• ๋Œ€ํ•œ์ „๊ธฐํ•™ํšŒ
Mobile QR Code QR CODE : The Transactions of the Korean Institute of Electrical Engineers
  • COPE
  • kcse
  • ํ•œ๊ตญ๊ณผํ•™๊ธฐ์ˆ ๋‹จ์ฒด์ด์—ฐํ•ฉํšŒ
  • ํ•œ๊ตญํ•™์ˆ ์ง€์ธ์šฉ์ƒ‰์ธ
  • Scopus
  • crossref
  • orcid

  1. (Dept. of Biomedical Engineering Chonnam National University, South Korea)



Classifier, Machine Learning, Pain Assessment, Photoplethysmogram

1. ์„œ ๋ก 

์ˆ˜์ˆ  ํ›„ ํ†ต์ฆ์˜ ์ ์ ˆํ•œ ์กฐ์ ˆ์ด ํ™˜์ž ์˜ˆํ›„ ํ–ฅ์ƒ์— ํ•„์ˆ˜์ ์ด๋ผ๋Š” ๊ฒƒ์€ ๋„๋ฆฌ ์•Œ๋ ค์ง„ ์‚ฌ์‹ค์ด๋ฉฐ ์ฃผ๋กœ ์ง„ํ†ต์ œ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ†ต์ฆ์„ ์™„ํ™”ํ•œ๋‹ค. ์ด ๋•Œ ์ง„ํ†ต์ œ ํˆฌ์—ฌ๋Š” ์ •ํ•ด์ง„ ์‹œ๊ฐ„๋งˆ๋‹ค ์ฃผ๊ธฐ์ ์œผ๋กœ ์ด๋ฃจ์–ด์ง€๋Š” ๊ฒƒ์ด ์ผ๋ฐ˜์ ์ด๋‚˜ ํ†ต์ฆ์ž๊ฐ€์กฐ์ ˆ(Patient Controlled Analgesia, PCA) ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•˜๊ธฐ๋„ ํ•œ๋‹ค. ํ†ต์ฆ์ž๊ฐ€์กฐ์ ˆ์€ ํ†ต์ฆ์ด ์žˆ์„ ๋•Œ๋งˆ๋‹ค ์ •๋งฅ ๋˜๋Š” ๊ฒฝ๋ง‰ ์™ธ๊ฐ•์— ์„ค์น˜๋œ ํ†ต์ฆ์ž๊ฐ€์กฐ์ ˆ์žฅ์น˜๋ฅผ ํ†ตํ•ด ํ™˜์ž๊ฐ€ ์Šค์Šค๋กœ ์ง„ํ†ต์ œ๋ฅผ ํˆฌ์—ฌํ•˜๋Š” ๋ฐฉ๋ฒ•์ด๋ฉฐ, ์ฃผ๊ด€์  ๊ฒฝํ—˜์ธ ํ†ต์ฆ์„ ํ™˜์ž๊ฐ€ ์Šค์Šค๋กœ ๊ด€๋ฆฌํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ์žฅ์ ์ด ์žˆ๋‹ค [1]. ์ด์ƒ์˜ ๋ฐฉ๋ฒ•๋“ค์€ ํ†ต์ฆ ๊ด€๋ฆฌ์— ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๊ณ  ์žˆ์œผ๋‚˜ ๊ฐœ์ธ์˜ ํ†ต์ฆ ๋ฏผ๊ฐ๋„, ๊ณ ํ†ต ์ธ๋‚ด๋ ฅ ๋“ฑ์˜ ์ฐจ์ด๋กœ ์ธํ•ด ์ง„ํ†ต์ œ์˜ ๊ณผ๋‹ค, ๊ณผ์†Œ ํˆฌ์—ฌ๊ฐ€ ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ๊ฐœ์ธ์˜ ํ†ต์ฆ ํŠน์„ฑ์„ ์™„๋ฒฝํžˆ ๋ฐ˜์˜ํ•˜๋Š” ํ†ต์ฆ ์กฐ์ ˆ ๊ธฐ์ˆ ์€ ์—ฌ์ „ํžˆ ๊ทธ ์š”๊ตฌ๊ฐ€ ํฌ๋‹ค. ํ†ต์ฆ ์กฐ์ ˆ์˜ ์ฒซ ๋ฒˆ์งธ ๋‹จ๊ณ„๋Š” ํ†ต์ฆ ์œ ๋ฌด ๋ฐ ์ •๋„๋ฅผ ํŒŒ์•…ํ•˜๋Š”๋ฐ ์žˆ๋‹ค. ์ด์— ๋”ฐ๋ผ ๊ฒฝํ—˜์ ์œผ๋กœ ํ‰๊ฐ€๋˜๋˜ ํ†ต์ฆ ํ‰๊ฐ€์˜ ํ•œ๊ณ„๋ฅผ ๊ทน๋ณตํ•˜๊ธฐ ์œ„ํ•œ ์ƒ์ฒด์‹ ํ˜ธ ๊ธฐ๋ฐ˜ ์ •๋Ÿ‰์  ํ†ต์ฆ ํ‰๊ฐ€ ๊ธฐ์ˆ ์ด ์ง€์†์ ์œผ๋กœ ์ œ์•ˆ๋˜๊ณ  ์žˆ๋‹ค. ์ด ๊ธฐ์ˆ ๋“ค์€ ์ฃผ๋กœ ํ†ต์ฆ์— ์˜ํ•œ ์ž์œจ์‹ ๊ฒฝ๊ณ„ ๋ฐ˜์‘์„ ์ธก์ •ํ•˜๋Š”๋ฐ ๊ธฐ๋ฐ˜์„ ๋‘๊ณ  ์žˆ์œผ๋ฉฐ ์‹ฌ๋ฐ•๋ณ€์ด๋„, ๋™๊ณต ํฌ๊ธฐ ๋ณ€ํ™”, ํ”ผ๋ถ€์ „๊ธฐ๋ฐ˜์‘, ๊ด‘์šฉ์ ๋งฅํŒŒ(Photoplethysmogram, PPG) ํŒŒํ˜• ๋ณ€ํ™” ๋“ฑ์„ ํ†ตํ•ด ํ†ต์ฆ์˜ ๋ฐœ์ƒ ๋ฐ ์ •๋„๋ฅผ ์ถ”์ •ํ•œ๋‹ค [2ยญ5]. ์ด ์ค‘ ๊ด‘์šฉ์ ๋งฅํŒŒ๋Š” ์ž„์ƒ ํ™˜๊ฒฝ์—์„œ ์‚ฌ์šฉ๋„๊ฐ€ ๋†’๊ณ , ์‹ฌ๋ฐ• ์ˆ˜ ๋ณ€ํ™”, ํ˜ˆ๊ด€ ์ˆ˜์ถ• ๋ฐ ํŒฝ์ฐฝ ๋“ฑ์˜ ์ž์œจ์‹ ๊ฒฝ ํ™œ์„ฑ ์ •๋ณด๋ฅผ ๋ฐ˜์˜ํ•˜๊ณ  ์žˆ์œผ๋ฏ€๋กœ ํ†ต์ฆํ‰๊ฐ€์— ์‚ฌ์šฉ๋„๊ฐ€ ๋†’์•„ SPI(Surgical Pain Index, GE Healthcare, inc., Chicago, IL, USA) ๋“ฑ์˜ ํ†ต์ฆ ํ‰๊ฐ€ ์ง€ํ‘œ๋กœ ๊ฐœ๋ฐœ๋˜์–ด ์‚ฌ์šฉ๋˜๊ณ  ์žˆ๋‹ค [5]. ์ด์ƒ์˜ ๋ฐฉ๋ฒ• ์ค‘ ๋Œ€๋ถ€๋ถ„์€ ์‹ ํ˜ธ๋กœ๋ถ€ํ„ฐ ํ†ต์ฆ์„ ๋ฐ˜์˜ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ์•Œ๋ ค์ง„ ํŠน์ง• ๊ฐ’๋“ค์„ ์‚ฐ์ถœํ•œ ํ›„ ์ด์˜ ํšŒ๊ท€์‹์„ ๊ตฌํ•˜์—ฌ ํ†ต์ฆ์„ ์ˆ˜์น˜ํ™”ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•˜๊ณ  ์žˆ์œผ๋ฉฐ, ์ตœ๊ทผ์—๋Š” ๊ธฐ๊ณ„ํ•™์Šต์„ ์ ์šฉํ•˜์—ฌ ํ†ต์ฆ์„ ํ‰๊ฐ€ํ•˜๋ ค๋Š” ์—ฐ๊ตฌ ๋˜ํ•œ ์ˆ˜ํ–‰๋˜์–ด ์ง€๊ณ  ์žˆ๋‹ค [6,7]. ์ด ์—ฐ๊ตฌ์—์„œ๋Š” ๊ด‘์šฉ์ ๋งฅํŒŒ๋ฅผ ์‚ฌ์šฉํ•œ ํ†ต์ฆํ‰๊ฐ€์— ์žˆ์–ด ๋‹ค์–‘ํ•œ ๊ธฐ๊ณ„ํ•™์Šต ๋ถ„๋ฅ˜๊ธฐ๋“ค์˜ ํ†ต์ฆ ๋ถ„๋ฅ˜ ์„ฑ๋Šฅ์„ ๋น„๊ตํ•ด๋ณด๋Š” ๊ฒƒ์„ ๋ชฉ์ ์œผ๋กœ ํ•˜๋ฉฐ ์‹ค์ œ ์ˆ˜์ˆ  ์ „, ํ›„ ์ธก์ •๋œ ๊ด‘์šฉ์ ๋งฅํŒŒ์— ๋กœ์ง€์Šคํ‹ฑ ํšŒ๊ท€(Logistic Regression, LR), ๋žœ๋คํฌ๋ ˆ์ŠคํŠธ(Random Forest, RF), ๋‹ค์ธต ํผ์…‰ํŠธ๋ก (MultiยญLayer Perceptron, MLP), ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง(Convolutional Neural Network, CNN)์„ ์ ์šฉํ•˜์—ฌ ํ†ต์ฆ์„ ๋ถ„๋ฅ˜ํ•˜๊ณ  ๋ถ„๋ฅ˜๊ธฐ๋ณ„ ํ†ต์ฆ ๋ถ„๋ฅ˜ ์„ฑ๋Šฅ์„ ๋น„๊ต ํ‰๊ฐ€ํ•œ๋‹ค.

