개인화된 적혈구 수혈 요구량 예측을 위한 임상의사결정지원시스템 = Clinical Decision Support System for Personalized Red Blood Cell Demand

국문 초록 (Abstract)

혈액은 의료전달체계의 근간을 이루는 필수 의약품으로, 헌혈을 통해서만 수급이 가능한 특수성으로 인해 효율적 관리와 사용의 중요성이 강조되어 왔다. 수술 중 필연적으로 발생하는 출혈과 이로 인한 적혈구 수치 저하에 따르는 저산소증을 예방하기 위해, 수술실에서는 막대한 양의 적혈구 제제를 사용하며 이는 전체 적혈구 제제 소비량의 약 60%에 달한다. 수술 전 적정 혈액을 정밀하게 예측하는 것은, 혈액제제 폐기의 방지, 수술 환자의 안전 증진, 더 나아가 범 국가적 혈액자원의 효율적 관리에 이르는 높은 잠재력을 가지고 있다.
한편, 의료 분야에서 진단, 예후 예측, 위험도 예측 등 인공지능을 활용한 다양한 사례들이 보고되며, 일부 사례는 전문의 수준에 이르는 과업수행 성능을 보고하고 있다. 동시에 알고리즘 결정의 근거를 설명할 수 없는 블랙박스적 성격은 의료 인공지능이 임상 실무에 도입되는 것을 저해하는 주요 원인으로 지목되어왔다. 설명가능한 인공지능에 대한 연구의 진보를 통해 공학적 관점에서 기능 수준의 설명을 제공하는 SHapley Additive exPlanations (SHAP) 방식이 소개되어 의료분야 전반에서 널리 활용되고 있으나, 의료진의 시각에서 SHAP을 통한 설명이 충분한지에 대해서는 분야 연구자 간에 일치된 합의가 없었다.
이에 본 연구에서는 흉부외과 수술이 예정되어 있는 환자의 1) 수혈 요구량을 예측하는 기계학습 모델을 개발하고, 2) 이를 활용한 임상의사결정지원시스템 (Clinical Decision Support System, CDSS)으로 구현하고, 3) CDSS 결과부에 설명가능한 인공지능을 포함하여 표현방식에 따른 의료진의 조언 수용도와 신뢰도, 만족도를 비교하는 연구를 수행하였다. 본 연구는 수술 전 환자의 적혈구 수혈량 예측 CDSS의 개발과 검증 및 향후 의료분야에 적용 시 전략에 대한 CDSS의 우수 사례를 제시하였다는 점에서 의의를 가진다.
제1장에서는 흉부외과 수술 환자 데이터를 활용하여 수술 전 적정 수혈량을 예측하는 다수의 기계학습 모델을 개발하고, 가장 우수한 모델을 Precision MSBOS-Thoracic Surgery (pMSBOS-TS)로 선정하였다. 개발된 모델은 전통적으로 수술별 수혈 요구량의 평균값을 사용하여 도출한 최대혈액 신청량 (Maximum Surgical Blood Order Schedule, MSBOS) 방식과 비교하였으며, 예측 성능지표의 향상, 비용의 절감을 확인할 수 있었다.
제2장에서는 pMSBOS-TS 모델을 탑재한 CDSS에서 1) 인공지능 결과값만 제공한 경우, 2) SHAP 만을 제공한 경우 및 3) SHAP과 임상적 해석을 제공한 경우 의료진의 수용도, 신뢰도, 만족도를 비교하는 시뮬레이션 연구를 수행하였다. 인공지능 결과만 제시되었을 때 보다 추가적인 설명이 제시되었을 때 제안의 수용도, 신뢰도, 만족도 모두 향상되는 결과를 보였으며, 이러한 지표가 의료진의 수용도와 약한 양의 상관관계를 보인다는 점을 확인하였다.

