premenopausal and menopausal women. Journal of Menopausal Medicine. 2011;17(1):21-6. [22] Song T, Choi H, Lee S, Yeon M, Ko J, Lee C, et al. Performance of risk indices for prediction of osteoporosis in post- and perimenopausal women. Obstetrics & Gynecology Science. 2005;48(11): [23] Lee J, Kim E, Suk M, Kim E, Hwang L. Factors influencing osteoporosis. Korean Academy of Community Health Nursing. 2003;14(2): [24] Choi P, Min I. A predictive model for the employment of college graduates using a machine learning approach. Journal of Vocational Education & Training. 2018;21(1): [25] Lee G, Lee J. A classification of medical and advertising blogs using machine learning. Journal of Korea Academia-Industrial cooperation Society. 2018;19(11): [26] Eom J, Lee S, Kim B. A feasibility study on the improvement of diagnostic accuracy for energy-selective digital mammography using machine learning. Journal of Radiological Science and Technology. 2019;42(1):9-17. [27] Kim I, Lee K. Tree based ensemble model for developing and evaluating automated valuation models: The case of Seoul residential apartment. Journal of the Korean Data And Information Science Society. 2020;31(2): 구분 성명 소속 직위 제1 저자 이인자 동남보건대학교 부교수 교신저자 이준호 동남보건대학교 조교수 502 Journal of Radiological Science and Technology 43(6), 2020

서론 34 2

서론 34 2 34 2 Journal of the Korean Society of Health Information and Health Statistics Volume 34, Number 2, 2009, pp. 165 176 165 진은희 A Study on Health related Action Rates of Dietary Guidelines and Pattern of