2. ๊ธฐ๊ณ„ํ•™์Šต ๋ถ„๋ฅ˜๊ธฐ ์„ค๊ณ„

2.1 ๋กœ์ง€์Šคํ‹ฑ ํšŒ๊ท€

๋กœ์ง€์Šคํ‹ฑ ํšŒ๊ท€๋ถ„์„์€ ๋Œ€์ƒ๋“ค์ด ๋‘ ์ง‘๋‹จ ํ˜น์€ ๊ทธ ์ด์ƒ์˜ ์ง‘๋‹จ์œผ๋กœ ๋‚˜๋ˆ„์–ด์ง„ ๊ฒฝ์šฐ์— ๊ฐœ๋ณ„ ๊ด€์ธก ๊ฐ’๋“ค์ด ์–ด๋Š ์ง‘๋‹จ์œผ๋กœ ๋ถ„๋ฅ˜๋  ์ˆ˜ ์žˆ๋Š”๊ฐ€๋ฅผ ๋ถ„์„ํ•˜๊ณ  ์ด๋ฅผ ์˜ˆ์ธกํ•˜๋Š” ๋ชจ๋ธ์„ ๋ชจํ˜•ํ™”ํ•œ ๊ฒƒ์ด๋‹ค. ๋กœ์ง€์Šคํ‹ฑ ํšŒ๊ท€ ๋ถ„์„์€ ํ™•๋ฅ  ๋ชจ๋ธ๋กœ์„œ ์ข…์†๋ณ€์ˆ˜์™€ ๋…๋ฆฝ๋ณ€์ˆ˜ ๊ฐ„์˜ ์ธ๊ณผ๊ด€๊ณ„๋ฅผ ์‹œ๊ทธ๋ชจ์ด๋“œ(sigmoid) ํ•จ์ˆ˜๋ฅผ ์ด์šฉํ•˜์—ฌ ์ถ”์ •ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ์ž…๋ ฅ ๋ณ€์ˆ˜ x์— ๋Œ€ํ•˜์—ฌ ์‹(1)๊ณผ ๊ฐ™์€ ์ˆ˜์‹์œผ๋กœ ํ‘œํ˜„๋œ๋‹ค.

(1)
$sigmoid(x)=\dfrac{1}{1-e^{-x}}$

2.2 ๋žœ๋คํฌ๋ ˆ์ŠคํŠธ

์ฃผ๋กœ ๋ถ„๋ฅ˜ ๋˜๋Š” ํšŒ๊ท€ ๋ถ„์„์— ์‚ฌ์šฉ๋˜๋Š” ๋žœ๋คํฌ๋ ˆ์ŠคํŠธ ๋ถ„๋ฅ˜๊ธฐ๋Š” ๊ทธ๋ฆผ 1๊ณผ ๊ฐ™์ด ํ•™์Šต๋œ ๋‹ค์ˆ˜์˜ ์˜์‚ฌ ๊ฒฐ์ • ํŠธ๋ฆฌ(decision tree)๋ฅผ ๊ฒฐํ•ฉํ•˜์—ฌ ํ•œ ๊ฐœ์˜ ์ตœ์ข…์˜ˆ์ธก๋ชจํ˜•์„ ์ƒ์„ฑํ•˜๋Š” ์•™์ƒ๋ธ”(ensemble) ๋ถ„๋ฅ˜ ๊ธฐ๋ฒ•์ด๋‹ค [8]. ํ•™์Šต๊ณผ์ •์—์„œ ๋žœ๋คํฌ๋ ˆ์ŠคํŠธ๋Š” ๊ฐ ํŠธ๋ฆฌ์˜ ๋…ธ๋“œ(node)๋งˆ๋‹ค ์ตœ์ ์˜ ํŒ๋ณ„์‹๊ณผ ์ž„๊ณ„๊ฐ’์„ ๊ฒฐ์ •ํ•œ๋‹ค. ๋žœ๋คํฌ๋ ˆ์ŠคํŠธ๋Š” ์„œ๋กœ ๋‹ค๋ฅธ ํŠน์„ฑ์„ ๊ฐ€์ง€๋Š” ์—ฌ๋Ÿฌ ๊ฐœ์˜ ์˜์‚ฌ๊ฒฐ์ • ํŠธ๋ฆฌ๋ฅผ ๊ตฌ์„ฑํ•จ์œผ๋กœ์จ ๊ฐ ํŠธ๋ฆฌ ๊ฐ„์˜ ์ƒ๊ด€๊ด€๊ณ„๋ฅผ ์ตœ์†Œํ™”ํ•˜์—ฌ ๊ณผ์ ํ•ฉ(overfitting) ํ˜„์ƒ์„ ์ค„์ผ ์ˆ˜ ์žˆ๊ณ , ๊ฒฐ๊ณผ์ ์œผ๋กœ ํŠธ๋ฆฌ๋“ค์˜ ๊ฒฐ๊ณผ ํ•ฉ์‚ฐ์„ ํ†ตํ•ด ์‹ ๋ขฐ๋„ ๋†’๊ณ  ์•ˆ์ •์ ์ธ ๋ถ„๋ฅ˜ ๊ฒฐ๊ณผ๋ฅผ ๋„์ถœํ•  ์ˆ˜ ์žˆ๋Š” ์žฅ์ ์ด ์žˆ๋‹ค. ๋˜ํ•œ ๋งŽ์€ ๋…๋ฆฝ๋ณ€์ˆ˜๋ฅผ ๋ชจ๋‘ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ๋Œ€๋Ÿ‰์˜ ๋ฐ์ดํ„ฐ์—์„œ ์ค‘์š”ํ•œ ๋ณ€์ˆ˜๋ฅผ ์ฐพ์•„๋‚ด๊ณ  ์ถ”์‚ฐ ๋ชจ๋ธ์„ ๊ตฌ์„ฑํ•˜๋Š”๋ฐ ์šฉ์ดํ•˜๋‹ค. ๋žœ๋คํฌ๋ ˆ์ŠคํŠธ ๋ถ„๋ฅ˜๊ธฐ๋ฅผ ๊ตฌํ˜„ํ•  ๋•Œ ์„ค์ •ํ•ด์•ผ ํ•˜๋Š” ์ฃผ์š” ๋งค๊ฐœ ๋ณ€์ˆ˜๋กœ๋Š” ํŠธ๋ฆฌ์˜ ๊ฐœ์ˆ˜์™€ ํ•˜๋‚˜์˜ ํŠธ๋ฆฌ์—์„œ ๋ช‡ ๊ฐœ์˜ ๋…ธ๋“œ๋กœ ์ˆ˜ํ–‰ํ•  ๊ฒƒ์ธ์ง€ ๊ฒฐ์ •ํ•˜๋Š” ์ตœ๋Œ€ ํ—ˆ์šฉ ๊นŠ์ด๊ฐ€ ์žˆ๋‹ค. ๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ์ตœ๋Œ€ ํ—ˆ์šฉ ๊นŠ์ด(max depth)๋ฅผ 5๋กœ ์„ค์ •ํ•˜๊ณ , ํŠธ๋ฆฌ์˜ ๊ฐœ์ˆ˜๋ฅผ 10์œผ๋กœ ์„ค์ •ํ•˜์—ฌ ์—ฌ๋Ÿฌ ๊ฐœ์˜ ๊ฒฐ์ • ํŠธ๋ฆฌ๋ฅผ ํ•™์Šตํ•˜๋„๋ก ํ–ˆ๋‹ค.