다국어 초록 (Multilingual Abstract)

Blood product is a fundamental medicinal component within the healthcare system, and efficient management and use of these are important. Red blood cells serve as essential agents in preventing tissue hypoxia, accounting for approximately 60% of the total red blood cell products used in operating rooms. Preparing appropriate amount of red blood cells before surgery not only allows for the efficient utilization of limited medical resources, but also helps avoid situations where inadequately prepared blood may adversely impact patient safety.
The aim of this study is to develop and validate machine learning models for predicting personalized red blood cell demand among patients with thoracic surgery. We’ll also compare the acceptance, trust, and satisfaction of medical practitioners based on the explainable method when applied as a clinical decision support system. In addition, we aimed to analyze the correlation between acceptance and trust, as well as acceptance and satisfaction.
In Chapter 1, we developed machine learning models to predict the appropriate red blood cells before surgery. These models show higher performance compared to the Maximum Surgical Blood Order Schedule method based on the average blood transfusion demand per surgery. Precisely predicting the necessary amount of blood before surgery showcased the potential to enhance the efficient use of medical resources and ensure the timely transfusion of surgical patients, thereby increasing safety.
The field of medicine has witnessed various applications of artificial intelligence, such as in diagnosis, prognosis, and risk prediction, all demonstrating high performance. Algorithms like deep learning and ensemble techniques have shown superior performance compared to statistical methods like regression analysis. However, their inexplicable black-box nature impedes their integration into the medical field. Research on explainable AI has shown that providing explanations alongside results influences the acceptance and reliability of outcomes positively. Nevertheless, there remains ambiguity regarding the most suitable explanation methods thus far.
In Chapter 2, a simulation study was conducted to compare the acceptance, trustworthiness, and satisfaction of medical practitioners based on the results of the AI model integrated into the CDSS and the methods of explanation. We compared the results between only the AI result was provided and these results were accompanied by either explainable AI or clinical interpretations. The findings indicated that providing additional explanations significantly improved all of these factors compared to presenting only the AI results. Moreover, when clinical interpretations were provided alongside explainable AI methods, the evaluation metrics were notably higher.
This research showcased the development and validation of a CDSS for predicting preoperative red blood cell transfusion volumes in patients undergoing surgery. Additionally, it highlighted exemplary cases for the strategic application of CDSS in future healthcare settings.

목차 (Table of Contents)

• 제1장 흉부외과 수술 환자에서 개인화된 적혈구 수혈 요구량 예측을 위한 기계학습 알고리즘의 개발 및 검증 1
• 1.1. 배경 1
• 1.2. 연구 목적 3
• 1.3. 방법 3
• 1.3.1. 연구 설계 3
• 1.3.2. 연구 대상 4
• 1.3.3. 결과변수 4
• 1.3.4. 통계분석 4
• 1.4. 결과 8
• 1.4.1. 연구 대상자 특성 8
• 1.4.2. 계산된 MSBOS 13
• 1.4.3. 선형 회귀 분석 16
• 1.4.4. 모델 선정 및 예측 결과 비교 18
• 1.4.5. 변수 중요도의 시각화 23
• 1.4.6. 민감도 분석 23
• 1.5. 논의 25
• 1.6. 결론 27
• 제2장 임상의사결정지원시스템의 설명 방법에 따른 의료진의 수용도, 신뢰도, 만족도 비교 연구 29
• 2.1. 배경 29
• 2.2. 연구 목적 32
• 2.3. 방법 32
• 2.3.1. 수술 전 적혈구 요구량 예측을 위한 CDSS 32
• 2.3.2. 연구 설계 및 참가자 모집 33
• 2.3.3. 결과 평가 지표 (Outcome) 35
• 2.3.4. 통계 분석 39
• 2.4. 결과 40
• 2.4.1. 참가자 특성 40
• 2.4.2. 수용도 (WOA) 41
• 2.4.3. 신뢰도 42
• 2.4.4. 만족도 45
• 2.4.5. 수용도와 신뢰도, 만족도의 상관관계 분석 47
• 2.5. 논의 48
• 2.5.1. 주요 발견 48
• 2.5.2. 선행문헌과 비교 49
• 2.5.3. 제한점 51
• 2.6. 결론 52

