ACL網羅的サーベイ報告会2019

Learning from the Experience of Doctors: Automated Diagnosis of Appendicitis Based on Clinical Notes

Transcript

1. "$-໢ཏతαʔϕΠใࠂձ ҩྍݴޠσʔλͱඇݴޠҩྍσʔλͷڠಇ 2019. 11. 2 Yuta Nakamura
2. ࣗݾ঺հ தଜ ༏ଠ (ͳ͔ΉΒ Ώ͏ͨ) ɾ౎಺ۈ຿ͷ์ࣹઢՊҩ ɾ౦ژେֶେֶӃ ҩֶത࢜՝ఔ1೥ (ࣗવݴޠॲཧΛར༻ͨ͠ը૾਍அࢧԉ) @iBotamon
3. લஔ͖
4. /-1͔Βͷࢹ఺ ҩྍػؔ಺ͷσʔλΛ࢖͏λεΫ ྫɿ਍அࢧԉ ← ࠓճ͸ͪ͜Β จॻ࡞੒ࢧԉ ඇߏ଄Խσʔλͷߏ଄Խ ҩྍػؔ֎ͷσʔλΛ࢖͏λεΫ ྫɿҩֶݚڀࢧԉ SNS͔Βͷༀ෺༗֐ࣄ৅ͷൃݟʢࢢൢޙௐࠪʣ SNS͔Βͷ࣬පྲྀߦ༧ଌ ݈߁૬ஊͷͨΊͷର࿩γεςϜ
5. ҩྍ͔Βͷࢹ఺ ͳͥ “ඇݴޠσʔλͱͷڠಇ” ͕໘ന͍͔ʁ ैདྷͷྟচݚڀ ʹ ໖ີʹσβΠϯͨ͠ྟচࢼݧ ɾհೖ܈ / ରর܈ Λ৻ॏʹίϯτϩʔϧ ʴ ৽͍͠ྟচݚڀ ʹ େن໛σʔλ͔Β஌ݟΛಘΔ ɾҩྍݱ৔Ͱ೔ʑ஝ੵ͞Ε͍ͯ͘σʔλΛར༻ (Real World Data) ɾݱࡏ͸ඇݴޠσʔλΛ࢖͏΋ͷ͕ଟ͍ ↑ ࠓޙNLP΋ؔΘ͍ͬͯ͘Մೳੑ େ ҩༀ඼։ൃʹ͓͚ΔϦΞϧϫʔϧυσʔλ׆༻΁ͷظ଴ –੡ༀاۀͷࢹ఺ΑΓ- ༀࡎӸֶ. 2019; 24(1): 19-30.
6. ࿦จ঺հ
7. ࿦จͱ໰୊ઃఆ ٸੑ஬ਨԌΛɼ ɾٹٸ֎དྷͷΧϧςจॻʢʹݴޠσʔλʣ ɾٹٸ֎དྷͷ݂ӷݕࠪσʔλʢʹඇݴޠσʔλʣ ͔Βࣗಈ਍அ͢Δ BioNLP 2019: 18th ACL Workshop on Biomedical Natural Language Processing Steven Kester Yuwono et al. Learning from the Experience of Doctors: Automated Diagnosis of Appendicitis Based on Clinical Notes
8. TL;DR ᶃΧϧςจॻΛΫϦʔχϯάͤͣʹ࢖༻ ᶄඇݴޠσʔλ΋࢖༻ → ҩࢣͱ΄΅ಉ౳ͷ਍அੑೳ
9. Background: ٸੑ஬ਨԌ appendicitis ɾ஬ਨʹੜͨ͡ײછ঱ɼ͍ΘΏΔ “໡௎” ɾ঱ঢ়ɿయܕతʹ͸Έ͓ͧͪ → ӈԼෲ෦ʹҠಈ͢Δ௧Έ ɾ࣏ྍɿ߅ەༀ (߅ੜ෺࣭) or खज़ ʰ࠷৽ΨΠυϥΠϯ४ڌ ফԽث࣬ױ਍அɾ࣏ྍࢦ਑ʱ. தࢁॻళ. pp.338-340.
10. Background: ٸੑ஬ਨԌ appendicitis ɾ਍அɿ (1) ঱ঢ়΍֤छݕࠪΛ;·͑૯߹తʹ൑அ (2) είΞϦϯάγεςϜΛར༻ 9-10: very probable 7-8: probable 5-6: possible ʰ࠷৽ΨΠυϥΠϯ४ڌ ফԽث࣬ױ਍அɾ࣏ྍࢦ਑ʱ. தࢁॻళ. pp.338-340.
11. Background: ਍அ→࣏ྍ·ͰͷྲྀΕ དྷӃ ֎དྷ਍࡯ ɾ໰਍಺༰ɼ਎ମॴݟ ← ࢖͏ (ݴޠσʔλ) ɾ݂ӷݕࠪ ← ࢖͏ (ඇݴޠσʔλ) ɾը૾ݕࠪ (௒Ի೾, Xઢ, CT) ← ࢖Θͳ͍ ஬ਨԌͷՄೳੑߴ ஬ਨԌͷՄೳੑ௿ ೖӃ → ͞ΒͳΔݕࠪ ࣏ྍ → ୀӃ ɾ࠷ऴతͳ਍அ ← ୀӃαϚϦʔʹهࡌ ← Ground Truth ஬ਨԌ ஬ਨԌͰͳ͍ ࣏ྍ → ୀӃ
12. ໰୊ઃఆ (࠶ܝ) ٸੑ஬ਨԌΛɼ ɾٹٸ֎དྷͷΧϧςจॻ ɾٹٸ֎དྷͷ݂ӷݕࠪσʔλ ͔Βࣗಈ਍அ͢Δ free-textͷٹٸ֎དྷΧϧς + ݂ӷݕࠪ਺஋σʔλ ↓ ٸੑ஬ਨԌͷ probability score (0ʙ1) ೖྗ ग़ྗ
13. Materials & Methods σʔληοτ ɾγϯΨϙʔϧࠃཱେֶෟଐපӃͷ ٹٸ֎དྷΧϧς & ୀӃαϚϦʔ 180,000 ૊ ɾClass: positive 1.6%, negative 98.4% ࢦඪ ɾF0.5 score (precisionΛॏࢹ)
14. ͪͳΈʹʜ ҩྍݴޠσʔλ͸かなりԚ͍ ๩͍͠਍ྍͷ߹ؒʹॻ͔ΕΔ ↓ ɾஅยత ɾՕ৚ॻ͖ ɾུޠ ɾޡࣈɼ୤ࣈ <num> nkda complain of rif pain – since this afternoon - no vomiting / diarrhea - no fever o/e: afebrile vitals stable h <num> l clear a soft, rif tenderness, rebound positive fbc renal panel tw <num>
15. Convolution Encoder-Decoder CoNLL-2014 จ๏ޡΓగਖ਼λεΫ F0.5 45.36 (baseline:38.54) Chollampatt S et al. A Multilayer Convolutional Encoder-Decoder Neural Network for Grammatical Error Correction. AAAI 2018.
16. Embedding (word2vec (300d) Λ train dataͰ܇࿅) 1D Convolution (window = 3, filter = 300) Residual Connection 1 LSTM Residual Connection 2 Self Attention ݂ӷݕࠪͷ਺஋σʔλΛ جఈؔ਺ʹ௨ͯ͠concat Affine + Sigmoid score pad Convolutional Residual Recurrent Model (CR2 Model)
17. Result Baseline 1 ٹٸ֎དྷͰͷҩࢣͷ਍அ Alvarado score ≧ 7఺ CR2 model Baseline 1 Baseline 2 Model F0.5 0.686 F0.5 0.539 F0.5 0.579 '1͸গͳ͍͕ '/͸ଟ͍
18. Result Baseline 1 ٹٸ֎དྷͰͷҩࢣͷ਍அ Alvarado score ≧ 7఺ CR2 model Baseline 1 Baseline 2 Model F0.5 0.686 0.900 F0.5 0.539 0.821 F0.5 0.579 0.895 1PTJUJWF
    