๊ทธ๋ฆผ. 1. ๋žœ๋คํฌ๋ ˆ์ŠคํŠธ์˜ ๊ตฌ์กฐ

Fig. 1. Structure of Random Forest

../../Resources/kiee/KIEE.2019.68.12.1626/fig1.png

2.3 ๋‹ค์ธต ํผ์…‰ํŠธ๋ก 

์ธ๊ณต์‹ ๊ฒฝ๋ง(Artificial Neural Network, ANN)์€ ์ธ๊ฐ„์˜ ๋‘๋‡Œ์˜ ๋Œ€๋Ÿ‰ ๋ณ‘๋ ฌ์„ฑ์„ ๋ชจ๋ฐฉํ•œ ๊ฒƒ์œผ๋กœ ๋ฐ˜๋ณต์ ์ธ ํ•™์Šต์„ ํ†ตํ•ด ๋ฐ์ดํ„ฐ์— ์‚ฌ์ด์˜ ํŒจํ„ด์„ ์ฐพ์•„๋‚ด๋Š” ์•Œ๊ณ ๋ฆฌ์ฆ˜์ด๋‹ค. ์ž…์ถœ๋ ฅ์ธต๋งŒ์œผ๋กœ ๊ตฌ์„ฑ๋œ ๋‹จ์ธต ๊ตฌ์กฐ ์ธ๊ณต์‹ ๊ฒฝ๋ง์ธ ๊ฒฝ์šฐ์—๋Š” ์„ ํ˜• ๋ถ„๋ฆฌ๊ฐ€ ๊ฐ€๋Šฅํ•œ ๋ฌธ์ œ์—๋งŒ ์ ์šฉ ๊ฐ€๋Šฅํ•˜๋‹ค๋Š” ํ•œ๊ณ„๊ฐ€ ์žˆ๋Š”๋ฐ ์ด๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ๊ณ ์•ˆ๋œ ์‹ ๊ฒฝ๋ง์œผ๋กœ ๋‹ค์ธต ํผ์…‰ํŠธ๋ก ์ด ์žˆ๋‹ค. ๋‹ค์ธต ํผ์…‰ํŠธ๋ก ์€ ์ž…๋ ฅ์ธต(input layer)๊ณผ ์ถœ๋ ฅ์ธต(output layer) ์‚ฌ์ด์— ๋‹ค์ˆ˜์˜ ์€๋‹‰์ธต(hidden layer)์ด ์กด์žฌํ•˜๋Š” ์‹ ๊ฒฝ๋ง์œผ๋กœ ์€๋‹‰์ธต๊ณผ ๊ฐ ๋‰ด๋Ÿฐ ์‚ฌ์ด์˜ ์ž…์ถœ๋ ฅ ํŠน์„ฑ์„ ๋น„์„ ํ˜•์œผ๋กœ ์ฒ˜๋ฆฌํ•จ์œผ๋กœ์จ ๋ณต์žกํ•œ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•  ์ˆ˜ ์žˆ๋‹ค. ์ด ์—ฐ๊ตฌ์—์„œ๋Š” ์ž…์ถœ๋ ฅ์ธต๊ณผ 2๊ฐœ์˜ ์€๋‹‰์ธต์œผ๋กœ ๊ตฌ์„ฑ๋œ ๋‹ค์ธต ํผ์…‰ํŠธ๋ก ์„ ์„ค๊ณ„ํ•˜์˜€๋‹ค (๊ทธ๋ฆผ 2). ์€๋‹‰์ธต์€ 2048๊ฐœ ๋…ธ๋“œ๋ฅผ ๊ฐ€์ง€๋Š” 2๊ฐœ ์ธต์œผ๋กœ ๊ตฌ์„ฑํ•˜์˜€์œผ๋ฉฐ, ์ถœ๋ ฅ์ธต์€ ์›ยญํ•ซ ์ธ์ฝ”๋”ฉ(oneยญhot encoding)์— ๋Œ€์‘๋˜๋„๋ก 2๊ฐœ๋กœ ์„ค์ •ํ•˜์˜€๋‹ค. ์ถœ๋ ฅ์ธต์„ ์ œ์™ธํ•œ ๋‚˜๋จธ์ง€ ์ธต์—์„œ๋Š” ํ™œ์„ฑํ™” ํ•จ์ˆ˜๋กœ ReLU(Rectified Linear Unit)๋ฅผ ์‚ฌ์šฉํ•˜์˜€๊ณ , ์ถœ๋ ฅ์ธต์—์„œ๋Š” ์†Œํ”„ํŠธ๋งฅ์Šค(softmax) ํ•จ์ˆ˜๋ฅผ ์‚ฌ์šฉํ•˜์˜€๋‹ค. ๋ถ„๋ฅ˜๊ธฐ์˜ ํ•™์Šต ํšจ์œจ์„ฑ์„ ๋†’์ด๊ธฐ ์œ„ํ•ด epoch์„ 10์œผ๋กœ ์„ค์ •ํ•˜์˜€๊ณ  ๋™์‹œ์— ํ•œ๋ฒˆ ํ•™์Šต์„ ์ˆ˜ํ–‰ํ•  ๋•Œ์˜ ๋ฐ์ดํ„ฐ ์ˆ˜์ธ ๋ฐฐ์น˜ ํฌ๊ธฐ(batch size)๋Š” 20์œผ๋กœ ์„ค์ •ํ•˜์˜€์œผ๋ฉฐ ๋ฐ์ดํ„ฐ์˜ ๊ณผ์ ํ•ฉ ๋ฐฉ์ง€๋ฅผ ์œ„ํ•ด ํƒˆ๋ฝ์œจ(dropout rate)์„ 0.5๋กœ ์„ค์ •ํ–ˆ๋‹ค.