참고문헌 (Reference)

1. Nosocomial anemia, Eyster E, Bernene J., 223(1):73–4. PMID:4739092, , 1973

2. Python reference manual, Drake Jr FL, Van Rossum G, Centrum voor Wiskunde en Informatica Amsterdam, , 1995

3. Artificial intelligence in medicine, Tremblay, J., Hamet, P., Metabolism Elsevier Inc.69:S36–S40. PMID:28126242, , 2017

4. The role of trust in automation reliance, Dzindolet MT, Peterson SA, Beck HP, Pierce LG, Pomranky RA, 58(6):697– 718. doi: https://doi. org/10.1016/S10715819(03)000387, , 2003

5. Blood transfusion during the COVID19 outbreak, Chen F, Ren M, Wang X, Li L, Lei H, Cai X, Blood Transfusion Edizioni SIMTI SIMTI Servizi Srl18(2):79–82. PMID:32267830, , 2020

6. Patient blood management – The new frontier, Isbister J, Van Aken H., Shander A, Hofmann A, Best Pract Res Clin Anaesthesiol27(1):5–10. PMID:23590911, , 2013

7. Machine learning–based prediction of transfusion, Mitterecker A, Tschoellitsch T, Hochreiter S, Meier J, Leahy MF, Lloyd A, Böck C, Hofmann A, Schwarzbauer K, Trentino KM, Transfusion (Paris) John Wiley & Sons, Ltd60(9):1977–1986. PMID:32596877, , 2020

8. A Survey of Methods for Explaining Black Box Models, Pedreschi D., Turini F, Guidotti R, Ruggieri S, Monreale A, Giannotti F, ACM Comput Surv New York, NY, USA: Association for Computing Machinery51(5). doi: 10.1145/3236009, , 2018

9. Introduction to artificial intelligence in medicine, Brodie R., Mintz Y, Minimally Invasive Therapy & Allied Technologies Taylor & Francis28(2):73–81. doi: 10.1080/13645706.2019.1575882, , 2019

10. DARPA’s explainable artificial intelligence program, Gunning D, Aha DW, AI Mag40(2):44–58. doi: 10.1609/aimag. v40i2.2850, , 2019

11. Why should i trust you? predictions of any classifier, Guestrin C., Singh S, Ribeiro MT, Explaining Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining1317Augu:1135–1144. doi: 10.1145/2939672.2939778, , 2016

12. A HumanGrounded Evaluation of SHAP for Alert Processing, Weerts HJP, van Ipenburg W, Pechenizkiy M., Available from http//arxiv. org/abs/1907.03324, , 2019

13. R: A Language and Environment for Statistical Computing, R Core Team, Vienna, Austria: R Foundation for Statistical Computing Available from https//www rproject. org/accessed Oct 25 2021, , 2017

14. Artificial intelligence in oncology: Path to implementation, Arriaga YE, Chua IS, Levitan NA, Korach ZT, Jackson GP, GazielYablowitz M, Bates DW, Kehl KL, Hassett M., Cancer Med John Wiley & Sons, Ltd10(12):4138–4149. doi: https://doi. org/10.1002/cam4.3935, , 2021

15. A Unified Approach to Interpreting Model Predictions editors, Luxburg U Von, Vishwanathan S, Wallach H, Lundberg SM, Garnett R, Fergus R, Guyon I, Bengio S, Lee SI, In Adv Neural Inf Process Syst Curran Associates, Inc. Available from https//proceedings neurips cc/paper_files/paper/2017/file/8a20a862197 8632d76c43dfd28b67767Paper. pdf, , 2017

16. Towards a rigorous science of interpretable machine learning, Kim, B., Doshi-Velez, F., Ml:1–13 Available from http//arxiv. org/abs/1702.08608, , 2017

17. Red blood cell transfusion policy: A critical literature review, Mengoli C, Liumbruno GM, Vaglio S, Franchini M, Marano G, Pupella S, Muñoz M, Blood Transfusion15(4):307–317. PMID:28661855, , 2017

