目的 探索构建基于深度学习的心肺复苏术(cardiopulmonary resuscitation,CPR)实时人工智能辅助培训和考核系统,为CPR培训和考核提供有效便捷的方法。方法 2021年3月至2022年3月获取郑州大学第一附属医院临床技能培训中心90名学员的心肺复苏培训和考试数据,随机分为训练组(72例)与测试组(18例)。应用训练组视频资料和深度学习模型,构建CPR人工智能实时辅助培训和考核系统,对两组学员视频中身体的关键点进行检测和定位,提取CPR的考核指标,实时展示在移动终端上。比较两组学员的考核指标,并与传感器显示的考核指标进行比较,同时比较人工智能评分与考官评分,采用t检验、非参数检验或方差分析。结果 视频人工智能分析结果显示,训练组和测试组每轮按压时间分别为(16.82±1.24)秒、(16.88±1.22)秒,按压深度分别为[5.85(5.60,6.00)]厘米、[5.80(5.50,6.10)] 厘米,按压频率分别为(107.64±8.13) 次/分钟、(107.11±8.49)次/分钟,差异均无统计学意义(P>0.05)。人工智能分析结果和传感器测量结果显示,按压时间分别为(16.83±1.23)秒、(16.80±1.21)秒,按压频率分别为(107.42±8.07) 次/分钟、(107.71±8.00)次/分钟,差异均无统计学意义(P>0.05)。人工智能与3位考官的评价分别为(68.92±3.20)分,(68.92±3.28)分,(68.83±3.24)分,(68.84±3.25)分,差异均无统计学意义(P>0.05)。结论 CPR人工智能实时辅助培训和考核系统能够通过对学员CPR操作时身体关键点的检测和定位,较为准确地实时检测学员CPR操作情况并在移动终端显示,这有助于学员CPR的培训和考核。
Objective To explore the development of a real-time artificial intelligence (AI) assisted training and assessment system for cardiopulmonary resuscitation (CPR) based on deep learning, providing an effective and convenient method for CPR training and assessment. Methods Video data and sensor data of 90 trainees undergoing CPR training and examination at the Clinical Skills Training Center of the First Affiliated Hospital of Zhengzhou University from March 2021 to March 2022 were collected and randomly divided into a training group (72 participants) and a testing group (18 participants). Using the training group video data and deep learning models, an AI real-time assisted training and assessment system for CPR was constructed to detect and locate the key points of trainees′ bodies during CPR and extract the assessment indicators, which were displayed in real-time on mobile devices. The assessment indicators obtained from the video analysis by AI in the two groups of trainees were compared, as well as the comparison between the assessment indicators obtained from the video analysis and the sensor data and the comparison between the scores given by the AI system and the human examiners. T-test, non-parametric test or analysis of variance was used for statistical analyses. Results The AI analysis of the video data showed that the compression time, depth, and frequency for both the training and testing groups were (16.82±1.24) and (16.88±1.22) seconds, [5.85(5.60, 6.00)] and [5.80(5.50, 6.10)] cm, (107.64±8.13) and (107.11±8.49) times/minute, respectively, with no statistical difference (P>0.05). The assessment indicators obtained from the video analysis by AI and the sensor data showed that the compression time and frequency for both the training and testing groups were (16.83±1.23) and (16.80±1.21) seconds, (107.42±8.07) and (107.71±8.00) times/minute, respectively, with no statistical difference (P>0.05) between the two groups. The scores given by AI system and the human examiner were (68.92±3.20), (68.92±3.28), (68.83±3.24), (68.84±3.25), respectively. There was no statistical difference between the scores given by the AI system and the human examiners (P>0.05). Conclusions The real-time AI assisted training and assessment system for CPR detects and locates the key points of trainees′ bodies during CPR, and accurately displays the situation of trainees′ CPR in real-time on mobile devices, which is helpful for CPR training and assessment.
[1] 国家心血管病中心. 中国心血管健康与疾病报告2020[M] . 北京:科学出版社, 2021:96-97.
[2] Böttiger BW, Lockey A, Aickin R, et al. Up to 206 million people reached and over 5.4 million trained in cardiopulmonary resuscitation worldwide: the 2019 International Liaison Committee on Resuscitation World Restart a Heart initiative[J] . J Am Heart Assoc, 2020,9(15):e017230. DOI: 10.1161/JAHA.120.017230.
[3] Kiguchi T, Okubo M, Nishiyama C, et al. Out-of-hospital cardiac arrest across the world: first report from the International Liaison Committee on Resuscitation (ILCOR)[J] . Resuscitation, 2020,152:39-49. DOI: 10.1016/j.resuscitation.2020.02.044.
[4] 冀娟. 全国心肺复苏普及进亿家精准健康工程实施体会[J] .世界最新医学信息文摘(电子版),2018(97):285-285,287.
[5] 陶秀萍. 公众急救技能培训模式之“情境教育”[J] .中华卫生应急电子杂志,2018,4(3):179-181. DOI: 10.3877/cma.j.issn.2095-9133.2018.03.013.
[6] Dick-Smith F, Power T, Martinez-Maldonado R, et al. Basic life support training for undergraduate nursing students: an integrative review[J] . Nurse Educ Pract, 2021,50:102957. DOI: 10.1016/j.nepr.2020.102957.
[7] 靳杰, 董文龙, 樊毫军, 等. 心肺复苏模拟人研究现状[J] .医疗卫生装备,2021,42(8):80-84. DOI: 10.19745/j.1003-8868.2021173.
[8] Elola A, Aramendi E, Irusta U, et al. Convolutional recurrent neural networks to characterize the circulation component in the thoracic impedance during out-of-hospital cardiac arrest[J] . Annu Int Conf IEEE Eng Med Biol Soc, 2019,2019:1921-1925. DOI: 10.1109/EMBC.2019.8857758.
[9] Esteva A, Kuprel B, Novoa RA, et al. Dermatologist-level classification of skin cancer with deep neural networks[J] . Nature, 2017,542(7639):115-118. DOI: 10.1038/nature21056.
[10] Bazarevsky V, Kartynnik Y, Vakunov A, et al. Blazeface: sub-millisecond neural face detection on mobile GPUs[EB/OL] . [2022-06-01] . https://arxiv.org/abs/1907.05047.
[11] Bazarevsky V, Zhang F. On-device, real-time hand tracking with Mediapipe[EB/OL] . (2019-08-19)[2022-06-01] . https://ai.googleblog.com/2019/08/on-device-real-time-hand-tracking-with.html.
[12] Pulli K, Baksheev A, Kornyakov K, et al. Real-time computer vision with OpenCV[J] . Communications of the ACM, 2012,55(6):61-69. DOI: 10.1145/2184319.2184337.
[13] Lugaresi C, Tang J, Nash H, et al. Mediapipe: a framework for building perception pipelines[EB/OL] .[2022-06-01] . https://arxiv.org/abs/1906.08172.
[14] 中国医师协会毕业后医学模拟教育专家委员会, 中国医药教育协会医学模拟教育专业委员会. 医学模拟中心建设标准专家共识(2017)[J] .中华医学教育杂志,2017,37(6):803-807. DOI: 10.3760/cma.j.issn.1673-677X.2017.06.002.