๊ทธ๋ฆผ. 2. ์—ฐ๊ตฌ์— ์‚ฌ์šฉ๋œ ๋‹ค์ธต ํผ์…‰ํŠธ๋ก  ๊ตฌ์กฐ

Fig. 2. MultiยญLayer Perceptron structure used in this study

../../Resources/kiee/KIEE.2019.68.12.1626/fig2.png

2.4 ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง

ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง์€ ์‹œ๊ฐ ์ฒ˜๋ฆฌ ๊ณผ์ •์„ ๋ชจ๋ฐฉํ•œ ๊ฒƒ์œผ๋กœ ์ฃผ๋กœ ์ด๋ฏธ์ง€๋‚˜ ์˜์ƒ ๋“ฑ์˜ ๊ณต๊ฐ„์  ๋ฐ์ดํ„ฐ์˜ ํŒจํ„ด ์‹๋ณ„์— ํšจ๊ณผ์ ์ธ ์‹ ๊ฒฝ๋ง์ด๋‹ค. ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง์€ ์ด๋ฏธ์ง€์˜ ํŠน์ง•์„ ์ถ”์ถœํ•˜๋Š” ๋ถ€๋ถ„๊ณผ ์ด๋ฏธ์ง€๋ฅผ ๋ถ„๋ฅ˜ํ•˜๋Š” ๋ถ€๋ถ„์œผ๋กœ ๊ตฌ์„ฑ๋˜๋ฉฐ ๋ฐ์ดํ„ฐ์˜ ๊ณต๊ฐ„ ์ •๋ณด๋ฅผ ์œ ์ง€ํ•˜๋ฉด์„œ ์ธ์ ‘ ์ด๋ฏธ์ง€์™€์˜ ํŠน์ง•์„ ํšจ๊ณผ์ ์œผ๋กœ ์ธ์‹ํ•˜๊ณ  ๊ฐ•์กฐํ•˜๋Š” ๋ฐฉ์‹์œผ๋กœ ํ•™์Šต๋œ๋‹ค. ๋˜ํ•œ ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง ์ค‘ ํ•˜๋‚˜์ธ ์ผ์ฐจ์› ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง์€ ์ผ์ฐจ์› ๊ฒฉ์ž ํ˜•ํƒœ ๋ฐ์ดํ„ฐ์˜ ํŒจํ„ด ์ธ์‹์— ์ข‹์€ ์„ฑ๋Šฅ์„ ๋ณด์ด๊ณ  ์žˆ์–ด ์‹ ํ˜ธ ๋ฐ์ดํ„ฐ๋ฅผ ๋ถ„์„ํ•˜๋Š”๋ฐ ๊ฐ€์žฅ ํ™œ๋ฐœํžˆ ์‚ฌ์šฉ๋˜๊ณ  ์žˆ๋‹ค. ์ด ์—ฐ๊ตฌ์—์„œ๋Š” ํ†ต์ฆ ์œ ๋ฌด๋ฅผ ๋ถ„๋ฅ˜ํ•˜๊ธฐ ์œ„ํ•ด 212๊ฐœ์˜ ํ†ต์ฆํ›„๋ณด์ง€ํ‘œ๋ฅผ ์ผ์ฐจ์› ์ž…๋ ฅ์œผ๋กœ ํ•˜๋Š” 1์ฐจ์› ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง์„ ์„ค๊ณ„ํ•˜์˜€๋‹ค. ํ•ฉ์„ฑ๊ณฑ ์ธต(convolution layer) ํ•„ํ„ฐ์˜ ํฌ๊ธฐ๋Š” 1ร—2๋กœ ์„ค์ •ํ•˜์˜€์œผ๋ฉฐ, ์ŠคํŠธ๋ผ์ด๋“œ(stride)๋Š” 1๋กœ ์„ค์ •ํ•˜์—ฌ ์—ฌ๋Ÿฌ ๋ฒˆ์˜ ํ•ฉ์„ฑ๊ณฑ ์—ฐ์‚ฐ์„ ํ†ตํ•ด ์ธ์ ‘ํ•œ ๋ฐ์ดํ„ฐ๋“ค์˜ ํŠน์ง•์„ ์ถ”์ถœํ–ˆ๋‹ค. ์™„์ „ ์—ฐ๊ฒฐ์ธต(fully connected layer)์—์„œ๋Š” ์•ž์—์„œ ์ถ”์ถœํ•œ ํŠน์ง•๋“ค์„ 512๊ฐœ์˜ ๋…ธ๋“œ๋ฅผ ๊ฐ€์ง€๋Š” ์€๋‹‰์ธต์„ ๊ฑฐ์ณ ReLU ํ™œ์„ฑํ™” ํ•จ์ˆ˜๋ฅผ ํ†ตํ•ด ์ตœ์ข…์ ์œผ๋กœ 1๊ฐœ์˜ ๋…ธ๋“œ๋กœ ์ถœ๋ ฅ๋˜์–ด 0.5๋ฅผ ๊ธฐ์ค€์œผ๋กœ 0๊ณผ 1๋กœ ์ด์ง„ ๋ถ„๋ฅ˜ํ•˜์˜€๋‹ค. ๋‹ค์ธต ํผ์…‰ํŠธ๋ก ๊ณผ ๋™์ผํ•˜๊ฒŒ ํ•™์Šต epoch์€ 10, ๋ฐฐ์น˜ ํฌ๊ธฐ๋Š” 20์œผ๋กœ ์„ค์ •ํ•˜์—ฌ ํ•™์Šต๋˜์—ˆ๋‹ค. ์‚ฌ์šฉ๋œ ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง์˜ ๊ตฌ์กฐ๋Š” ๊ทธ๋ฆผ 3๊ณผ ๊ฐ™๋‹ค.

๊ทธ๋ฆผ. 3. ์—ฐ๊ตฌ์— ์‚ฌ์šฉ๋œ ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง ๊ตฌ์กฐ

Fig. 3. Convolutional Neural Network structure used in this study

../../Resources/kiee/KIEE.2019.68.12.1626/fig3.png

3. ๋ถ„๋ฅ˜ ์„ฑ๋Šฅ ํ‰๊ฐ€

3.1 ํ†ต์ฆ ๋ฐ์ดํ„ฐ

๋ณธ ์—ฐ๊ตฌ๋Š” ์„œ์šธ์•„์‚ฐ๋ณ‘์› ์ž„์ƒ์—ฐ๊ตฌ์‹ฌ์˜์œ„์›ํšŒ(Institutional Review Board, IRB No.: 2016ยญ0477)์˜ ์Šน์ธ์„ ๋ฐ›์•˜์œผ๋ฉฐ ๊ตญ์ œ์ž„์ƒ์‹œํ—˜๋“ฑ๋กํ”Œ๋žซํผ์— ๋“ฑ๋ก๋˜์—ˆ๋‹ค (http://cris.nih.go.kr, KCT00 02080). ์ž„์ƒ์‹œํ—˜์—๋Š” ์ด 81๋ช…์˜ ํ”ผํ—˜์ž๊ฐ€ ์ฐธ์—ฌํ•˜์˜€์œผ๋ฉฐ ์ตœ์ข…์ ์œผ๋กœ 73๋ช…์˜ ๋ฐ์ดํ„ฐ๊ฐ€ ์‚ฌ์šฉ๋˜์—ˆ๋‹ค. ๋ฐ์ดํ„ฐ๋Š” ์ˆ˜์ˆ  ์ „๊ณผ ์ˆ˜์ˆ  ํ›„ ๊ฐ๊ฐ ์ธก์ •๋˜์—ˆ์œผ๋ฉฐ ํ†ต์ฆ ํ‰๊ฐ€ ๋„๊ตฌ์ธ VAS(visual analogue scale)๋ฅผ ๊ธฐ๋กํ•œ ํ›„, ๊ด‘์šฉ์ ๋งฅํŒŒ์™€ SPI๋ฅผ ๋™์‹œ์— ๊ธฐ๋กํ•˜์˜€๋‹ค. VAS๋Š” ํ†ต์ฆ์˜ ๊ฐ•๋„์— ๋”ฐ๋ผ 0ยญ100 ์‚ฌ์ด์˜ ๊ฐ’์œผ๋กœ ์ธก์ •๋˜๋Š”๋ฐ, ์ˆ˜์ˆ  ์ „ ๋ชจ๋“  ํ”ผํ—˜์ž์˜ VAS๋Š” 0์ ์œผ๋กœ ํ†ต์ฆ์ด ์—†๋Š” ์ƒํƒœ์˜€๊ณ , ์ˆ˜์ˆ  ํ›„ VAS๋Š” 67.6ยฑ1.0์ ์œผ๋กœ ๋ชจ๋“  ํ”ผํ—˜์ž๊ฐ€ 60 ์  ์ด์ƒ์˜ VAS๋ฅผ ๋ณด์˜€๋‹ค. ์ด์— ๋”ฐ๋ผ ๋ถ„๋ฅ˜๊ธฐ ์ž…๋ ฅ์„ ์œ„ํ•œ ํ†ต์ฆ ๋ ˆ์ด๋ธ”์„ ์ˆ˜์ˆ  ์ „ ๋ฐ์ดํ„ฐ๋Š” 0, ์ˆ˜์ˆ  ํ›„ ๋ฐ์ดํ„ฐ๋Š” 1๋กœ ์„ค์ •ํ•˜์˜€๋‹ค. ๊ด‘์šฉ์ ๋งฅํŒŒ๋Š” ์ˆ˜์ˆ  ์ง์ „ 6๋ถ„๊ฐ„ ํ†ต์ฆ์ด ์—†๋Š” ์ƒํƒœ์™€ ์ˆ˜์ˆ  ์งํ›„ 6๋ถ„๊ฐ„ ์ง„ํ†ต์ œ ํˆฌ์—ฌ ์ „ ์ƒํƒœ์—์„œ ๊ธฐ๋ก๋˜์—ˆ๋‹ค. ๊ด‘์šฉ์ ๋งฅํŒŒ๋Š” 300Hz์˜ ํ‘œ๋ณธํ™” ์ฃผํŒŒ์ˆ˜๋กœ ๊ธฐ๋ก๋˜์—ˆ๊ณ , SPI๋Š” 10์ดˆ๋งˆ๋‹ค ์ถœ๋ ฅ๋˜๋Š” ๊ฐ’์„ ๊ธฐ๋กํ•˜์˜€๋‹ค.

