改进型深度学习网络分析婴儿培养箱质控数据变化趋势的预防性维护

Preventive maintenance of infant incubators based on quality control data trend analysis using an improved deep learning network

摘要

目的：开发一种改进型深度学习网络，通过分析婴儿培养箱质控数据的微小变化趋势，将传统“阈值判定型”的被动维护模式转变为“趋势预测型”的主动干预机制，从而改进高风险医疗设备的预防性维护策略。方法：以婴儿培养箱相邻两次质控检测参数的差值构建特征矩阵，将数据划分为正常运行与故障预警类别。构建改进型TabNet深度学习网络，通过差异敏感特征注意力机制和多尺度特征融合模块增强对参数微小变化的识别能力。多维度评估模型的分类性能、预测准确性、风险量化准确度及设备维护决策价值，并与原始TabNet、深度神经网络及传统机器学习算法进行对比验证。通过特征重要性分析识别影响故障预测的关键参数，揭示婴儿培养箱性能退化的内在机理。结果：改进型TabNet模型在独立测试集上达到92.3%的准确率和75.0%的召回率，显著优于原始TabNet模型87.5%的准确率和7.1%的召回率，同时超越深度神经网络及传统机器学习方法。决策曲线分析进一步证实其在广泛风险阈值范围内均提供最大净获益。特征重要性分析显示，升温时间和相对湿度是预测模型中最具贡献的指标，其累积贡献率接近40%，而操作门附近D区的温度参数异常也显示出重要的预警价值，为设备早期性能退化的识别提供关键信息。结论：本研究成功构建基于改进型深度学习网络的婴儿培养箱预防性维护预测模型，实现对设备性能渐变式衰减的早期识别，推动医疗设备管理从被动响应向主动干预转变，对保障早产儿治疗环境安全具有重要的价值。

Abstract

Objective To develop an improved deep learning network to convert traditional "threshold-based" passive maintenance into "trend prediction-based" proactive intervention through the analysis of subtle changes in infant incubator quality control data, thereby optimizing preventive maintenance strategies for high-risk medical equipment. Methods Feature matrices were constructed from differences between adjacent quality control parameters of infant incubators, with data categorized into normal-condition and fault-warning classes. An improved TabNet deep learning network was developed to improve the detection of subtle parameter changes by incorporating differential-sensitive feature attention mechanisms and multi-scale feature fusion modules. Model performance was comprehensively evaluated across multiple dimensions, including classification capability, prediction accuracy, risk quantification, and decision-making value for equipment maintenance. The proposed model was then compared with original TabNet, deep neural networks, and traditional machine learning algorithms for validation. Feature importance analysis was used to identify key parameters affecting fault prediction and reveal the mechanisms underlying incubator performance degradation. Results The improved TabNet model achieved 92.3% accuracy and 75.0% recall rate on the independent test set, significantly higher than 87.5% accuracy and 7.1% recall rate of the original TabNet, while also outperforming deep neural networks and traditional machine learning methods. Decision curve analysis further confirmed its maximum net benefit across a wide range of risk thresholds. Feature importance analysis identified heating time and relative humidity as the most contributive indicators, with a cumulative contribution rate approaching 40%. Temperature abnormalities near the operation door (zone D) also exhibited significant warning value, providing crucial information for the early detection of equipment performance degradation. Conclusion A preventive maintenance prediction model for infant incubators is successfully constructed based on an improved deep learning network, enabling early identification of progressive equipment performance degradation, and transforming medical equipment management from passive troubleshooting to proactive intervention, with significant value for safeguarding the safety of premature infants' treatment environment.

关键词

婴儿培养箱 / 深度学习 / 质控数据 / 趋势预测 / 预防性维护

Key words

infant incubator / deep learning / quality control data / trend prediction / preventive maintenance

参考文献

[1] 陈沉, 李鑫, 刘玮 . 基于真实世界数据的婴儿培养箱临床应用评价［J］. 医疗装备, 2024, 37(15): 1-5.

[2] Chen C, Li X, Liu W . The evaluation of clinical application of infant incubators based on Real—World data［J］. Medical Equipment, 2024, 37(15): 1-5.