18. Tell Me a Story! Narrative Driven XAI with Large Language Models, Dams C, Vergouwen M., Hinns J, Martens D, Available from http//arxiv. org/abs/2309.17057, , 2023

19. The Need for User Models in Generating Expert System Explanations, Finin T., Kass R, 1, , 1988

20. Machine learning for predicting preoperative red blood cell demand, Feng Y, Xu Z, Sun X, Wang D, Yu Y, Transfusion Medicine John Wiley & Sons, Ltd31(4):262–270. PMID:34028930, , 2021

21. 67 Consistent Individualized Feature Attribution for Tree Ensembles, Lee SI, Lundberg SM, Erion GG, (2) Available from http://arxiv. org/abs/1802.03888, , 2018

22. Causability and explainability of artificial intelligence in medicine, Holzinger A, Müller H., Zatloukal K, Langs G, Denk H, WIREs Data Mining and Knowledge Discovery John Wiley & Sons, Ltd9(4):e1312. doi: https://doi. org/10.1002/widm.1312, , 2019

23. Unlocking the power of artificial intelligence and big data in medicine, Lovis C., 21(11):1–9. PMID:31702565, , 2019

24. An efficient explanation of individual classifications using game theory, Kononenko I, Štrumbelj E, 11:1–18, , 2010

25. Ethics and Governance of Artificial Intelligence for Health: WHO guidance, World Health Organization, World Health Organization ISBN:9789240029200, , 2021

26. Effects of the COVID19 pandemic on supply and use of blood for transfusion, Prati D, New H V, Thachil J, Massey E, Brunskill S, Apelseth TO, SoOsman C, Germain M, Doree C, Cardigan R, Stanworth SJ, Shehata N, Goldman M, 7(10):e756–e764. PMID:32628911, , 2020

27. Interpretable decision sets: A joint framework for description and prediction, Bach SH, Lakkaraju H, Leskovec J., Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining1317Augu:1675–1684. doi: 10.1145/2939672.2939874, , 2016

28. Maximum surgical blood order schedule in a newly setup tertiary care hospital, Thabah R, Sailo LT, Lanleila M, Yunus M, Lyngdoh NM, Bhattacharyya P., Bardoloi J, Anaesthesia, Pain and Intensive Care17(1):28–32, , 2013

29. Taking Advice: Accepting Help, Improving Judgment, and Sharing Responsibility, Fischer I, Harvey N, Organ Behav Hum Decis Process70(2):117–133. doi: 10.1006/obhd.1997.2697, , 1997

30. A Clinical Decision Support System for the Prediction of Quality of Life in ALS, Wei L, Mooney C., Heverin M, Antoniadi AM, Galvin M, Hardiman O, 12(3). doi: 10.3390/jpm12030435, , 2022

31. Decompositional rule extraction from support vector machines by active learning, Baesens B, Gestel T Van, Martens D, IEEE Trans Knowl Data Eng21(2):178–191. doi: 10.1109/TKDE.2008.131, , 2009

32. Peeking Inside the Black-Box: A Survey on Explainable Artificial Intelligence (XAI), Adadi, A., Berrada, M., IEEE Access IEEE6:52138–52160. doi: 10.1109/ACCESS.2018.2870052, , 2018

33. The utility and prediction of allogeneic blood transfusion use in orthopedic surgery, Brien WF, Gilbert K., Larocque B, 13(2):124–131. PMID:10218236, , 1999

34. European union regulations on algorithmic decision making and a right to explanation., Flaxman S., Goodman B, AI Mag38(3):50– 57. doi: 10.1609/aimag. v38i3.2741, , 2017

35. Machine learning explainability in nasopharyngeal cancer survival using LIME and SHAP, Almangush A, Elmusrati M, Mäkitie AA, Leivo I, Alabi RO, 13(1):8984. doi: 10.1038/s41598023357950, , 2023

36. The maximum surgical blood order schedule and surgical blood use in the United States, Friedman BA, Chadwick AR, Kingdon KI, Oberman HA, Transfusion (Paris)16(4):380–7. PMID:951737, , 1976