ํ†ต์ฆ ํ‰๊ฐ€ ๋ถ„๋ฅ˜์— ์‚ฌ์šฉ๋œ ์ž…๋ ฅ ๋ฐ์ดํ„ฐ๋กœ๋Š” ์ด์ „ ์—ฐ๊ตฌ์—์„œ ๋„์ถœ๋œ ํ†ต์ฆ ํ›„๋ณด ์ง€ํ‘œ๋“ค์„ ์‚ฌ์šฉํ•˜์˜€๋‹ค [9]. ์„ ํ–‰ ์—ฐ๊ตฌ์—์„œ๋Š” ์ง„ํญ๊ณผ ๊ฐ™์€ ๊ณต๊ฐ„์  ํŠน์„ฑ ๋ฐ ๋ฐ•๋™ ๊ฐ„๊ฒฉ, ํŒŒํ˜• ๊ตฌ๊ฐ„ ๋“ฑ์˜ ์‹œ๊ฐ„์  ํŠน์„ฑ ๋“ฑ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ด‘์šฉ์ ๋งฅํŒŒ ํ˜•ํƒœ ๋ถ„์„์„ ํ†ตํ•ด 23๊ฐœ์˜ ๊ธฐ๋ณธ ์ง€ํ‘œ๋ฅผ ๋„์ถœํ•˜์˜€๊ณ , ๊ฐœ์ธ ๊ฐ„์˜ ์ฐจ์ด๋ฅผ ๋ณด์ •ํ•˜๊ธฐ ์œ„ํ•œ ์ •๊ทœํ™”๋ฅผ ํ†ตํ•ด 28๊ฐœ์˜ ์ •๊ทœํ™” ์ง€ํ‘œ๋ฅผ ์ถ”๊ฐ€ ์ถ”์ถœํ•˜์—ฌ ํ•œ ๋ฐ•๋™์—์„œ ์ด 53๊ฐœ์˜ ์ง€ํ‘œ๋ฅผ ์ƒ์„ฑํ•˜์˜€๋‹ค. ์ด ํ›„ ์‹œ๊ณ„์—ด ๋ถ„์„์— ์‚ฌ์šฉ๋  ์ˆ˜ ์žˆ๋Š” ๋ถ„์„๋ฒ•์ธ ํ‰๊ท (Average, AV), ํ‘œ์ค€ํŽธ์ฐจ(Standard Deviation, SD), ์ธ์ ‘ํ•œ ์ง€ํ‘œ ๊ฐ’ ์ฐจ์ด์˜ ํ‘œ์ค€ํŽธ์ฐจ(Standard Deviation of the Successive Deviation, SDSD), ์ธ์ ‘ํ•œ ์ง€ํ‘œ ๊ฐ’ ์ฐจ์ด์˜ ํ‰๊ท ์ œ๊ณฑ๊ทผ(Root Mean Square of Successive Difference, RMSSD)๋“ฑ์„ 53๊ฐœ ์ง€ํ‘œ์— ์ ์šฉํ•˜์˜€๊ณ  ์ตœ์ข…์ ์œผ๋กœ 212๊ฐœ์˜ ์ง€ํ‘œ๋ฅผ ๋„์ถœํ•˜์˜€๋‹ค. ์ด ์—ฐ๊ตฌ์—์„œ๋Š” 212๊ฐœ์˜ ์ง€ํ‘œ์— ๋Œ€ํ•˜์—ฌ ๊ฐ ์ง€ํ‘œ๋ณ„๋กœ zยญscore ์ •๊ทœํ™” ๊ณผ์ •์„ ์ˆ˜ํ–‰ํ•˜๊ณ  ์ž…๋ ฅ ๋ฐ์ดํ„ฐ๋กœ ์‚ฌ์šฉํ•˜์˜€๋‹ค.

3.2 ์ค‘์ฒฉ ๊ต์ฐจ ๊ฒ€์ฆ

๊ฐœ๋ฐœ๋œ ํ†ต์ฆ ๋ถ„๋ฅ˜๊ธฐ ๋ชจ๋ธ ์„ฑ๋Šฅ์˜ ํ†ต๊ณ„์  ์‹ ๋ขฐ์„ฑ์„ ํ™•๋ณดํ•˜๊ธฐ ์œ„ํ•ด ์ค‘์ฒฉ ๊ต์ฐจ ๊ฒ€์ฆ(Nested cross validation)์„ ์‚ฌ์šฉํ•˜์˜€๋‹ค. ์ค‘์ฒฉ ๊ต์ฐจ ๊ฒ€์ฆ์€ k๊ฒน ๊ต์ฐจ ๊ฒ€์ฆ(kยญfold cross validation)๋ฐฉ๋ฒ•์„ ์™ธ๋ถ€ ๋ฃจํ”„(outer loop)์™€ ๋‚ด๋ถ€ ๋ฃจํ”„(inner loop) ๊ฐ๊ฐ ์ ์šฉํ•˜์—ฌ ๊ฒ€์ฆ์„ ์ˆ˜ํ–‰ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ์†Œ๊ทœ๋ชจ์˜ ๋ฐ์ดํ„ฐ๋กœ ๊ตฌํ˜„ํ•œ ๋ชจ๋ธ ๊ฐ„ ๋น„๊ต์— ์šฉ์ดํ•˜๋‹ค [10].