[3] Kinney MV, Lawn JE, Howson CP, et al. 15 Million preterm births annually: what has changed this year?［J］. Reprod Health, 2012, 9: 28.

[4] Özdemirci E, Özarslan Yatak M, Duran F, et al. Reliability assessments of infant incubator and the analyzer［J］. Gazi U J Sci, 2014, 27(4): 1169-1175.

[5] 邹金华, 齐宏亮, 李子好, 等 . 婴儿培养箱质量控制及其规范研究［J］. 现代仪器与医疗, 2023, 29(1): 60-65.

[6] Zou JH, Qi HL, Li ZH, et al. Study on quality control and standardization of infant incubator［J］. Modern Instruments & Medical Treatment, 2023, 29(1): 60-65.

[7] 沈国平 . 医疗器械分类目录查询系统［J］. 中国医疗器械信息, 2018, 24(24): 157-159.

[8] Shen GP . Classification catalogue query system for medical devices ［J］. China Medical Device Information, 2018, 24(24): 157-159.

[9] 虞华, 尹建兵, 张兰, 等 . 婴儿培养箱产品不良事件探讨［J］. 中国医疗器械杂志, 2021, 45(3): 335-339.

[10] Yu H, Yin JB, Zhang L, et al. Study on adverse events of infant incubator products［J］. Chinese Journal of Medical Instrumentation, 2021, 45(3): 335-339.

[11] 曲彦, 牛惠芳 . 基于真实世界数据的婴儿培养箱临床使用风险分析与建议［J］. 中国医疗设备, 2024, 39(3): 98-102.

[12] Qu Y, Niu HF . Risk analysis and recommendations for clinical use of infant incubators based on real—world data［J］. China Medical Devices, 2024, 39(3): 98-102.

[13] 麦达成, 劳婉仪, 肖翔, 等 . 婴儿培养箱三级质量控制管理办法研究与实践［J］. 中国医疗设备, 2022, 37(1): 143-147.

[14] Mai DC, Lao WY, Xiao X, et al. Research and practice on management method of three—level quality control for infant incubator［J］. China Medical Devices, 2022, 37(1): 143-147.

[15] 王才有, 汤学军, 董方杰, 等 . 全国三级医院信息化情况调查研究［J］. 中国卫生信息管理杂志, 2016, 13(4): 342-347.

[16] Wang CY, Tang XJ, Dong FJ, et al. Investigation and research on informatization of the tertiary hospital in China［J］. Chinese Journal of Health Informatics and Management, 2016, 13(4): 342-347.

[17] 张文媛 . 医疗设备的维保成本与管理［J］. 医疗装备, 2018, 31(3): 100-101.

[18] Zhang WY . Maintenance cost and management of medical equipment ［J］. Medical Equipment, 2018, 31(3): 100-101.

[19] Marakala V, Sriramakrishnan GV, Jakka G, et al. Use of deep learning application in medical devices［C］// 2022 4th International Conference on Inventive Research in Computing Applications (ICIRCA). Piscataway, NJ, USA: IEEE, 2022: 935-939.

[20] Chen XW, Lin XT . Big data deep learning: challenges and perspectives ［J］. IEEE Access, 2014, 2: 514-525.

[21] Ahmed SF, Alam MS, Hassan M, et al. Deep learning modelling techniques: current progress, applications, advantages, and challenges ［J］. Artif Intell Rev, 2023, 56(11): 13521-13617.

[22] 牟丽, 夏英华, 何群, 等 . 我国智慧医疗建设现状、问题及对策研究［J］. 中国医院, 2021, 25(1): 24-26.

[23] Mou L, Xia YH, He Q, et al. Research on the status quo, problems and countermeasures of smart medical construction in China［J］. Chinese Hospitals, 2021, 25(1): 24-26.

[24] 国家质量监督检验检疫总局 . 婴儿培养箱校准规范: JJF 1260—2010 ［S］. 北京: 中国质检出版社, 2010.