37. Reducing intraoperative red blood cell unit wastage in a large academic medical center, Booth GS, Dmochowski RR, Young PP, Austin TM, Paroskie A, Whitney GM, Pilla MA, Sandberg WS, France DJ, Deegan RJ, Woods MC, Transfusion (Paris) Blackwell Publishing Inc.55(11):2752–2758. PMID:26202213, , 2015

38. COVID19 pandemic and transfusion medicine: the worldwide challenge and its implications, Al Mahmasani L, Finianos A, Taher A., Hodroj MH, Ann Hematol100(5):1115–1122. doi: 10.1007/s00277021 04441y, , 2021

39. STATISTICAL METHODS FOR ASSESSING AGREEMENT BETWEEN TWO METHODS OF CLINICAL MEASUREMENT, Martin Bland J, Altman DouglasG, 327(8476):307–310. doi: https://doi. org/10.1016/S01406736(86)908378, , 1986

40. A Severe Sepsis Mortality Prediction Model and Score for Use with Administrative Data Dee, Rosenthal GE, Sarrazin MSV, 307(13):1433– 1435. doi: 10.1097/CCM.0000000000001392. A, , 2012

41. Intelligible models for healthcare: Predicting pneumonia risk and hospital 30day readmission, Gehrke J, Elhadad N., Sturm M, Lou Y, Koch P, Caruana R, Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining2015Augus:1721–1730. doi: 10.1145/2783258.2788613, , 2015

42. Use of Maximum Surgical Order Schedule (MSBOS) among pediatric patients to optimize blood utilization, Guzman JPS, Gepte MBP, Resurreccion LL, 15(1):4. doi: 10.1186/s4315901900059, , 2019

43. Blood Transfusion Policies in Elective General Surgery: How to Optimise CrossMatchtoTransfusion Ratios, Pollard C, Pattenden C, Hollobone C, Hall TC, Dennison AR, Transfusion Medicine and Hemotherapy40(1):27–31. PMID:23637646, , 2013

44. Who is afraid of black box algorithms? On the epistemological and ethical basis of trust in medical AI, Durán JM, Jongsma KR, 47(5):329–335. doi: 10.1136/medethics2020106820, , 2021

45. An empirical evaluation of the comprehensibility of decision table, tree and rule based predictive models, Baesens B, Vanthienen J, Huysmans J, Mues C, Dejaeger K, Decis Support Syst Elsevier B. V.51(1):141– 154. doi: 10.1016/j. dss.2010.12.003, , 2011

46. Artificial intelligence framework for simulating clinical decisionmaking: A Markov decision process approach, Bennett CC, Hauser K., 57(1):9–19. PMID:23287490, , 2013

47. Economic impact of a machine learningbased strategy for preparation of blood products in brain tumor surgery, Oearsakul T, Tunthanathip T, Kaewborisutsakul A, Taweesomboonyat C., Saeheng S, 17(7):e0270916. PMID:35776752, , 2022

48. Prediction of Preoperative Blood Preparation for Orthopedic Surgery Patients: A Supervised Learning Approach, Shen CC, Lee PJ, Chang CM, Hung JH, Hu YH, 8(9). doi: 10.3390/app8091559, , 2018

49. Reducing surgical patient costs through use of an artificial neural network to predict transfusion requirements, Walczak S, Scharf JE, Decis Support Syst30(2):125–138. doi: 10.1016/S0167 9236(00)000932, , 2000

50. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead, Rudin, C., 1:206–215. doi: 10.1038/s422560190048x, , 2019

51. Explainable Artificial Intelligence (XAI): Concepts, taxonomies, opportunities and challenges toward responsible AI, Barbado A, Chatila R, Del Ser J, Herrera F., Benjamins R, Molina D, Bennetot A, Garcia S, DíazRodríguez N, Tabik S, GilLopez S, Barredo Arrieta A, Information Fusion58:82–115. doi: 10.1016/j. inffus.2019.12.012, , 2020