๋ณธ ์—ฐ๊ตฌ์—์„œ ์ˆ˜ํ–‰๋œ ์ค‘์ฒฉ ๊ต์ฐจ ๊ฒ€์ฆ์€ ์™ธ๋ถ€ ๋ฃจํ”„์™€ ๋‚ด๋ถ€ ๋ฃจํ”„ ๋ชจ๋‘ 5๊ฐœ์˜ ์ง‘๋‹จ์œผ๋กœ ๋‚˜๋ˆ„์–ด ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ–ˆ๋‹ค. ๋จผ์ € ์™ธ๋ถ€ ๋ฃจํ”„๋ฅผ ๊ตฌ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ์ „์ฒด ๋ฐ์ดํ„ฐ๋ฅผ 5๊ฐœ์˜ ์ง‘๋‹จ์œผ๋กœ ๋‚˜๋ˆ„์–ด ํ•œ ๊ฐœ์˜ ์ง‘๋‹จ์€ ์‹œํ—˜ ์ง‘๋‹จ(test set)์œผ๋กœ, ๋‚˜๋จธ์ง€ ์ง‘๋‹จ์€ ๋‚ด๋ถ€ ๋ฃจํ”„๋ฅผ ๊ตฌ์„ฑํ•˜๋Š” ๊ฐœ๋ฐœ ์ง‘๋‹จ(development set)์œผ๋กœ ๊ตฌ๋ถ„ํ•˜์—ฌ 5๊ฐœ์˜ ์ง‘๋‹จ์ด ํ•œ ๋ฒˆ์”ฉ์€ ์‹œํ—˜ ์ง‘๋‹จ์œผ๋กœ ์‚ฌ์šฉ๋  ์ˆ˜ ์žˆ๋„๋ก ๋ฐ˜๋ณต ์ˆ˜ํ–‰ํ•˜์—ฌ ์ด 5๋ฒˆ ๋ฐ˜๋ณต์— ์˜ํ•œ ํ‰๊ท  ๋ถ„๋ฅ˜ ์„ฑ๋Šฅ์„ ๊ณ„์‚ฐํ•œ๋‹ค. ์ด์™€ ๋น„์Šทํ•˜๊ฒŒ ๋‚ด๋ถ€ ๋ฃจํ”„์˜ ์ง‘๋‹จ ๊ตฌ์„ฑ์€ ์‹œํ—˜ ๋ฐ์ดํ„ฐ ์ง‘๋‹จ์œผ๋กœ ๊ตฌ์„ฑ๋œ ๋ฐ์ดํ„ฐ๋ฅผ ์ œ์™ธํ•œ ๋‚˜๋จธ์ง€ 4๊ฐœ์˜ ๋ฐ์ดํ„ฐ ์ง‘๋‹จ์„ ๋‹ค์‹œ 5๊ฐœ์˜ ์ง‘๋‹จ์œผ๋กœ ์žฌ๊ตฌ๋ถ„ํ•˜์—ฌ ํ•œ ๊ฐœ์˜ ์ง‘๋‹จ์€ ๊ฒ€์ฆ(validation) ๋ฐ์ดํ„ฐ ๋‚˜๋จธ์ง€ ์ง‘๋‹จ์€ ํ•™์Šต(training) ๋ฐ์ดํ„ฐ๋กœ ์ค‘์ฒฉ ๋ถ„๋ฐฐํ•˜์—ฌ ํ•™์Šตยญ๊ฒ€์ฆ์„ ์ˆ˜ํ–‰ํ•œ๋‹ค (๊ทธ๋ฆผ 4). ๊ฒฐ๊ณผ ์ œ๊ณต์— ์žˆ์–ด, ์™ธ๋ถ€ ๋ฃจํ”„๋ฅผ ํ†ตํ•ด ๊ณ„์‚ฐ๋œ ์‹œํ—˜ ๋ฐ์ดํ„ฐ ํ‰๊ท  ๋ถ„๋ฅ˜ ์„ฑ๋Šฅ์„ ๋ถ„๋ฅ˜๊ธฐ์˜ ์ตœ์ข… ์„ฑ๋Šฅ ๊ฒฐ๊ณผ๋กœ ์‚ฌ์šฉ๋˜์—ˆ๊ณ , ๋‚ด๋ถ€ ๋ฃจํ”„๋ฅผ ํ†ตํ•ด ๊ณ„์‚ฐ๋œ ๊ฒ€์ฆ ๋ฐ์ดํ„ฐ ํ‰๊ท  ๋ถ„๋ฅ˜ ์„ฑ๋Šฅ์€ ๊ฐœ๋ฐœ(development) ์„ฑ๋Šฅ ๊ฒฐ๊ณผ๋กœ ์‚ฌ์šฉ๋˜์—ˆ๋‹ค. ์ตœ์ข… ๋ถ„๋ฅ˜๊ธฐ ์„ฑ๋Šฅ์€ ์ž„์˜ ์ถ”์ถœ(random sampling)์— ์˜ํ•œ ์ค‘์ฒฉ ๊ต์ฐจ ๊ฒ€์ฆ์„ 30ํšŒ ์ˆ˜ํ–‰ํ•œ ํ›„ ํ‰๊ท ์„ ๋‚ด์–ด ๊ณ„์‚ฐ๋˜์—ˆ๋‹ค.

๊ทธ๋ฆผ. 4. ์ค‘์ฒฉ ๊ต์ฐจ ๊ฒ€์ฆ (5๊ฒน)

Fig. 4. Nested cross validation (5ยญfold)

../../Resources/kiee/KIEE.2019.68.12.1626/fig4.png

3.3 ํ†ต์ฆ ๋ถ„๋ฅ˜๊ธฐ ์„ฑ๋Šฅ ํ‰๊ฐ€ ์ง€ํ‘œ

ํ†ต์ฆ ๋ถ„๋ฅ˜๊ธฐ์˜ ์„ฑ๋Šฅ ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด TP(True Positive), TN(True Negative), FP(False Positive), FN(False Negative)๋ฅผ ๋„์ถœํ•œ ํ›„ ์ด๋ฅผ ์ด์šฉํ•ด ์ •ํ™•๋„(Accuracy, AC), ๋ฏผ๊ฐ๋„(Sensitivity, SE), ํŠน์ด๋„(Specificity, SP), ์–‘์„ฑ์˜ˆ์ธก๋„(Positive Predictive Value, PPV)๋ฅผ ๊ตฌํ•ด ํ‰๊ฐ€ ์ง€ํ‘œ๋กœ ์‚ฌ์šฉํ–ˆ๋‹ค. ๊ฐ ์ง€ํ‘œ ๊ณ„์‚ฐ ๋ฐฉ๋ฒ•์€ ์‹(2)ยญ(5)์™€ ๊ฐ™๋‹ค.

(2)
$AC =\dfrac{TP +TN}{TP +FP+TN+FN}$

(3)
$SE =\dfrac{TP}{TP +FN}$

(4)
$SP =\dfrac{TN}{TN +FN}$

(5)
$PPV =\dfrac{TP}{TP +FP}$

4. ๊ฒฐ ๊ณผ

ํ‘œ 1์€ ๋ถ„๋ฅ˜ ๊ฒฐ๊ณผ๋ฅผ ๋ณด์—ฌ์ค€๋‹ค. ๊ฒฐ๊ณผ์—์„œ ์‹œํ—˜ ์ง‘ํ•ฉ์˜ ์ •ํ™•๋„๋Š” ๋กœ์ง€์Šคํ‹ฑ ํšŒ๊ท€, ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง, ๋‹ค์ธต ํผ์…‰ํŠธ๋ก , ๋žœ๋คํฌ๋ ˆ์ŠคํŠธ์ˆœ์„œ๋กœ ๋‚˜ํƒ€๋‚ฌ๋‹ค. ๋กœ์ง€์Šคํ‹ฑ ํšŒ๊ท€์˜ ๊ฒฝ์šฐ๋Š” ๋ชจ๋“  ์„ฑ๋Šฅ ์ง€ํ‘œ์—์„œ ์ „๋ฐ˜์ ์œผ๋กœ ๋†’์€ ์„ฑ๋Šฅ์„ ๋ณด์˜€์œผ๋‚˜ ๋ฏผ๊ฐ๋„์— ์žˆ์–ด์„œ๋Š” ์ƒ๋Œ€์ ์œผ๋กœ ๋‚ฎ์€ ๊ฒฐ๊ณผ๋ฅผ ๋ณด์˜€๋‹ค.

ํ‘œ 1. ๋ถ„๋ฅ˜๊ธฐ๋ณ„ ํ†ต์ฆ ๋ถ„๋ฅ˜ ์„ฑ๋Šฅ ๋ฐ ๋น„๊ต

Table 1. Comparison of the pain assessment performance of classifiers

Classifier

Metrics

Value

Development set

Test

set

Test set / Development

set (%)

LR

AC

0.855

0.835

97.6%

SE

0.786

0.771

98.0%

SP

0.924

0.900

97.3%

PPV

0.912

0.885

97.0%

AUC

0.856

0.828

96.7%

RF

AC

0.805

0.766

95.2%

SE

0.802

0.762

94.9%

SP

0.807

0.770

95.4%

PPV

0.807

0.768

95.3%

AUC

0.812

0.762

93.9%

MLP

AC

0.859

0.771

89.7%

SE

0.844

0.779

92.3%

SP

0.874

0.762

87.2%

PPV

0.870

0.766

88.1%

AUC

0.944

0.776

82.2%

CNN

AC

0.920

0.807

87.7%

SE

0.908

0.801

88.2%

SP

0.932

0.813

87.3%

PPV

0.930

0.811

87.2%

AUC

0.995

0.799

80.3%

LR: Logistic Regression, RF: Random Forest, MLP: MultiยญLayer Perceptron, CNN: Convolutional Neural Network, AC: Accuracy, SE: Sensitivity, SP: Specificity, PPV: Positive Predictive Value, AUC: Area under curve