[25] General Administration of Quality Supervision, Inspection and Quarantine. Calibration Specification for Baby Incubator: JJF 1260— 2010［S］. Beijing: China Quality Inspection Press, 2010.

[26] Li WJ . TabNet for high—dimensional tabular data: advancing interpretability and performance with feature fusion ［C］// 5th International Conference on Signal Processing and Machine Learning (CONF SPML 2025). Piscataway, NJ, USA: IEEE, 2025: 168-173.

[27] Arik SÖ, Pfister T . TabNet: attentive interpretable tabular learning［C］// Proceedings of the Thirty—Fifth AAAI Conference on Artificial Intelligence and the Thirty—Third Conference on Innovative Applications of Artificial Intelligence and the Eleventh Symposium on Educational Advances in Artificial Intelligence. Palo Alto, CA, USA: AAAI Press, 2021: 6679-6687.

[28] 李坤, 马兴华 . 预测性维护在医疗设备管理中的应用研究进展［J］. 中国医疗设备, 2025, 40(1): 154-163.

[29] Li K, Ma XH . Research progress on the application of predictive maintenance in medical device management［J］. China Medical Devices, 2025, 40(1): 154-163.

[30] 宋颖 . 医疗设备故障预测与预防维护策略研究［J］. 中国设备工程, 2024(20): 167-169.

[31] Song Y . Research on medical equipment fault prediction and preventive maintenance strategies［J］. China Plant Engineering, 2024(20): 167-169.

[32] 中华人民共和国国家卫生健康委员会 . 国家卫生健康委办公厅关于印发医院智慧管理分级评估标准体系(试行)的通知: 国卫办医函〔2021〕86号［EB/OL］. (2021—03—15). https://www.nhc.gov.cn/yzygj/c100068/202103/a14c60de4af9423cbbf45712e27e3cc8.shtml.

[33] National Health Commission of the People's Republic of China. Notice of the general office of the national health commission on issuing the graded evaluation standard system for smart hospital management (trial): national health commission medical letter［2021］No. 86［EB/ OL］. (2021—03—15). https://www.nhc.gov.cn/yzygj/c100068/202103/ a14c60de4af9423cbbf45712e27e3cc8.shtml.

[34] Kirana VC, Andayani DH, Pudji A, et al. Effect of closed and opened the door to temperature on PID—based baby incubator with kangaroo mode［J］. Indonesian J Electronic Electromed Eng Med Inform, 2025, 3(3): 121-127.

[35] Rashid FL, Basem A, Hussein AK, et al. Advancements and innovations in thermodynamics for infant incubators: a review［J］. Int J Heat Technol, 2023, 41(6): 1543-1553.

[36] Spahić L, Sredović U, Kurpejović Z, et al. Machine learning for improved medical device management: a focus on infant incubators ［J］. Technol Health Care, 2025, 33(4): 2034-2040.

[37] Kovačević Ž, Gurbeta Pokvić L, Spahić L, et al. Prediction of medical device performance using machine learning techniques: infant incubator case study［J］. Health Technol (Berl), 2020, 10(1): 151-155.

[38] Kholiq A, Syaifudin, Lamidi. Enhancement and maintenance of infant incubators at sukodono health centre in sidoarjo: efforts to ensure continuity of healthcare services［J］. Front Comm Serv Emp, 2024, 3(4): 125-131.

[39] 普利明, 马剑达, 林建子, 等 . DAS28CRP到DAS28ESR换算方程的建立和阈值的推导［J］. 新医学, 2022, 53(3): 182-187.

[40] Pu LM, Ma JD, Lin JZ, et al. Establishment of conversion equation from DAS28CRP to DAS28ESR and derivation of threshold of DAS28CRP［J］. Journal of New Medicine, 2022, 53(3): 182-187.

[41] Lim B, Zohren S . Time—series forecasting with deep learning: a survey ［J］. Philos Trans A Math Phys Eng Sci, 2021, 379(2194): 20200209.

基金资助

宁夏回族自治区卫生健康系统科研课题(2024-NWQP-B058)

北京大学第一医院宁夏妇女儿童医院院级科研课题(YJKY-2023-05)