    OFHBUJWF
    
    
    ͔ͭ /FHBUJWF΋͢΂ͯෲ෦࣬ױͷEBUBTFU
19. Result Attention Layer ͸ “஬ਨԌΒ͍͠঱ঢ়” ͱ “஬ਨԌΒ͘͠ͳ͍঱ঢ়” Λ ೝ͍ࣝͯͨ͠ ӈෲ෦௧ ӈ௎ࠎᜰͷ௧Έ ൓௓௧ ᙤ௧ Լཀྵ
20. Discussion ҙຯ͸͋Δͷ͔ʁ → ը૾ݕࠪΛҰ੾࢖͍ͬͯͳ͍ͷ͕ಛ௃ ஬ਨԌͷ਍ྍͰCT͕ࡱ૾͞ΕΔػձ͕ٸ૿͍ͯ͠Δ͕ ਍ྍͷoutcomeʹد༩͍ͯ͠ͳ͍ͱ͍͏ࢦఠ͕͋Δ → ແବͳը૾ݕࠪΛݮΒͤΔՄೳੑ Michael D Repplinger et al. J Am Coll Radiol. 2016; 13(9): 1050-1056. Sheraz R Markar et al. Int J Surg. 2014; 12(4): 357-360.
21. கࢮత͕ͩ਍அ΋೉͍͠ɼٸੑݺٵଅഭ঱ީ܈(ARDS) Λ ूத࣏ྍࣨͰૣظൃݟ͢ΔͨΊͷݚڀ Free-text ਍ྍه࿥ͷ΄͔, ଟ͘ͷඇݴޠҩྍσʔλΛ࢖༻ (ICD code, όΠλϧαΠϯ, ݂ӷݕࠪ, ݺٵঢ়ଶ, ҙࣝঢ়ଶ) → AUC 92.61ʙ93.59 ಉछͷ΄͔ͷ࿦จ BioNLP 2019: 18th ACL Workshop on Biomedical Natural Language Processing Emilia Apostolova et al. Combining Structured and Free-text Electronic Medical Record Data for Real-time Clinical Decision Support
22. ɾҩྍݴޠॲཧͱඇݴޠσʔλͷڠಇΛѻͬͨ ACL BioNLP Workshop ͷ࿦จΛ঺հ͠·ͨ͠ ɾࠓޙ·͢·͢ൃల͕ظ଴͞ΕΔྖҬ ·ͱΊ