๊ณผ์ ํ•ฉ ์—ฌ๋ถ€๋ฅผ ํŒ๋‹จํ•˜๊ธฐ ์œ„ํ•˜์—ฌ ๊ฐœ๋ฐœ ์ง‘ํ•ฉ๊ณผ ์‹œํ—˜ ์ง‘ํ•ฉ ๊ฐ„ ์„ฑ๋Šฅ ๊ฐ์†Œ ๋น„์œจ์„ ๋น„๊ตํ•œ ๊ฒฐ๊ณผ ๋กœ์ง€์Šคํ‹ฑ ํšŒ๊ท€, ๋žœ๋คํฌ๋ ˆ์ŠคํŠธ, ๋‹ค์ธต ํผ์…‰ํŠธ๋ก , ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง ์ˆœ์œผ๋กœ ๊ณผ์ ํ•ฉ์— ๊ฐ•์ธํ•œ ๊ฒฐ๊ณผ๋ฅผ ๋ณด์ด๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ๋‹ค. ๊ทธ๋ฆผ 5๋Š” ๊ฐ ๋ถ„๋ฅ˜๊ธฐ์˜ ROC(receiver operating characteristic)์™€ ๊ณก์„ ํ•˜๋ฉด์ (area under curve, AUC)๋ฅผ ๋„์‹œํ•œ ๊ฒƒ์ด๋‹ค. ์ด ๊ทธ๋ฆผ์—์„œ๋Š” ์ „์ฒด์ ์œผ๋กœ ๊ฐœ๋ฐœ ์ง‘ํ•ฉ์— ๋น„ํ•ด ์‹œํ—˜ ์ง‘ํ•ฉ์—์„œ AUC๊ฐ€ ๊ฐ์†Œํ•˜๋Š” ๊ฒƒ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ๋‹ค. ๋˜ํ•œ, ๊ฐœ๋ฐœ ์ง‘ํ•ฉ๊ณผ ์‹œํ—˜ ์ง‘ํ•ฉ ๊ฐ„ ๊ณผ์ ํ•ฉ์€ ๋‹ค์ธต ํผ์…‰ํŠธ๋ก , ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง์— ์„œ ํ›จ์”ฌ ๋” ํฌ๊ฒŒ ๋‚˜ํƒ€๋‚˜๋Š” ๊ฒƒ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ๋‹ค.

๊ทธ๋ฆผ. 5. ์—ฌ๋Ÿฌ ๋ถ„๋ฅ˜๊ธฐ์˜ Receiver Operating Curve

Fig. 5. Receiver Operating Curve of various classifiers (LR: Logistic Regression, RF: Random Forest, MLP: Multi-Layer Perceptron, CNN: Convolutional Neural Network, DE: Development set, TE: Test set)

../../Resources/kiee/KIEE.2019.68.12.1626/fig5.png

์ด์ƒ์˜ ๊ฒฐ๊ณผ๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์˜๋ฏธ๋กœ ํ•ด์„๋  ์ˆ˜ ์žˆ๋‹ค. ๋จผ์ € ์ด ์—ฐ๊ตฌ์˜ ๊ฒฐ๊ณผ๋Š” ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง์ด๋‚˜, ๋‹ค์ธต ํผ์…‰ํŠธ๋ก ์ด ๋‹ค๋ฅธ ๋ชจ๋ธ์— ๋น„ํ•ด ๊ฐœ๋ฐœ ์ง‘ํ•ฉ์—์„œ๋Š” ์œ ์‚ฌํ•˜๊ฑฐ๋‚˜ ๋†’์€ ์„ฑ๋Šฅ์„ ๋ณด์ด์ง€๋งŒ ๊ณผ์ ํ•ฉ๋˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์Œ์„ ๋ณด์—ฌ์ฃผ๊ณ  ์žˆ๋‹ค. ์ด๋Š” ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง์ด๋‚˜ ๋‹ค์ธต ํผ์…‰ํŠธ๋ก ์ด ๋‹ค๋ฅธ ๋ชจ๋ธ๊ณผ ๋น„๊ตํ•˜์—ฌ ๋ฐ์ดํ„ฐ ํ‘œํ˜„๋ ฅ์ด ๋†’๋‹ค๋Š” ๊ฒƒ์œผ๋กœ ํ•ด์„๋  ์ˆ˜ ์žˆ๋‹ค. ์ด๋Š” ๋ชจ๋ธ ๊ฐœ๋ฐœ์— ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง ๋“ฑ์„ ์‚ฌ์šฉํ•˜๋Š” ๊ฒฝ์šฐ ๋” ๋งŽ์€ ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ์ด ์š”๊ตฌ๋จ์„ ๋œปํ•˜๋ฉฐ, ๋™์‹œ์— ๋” ๋งŽ์€ ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ์„ ์‚ฌ์šฉํ•˜๋Š” ๊ฒฝ์šฐ ๊ณผ์ ํ•ฉ ๊ฐ์†Œ๋ฅผ ํ†ตํ•ด ๊ธฐ์กด ๋ชจ๋ธ๊ณผ ๋น„๊ตํ•˜์—ฌ ์„ฑ๋Šฅ์„ ๋”์šฑ ํ–ฅ์ƒ์‹œํ‚ฌ ๊ฒƒ์œผ๋กœ ์˜ˆ์ƒ๋œ๋‹ค. ๋˜ํ•œ, ์ด ์—ฐ๊ตฌ ๊ฒฐ๊ณผ๋ฅผ ํ•ด์„ํ•˜๋Š”๋ฐ ๊ณ ๋ ค๋˜์–ด์•ผ ํ•˜๋Š” ์š”์†Œ๋Š” ๋ถ„๋ฅ˜๊ธฐ์˜ ์ž…๋ ฅ์ด ์› ์‹ ํ˜ธ๊ฐ€ ์•„๋‹Œ ์ถ”์ถœ๋œ ํŠน์ง•์ด๋ผ๋Š” ๋ฐ ์žˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ๋กœ์ง€์Šคํ‹ฑ ํšŒ๊ท€, ๋žœ๋คํฌ๋ ˆ์ŠคํŠธ ๋“ฑ์€ ์ž…๋ ฅ ๊ฐ’๋“ค์˜ ์‹œ๊ณต๊ฐ„์  ์—ฐ๊ฒฐ์„ฑ์„ ๊ณ ๋ คํ•˜์ง€ ์•Š๋Š” ๋ฐ˜๋ฉด, ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง ๋“ฑ์€ ๋ฐ์ดํ„ฐ ์ž์ฒด์˜ ์‹œ๊ฐ„์ , ๊ณต๊ฐ„์  ์—ฐ๊ฒฐ์„ฑ์„ ํ†ตํ•ด ์ฃผ์š” ํŠน์ง•(feature)๋“ค์„ ์ถ”์ถœํ•˜์—ฌ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์œผ๋กœ ์•Œ๋ ค์ ธ ์žˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ด€์ ์—์„œ ์ด ์—ฐ๊ตฌ๋Š” ์„œ๋กœ ์ƒ๊ด€์„ฑ์ด ์—†๋Š” ์ถ”์ถœ๋œ ํŠน์ง• ๊ฐ’์„ ์ž…๋ ฅ์œผ๋กœ ํ•˜์˜€์œผ๋ฏ€๋กœ ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง์˜ ์žฅ์ ์ด ํฌ์„๋˜์—ˆ์„ ๊ฐ€๋Šฅ์„ฑ์ด ์žˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ์ด ์—ฐ๊ตฌ์—์„œ๋Š” ๋ถ„๋ฅ˜๊ธฐ ํŒŒ๋ผ๋ฏธํ„ฐ ์„ค์ •์— ์žˆ์–ด ๋žœ๋คํฌ๋ ˆ์ŠคํŠธ ๋ถ„๋ฅ˜๊ธฐ์˜ ์ตœ๋Œ€ ๊นŠ์ด์™€ ์ตœ๋Œ€ ๊ฐ€์ง€ ์ˆ˜, ๋‹ค์ธต ํผ์…‰ํŠธ๋ก ์˜ ์€๋‹‰์ธต์˜ ๊นŠ์ด์™€ ํญ, Dropout rate, ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง์˜ ํ•„ํ„ฐ ํฌ๊ธฐ, ์ธต ์ˆ˜ ๋“ฑ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ์ตœ์ ํ™” ์—†์ด ์‚ฌ์šฉํ•˜์˜€์œผ๋ฏ€๋กœ, ๊ฐ ๋ถ„๋ฅ˜๊ธฐ์—์„œ ๋„์ถœ๋œ ๊ฒฐ๊ณผ๋ฅผ ๊ฐ ๋ถ„๋ฅ˜๊ธฐ์—์„œ ๋‚ผ ์ˆ˜ ์žˆ๋Š” ์ตœ๋Œ€ ์„ฑ๋Šฅ์œผ๋กœ ํŒ๋‹จํ•  ์ˆ˜๋Š” ์—†๋‹ค.