52. Understanding the impact of explanations on advicetaking: a user study for AIbased clinical Decision Support Systems, Beretta A, Panigutti C, Pedreschi D., Giannotti F, Conference on Human Factors in Computing Systems Proceedings doi: 10.1145/3491102.3502104, , 2022

53. Establishment of Maximum Surgical Blood Order Schedule (MSBOS) and Evaluation of Effective Blood Usage in Major Surgery, Kim MH, Kang SY, Jeon Y La, Lee WI, Korean Society of Blood29(1):41–50. doi: 10.17945/kjbt.2018.29.1.41, , 2018

54. Machine Learning for the Prediction of Red Blood Cell Transfusion in Patients During or After Liver Transplantation Surgery, Gui R, Wang YJ, Chen ZW, Zhao QY, Wu J, Liu LP, Dong H, Luo YW, Front Med (Lausanne) Frontiers Media S. A.8:632210. PMID:33693019, , 2021

55. A comparison of usability methods for testing interactive health technologies: Methodological aspects and empirical evidence, Jaspers MWM, Int J Med Inform PMID:19046928, , 2009

56. Current challenges and future opportunities for xai in machine learningbased clinical decision support systems: A systematic review, Antoniadi AM, Du Y, Guendouz Y, Wei L, Mooney C., Mazo C, Becker BA, 11(11):1–23. doi: 10.3390/app11115088, , 2021

57. Explainable Machine Learning Model for Predicting FirstTime Acute Exacerbation in Patients with Chronic Obstructive Pulmonary Disease, Kor CT, Lin CH, Lin PR, Li YR, Lin SH, Wang BY, 12(2). doi: 10.3390/jpm12020228, , 2022

58. Development and validation of a machine learning method to predict intraoperative red blood cell transfusions in cardiothoracic surgery, Morrisey C, Tonna JE, Zhe S, Metcalf RA, Wang Z, Zimmerman J, Sharma V, 12(1):1355. PMID:35079127, , 2022

59. Exploring the Capabilities of a Lightweight CNN Model in Accurately Identifying Renal Abnormalities: Cysts, Stones, and Tumors, Using LIME and SHAP, Bhandari M, Yogarajah P, Kavitha MS, Condell J, 202313(5). doi: 10.3390/app13053125, , 2023

60. From RealWorld Patient Data to Individualized Treatment Effects Using Machine Learning: Current and Future Methods to Address Underlying Challenges, Alaa AM, van der Schaar M., Lambert C, Bica I, Clin Pharmacol Ther John Wiley & Sons, Ltd109(1):87–100. PMID:32449163, , 2021

61. Interpretable prediction of 3year allcause mortality in patients with heart failure caused by coronary heart disease based on machine learning and SHAP, Liu Y, Wang K, Han Q, Ren J, Zheng C, Yang H, Tian J, Zhang Y., Comput Biol Med Elsevier Ltd137:104813. PMID:34481185, , 2021

62. A machine learning model to predict the risk of 30day readmissions in patients with heart failure: A retrospective analysis of electronic medical records data, Hisamitsu T, Kojima G, Kakarmath S, Golas SB, Agboola S, Shibahara T, Sato J, Otaki H, Nakae T, Felsted J, Kvedar J, Jethwani K., BMC Med Inform Decis Mak 2018;18(1):1–17. PMID:29929496, , 2018

63. The Role of XAI in AdviceTaking from a Clinical Decision Support System: A Comparative User Study of Feature ContributionBased and Example59 Based Explanations, Antoniadi AM, McAuliffe FM, McNestry C, Mooney C., Du Y, 12(20). doi: 10.3390/app122010323, , 2022

64. Reducing Unnecessary CrossMatching: A PatientSpecific Blood Ordering System Is More Accurate in Predicting Who Will Receive a Blood Transfusion Than the Maximum Blood Ordering System, Palmer T, O ’ Reilly M, Greenfield MLVH, Wahr JA, 96(2):369–375. PMID:12538179, , 2003