5. ๊ฒฐ ๋ก 

์ด ์—ฐ๊ตฌ์˜ ๊ฒฐ๊ณผ์—์„œ๋Š” ๊ด‘์šฉ์ ๋งฅํŒŒ๋ฅผ ์‚ฌ์šฉํ•œ ํ†ต์ฆ ๋ถ„๋ฅ˜์— ์žˆ์–ด ์ถ”์ถœ๋œ ํŒŒํ˜• ํŠน์ง•์„ ์ž…๋ ฅ์œผ๋กœ ํ•˜๋Š” ๋ถ„๋ฅ˜์˜ ๊ฒฝ์šฐ ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง, ๋‹ค์ธต ํผ์…‰ํŠธ๋ก , ๋žœ๋คํฌ๋ ˆ์ŠคํŠธ ๋ถ„๋ฅ˜๊ธฐ์— ๋น„ํ•ด ๋กœ์ง€์Šคํ‹ฑ ํšŒ๊ท€๊ฐ€ ๊ฐ€์žฅ ๊ณผ์ ํ•ฉ์ด ์ ๊ณ  ๋†’์€ ๋ถ„๋ฅ˜ ์ •ํ™•๋„๋ฅผ ๋ณด์ด๋Š” ๊ฒƒ์„ ํ™•์ธํ•˜์˜€๋‹ค. ๋‹จ, ์ด ์—ฐ๊ตฌ์—์„œ๋Š” ์ „์ฒ˜๋ฆฌ๋ฅผ ํ†ตํ•ด ์„œ๋กœ ๋…๋ฆฝ์ ์ธ ํŒŒํ˜•์˜ ํŠน์ง•์„ ์ถ”์ถœํ•œ ํ›„ ์ž…๋ ฅ์œผ๋กœ ์‚ฌ์šฉํ•˜์˜€์œผ๋ฏ€๋กœ, ์ž…๋ ฅ ๋ฒกํ„ฐ ๊ฐ„ ์‹œ๊ณต๊ฐ„์  ์—ฐ๊ฒฐ์„ฑ์ด ์š”๊ตฌ๋˜์ง€ ์•Š๋Š” ๋กœ์ง€์Šคํ‹ฑ ํšŒ๊ท€๋‚˜ ๋žœ๋คํฌ๋ ˆ์ŠคํŠธ์— ์ข€ ๋” ์œ ๋ฆฌํ•œ ๊ฒฐ๊ณผ๊ฐ€ ๋„์ถœ๋˜์—ˆ์„ ์ˆ˜ ์žˆ๋‹ค. ๋”ฐ๋ผ์„œ, ๋ณด๋‹ค ์ •๊ตํ•œ ๋ถ„๋ฅ˜๊ธฐ ๋น„๊ต๋ฅผ ์œ„ํ•ด์„œ๋Š” ํŠน์ง•์„ ์ถ”์ถœํ•˜๊ธฐ ์ด์ „ ์› ์‹ ํ˜ธ๋ฅผ ์ž…๋ ฅ์œผ๋กœ ํ•  ๋•Œ, ๋ถ„๋ฅ˜ ์„ฑ๋Šฅ์— ๋Œ€ํ•œ ๋น„๊ต๊ฐ€ ์ˆ˜ํ–‰๋  ์ˆ˜ ์žˆ๋‹ค. ๋‹ค๋ฅธ ๊ด€์ ์—์„œ, ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง์ด๋‚˜ ๋‹ค์ธต ํผ์…‰ํŠธ๋ก ์€ ๋‹ค๋ฅธ ๋ถ„๋ฅ˜๊ธฐ์— ๋น„ํ•ด ๊ณผ์ ํ•ฉ๋˜๋Š” ๊ฒฝํ–ฅ์„ ๋ณด์˜€์œผ๋‚˜ ์ž…๋ ฅ์— ๋Œ€ํ•œ ํ‘œํ˜„๋ ฅ์ด ํฐ ๊ฒƒ์œผ๋กœ ํ•ด์„๋  ์ˆ˜ ์žˆ์œผ๋ฉฐ ํ•™์Šต์— ์‚ฌ์šฉ๋˜๋Š” ๋ฐ์ดํ„ฐ์˜ ์ฆ๊ฐ€๋ฅผ ํ†ตํ•ด ์„ฑ๋Šฅ์˜ ํ–ฅ์ƒ์„ ๊ธฐ๋Œ€ํ•  ์ˆ˜ ์žˆ์„ ๊ฒƒ์œผ๋กœ ํŒ๋‹จ๋œ๋‹ค. ๋”ฐ๋ผ์„œ, ํ–ฅํ›„ ๋ณด๋‹ค ํฐ ๊ฐœ์ฒด ์ˆ˜๋ฅผ ํ™œ์šฉํ•œ ๋น„๊ต๋ฅผ ํ†ตํ•ด ์ตœ์ ์˜ ๋ถ„๋ฅ˜๊ธฐ๋ฅผ ์„ ์ •ํ•  ์ˆ˜ ์žˆ์„ ๊ฒƒ์œผ๋กœ ๊ธฐ๋Œ€๋œ๋‹ค.

Acknowledgements

This research was supported by a grant of the Basic Science Research Program (NRF-2018R1A4A1025704, NRF-2018R1D1A 3B07046442) and grant of the through the National Research Foundation of Korea (NRF) funded by the Ministry of Science and ICT, Korea.

References

1 
J. A. Grass, 2005, Patient-controlled analgesia, Anesthesia & Analgesia, Vol. 101, No. 5S, pp. S44-S61DOI
2 
M.. Jeanne, C. Clรฉment, J. De Jonckheere, R. Logier, B. Tavernier, 2012, Variations of the analgesia nociception index during general anaesthesia for laparoscopic abdominal surgery, Journal of Clinical Monitoring and Computing, Vol. 26, No. 4, pp. 289-294DOI
3 
M. A. Connelly, J. T. Brown, G. L. Kearns, R. A. Anderson, S. D. St Peter, K. A. Neville, 2014, Pupillometry: a non-invasive technique for pain assessment in paediatric patients, Archives of Disease in Childhood, Vol. 99, No. 12, pp. 1125-1131DOI
4 
T. Ledowski, J. Bromilow, M. Paech, H. Storm, R. Hacking, S. Schug, 2006, Monitoring of skin conductance to assess postoperative pain intensity, British Journal of Anaesthesia, Vol. 97, No. 6, pp. 862-865DOI
5 
M. Huiku, et al., 2007, Assessment of surgicalstress during general anaesthesia, British Journal of Anaesthesia, Vol. 98, No. 4, pp. 447-455DOI
6 
R. F. Rojas, X. Huang, K.-L. Ou, 2019, A Machine Learning Approach for the Identification of a Biomarker of Human Pain using fNIRS, Scientific Reports, Vol. 9, No. 1, pp. 5645DOI
7 
S. Gruss, et al., 2015, Pain Intensity Recognition Rates via Biopotential Feature Patterns with Support Vector Machines, Public Library of Science One, Vol. 10, No. 10, pp. e0140330DOI
8 
L. Breiman, 2001, Random forests, Machine Learning, Vol. 45, No. 1, pp. 5-32DOI
9 
H. S. Seok, B.-M. Choi, G.-J. Noh, H. Shin, 2019, Post- operative Pain Assessment Model Based on Pulse Contour Characteristics Analysis, Institute of Electrical and Electronics Engineers Journal of Biomedical and Health Informatics, Vol. 23, No. 6, pp. 2317-2324DOI
10 
J. Lever, M. Krzywinski, N. Altman, 2016, Points of Significance: Model selection and overfitting, Nature Methods, Vol. 9, No. 13, pp. 703-704Google Search

์ €์ž์†Œ๊ฐœ

์ž„์ง€์—ฐ (Ji Yeon Yim)
../../Resources/kiee/KIEE.2019.68.12.1626/au1.png

J. Y. Yim is an undergraduate student of Department of Biomedical Engineering, Chonnam National University, Yeosu, Republic of Korea.

She is pursuing B. Eng degree since 2015.

์‹ ํ•ญ์‹ (Hangsik Shin)
../../Resources/kiee/KIEE.2019.68.12.1626/au2.png

H. Shin received the B.S., M.S. and Ph.D. degree in electrical and electronic engineering from the Department of Electrical and Electronics Engineering, Yonsei University, Seoul, Korea, in 2003, in 2005 and 2010, respectively.

In 2010, he joined the Digital Media and Communi- cation Research and Development Center of Samsung Electronics, Co. Ltd., Korea.

Since August 2013, he has been with the Department of Biomedical Engineering, Chonnam National University, Yeosu, Korea, where he is an Associate Professor.

His research area includes biomedical signal processing, physiological modeling and computer simulation, u-Healthcare and mobile healthcare Technologies.