Assessment of faults and abnormalities of machine signs of non-invasive ventilator for COPD pulmonary rehabilitation based on deep learning
10.3969/j.issn.1672-8270.2025.06.008
- VernacularTitle:基于深度学习的评估慢性阻塞性肺疾病肺康复治疗中无创呼吸机故障及患者体征异常研究
- Author:
Bo ZHAO
1
;
Xiaomei HAN
1
;
Hongwei FENG
1
Author Information
1. 复旦大学附属中山医院吴淞医院呼吸与危重症医学科 上海 200940
- Publication Type:Journal Article
- Keywords:
Convolutional neural network(CNN);
Chronic obstructive pulmonary disease(COPD);
Pulmonary rehabilitation;
Fault of non-invasive ventilator;
Abnormalities of machine sign
- From:
China Medical Equipment
2025;22(6):39-44
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To propose an alarm method for ventilator fault and abnormality of machine signs based on convolutional neural network(CNN)model,which aimed at the problem of low detection rate of ventilator faults and abnormalities of machine sign,and to analyze its application value in patients with chronic obstructive pulmonary disease(COPD)in using non-invasive ventilator.Methods:This study established a identification model(CNN-SG)for ventilator fault based on stochastic gradient descent by mini-batch method that was introduced by CNN network,and a multi-task CNN identification model for abnormalities of machine signs of ventilator was established on the basis of CNN model.The respiratory waveform data of 60 patients with chronic obstructive pulmonary disease(COPD)who admitted to Wusong Hospital,Zhongshan Hospital Affiliated to Fudan University from January 2019 to January 2024 were collected.These data were divided into train set(42 cases)and validation set(18 cases)as ratio of 7 to 3.The training set was used in the learning process of model,while the validation set was used to assess the model's performance for unknown data.One-hot encoding was used to represent fault states,and to predict whether occurred abnormal data failures within specific time intervals.Two independent convolutional networks were employed respectively to detect and identify invalid inspiratory effort and double-triggering abnormalities in abnormal machine signs.Model performance was assessed by using accuracy,precision,recall rate,negative predictive value(NPV),and specificity.Chi-square test was used to compare the differences of various assessment indicators between different models.The F1-score was calculated to comprehensively assess the performance of model.Results:In the validation set,the accuracy(99.98%),precision(99.90%),recall rate(99.97%),NPV(98.68%),and specificity(98.24%)of CNN-SG model were higher than those of CNN model in identify the fault of ventilator.In detecting invalid inspiratory effort,the accuracy,sensitivity,specificity and F1-scores of CNN-MTL model were respectively 98.83%,98.81%,97.68%and 98.85.The accuracy,sensitivity,specificity and F1-scores of CNN-MTL model were respectively 98.70%,98.83%,97.62%,and 98.75 in identifying double-triggering positive features,as well as 98.80%,98.81%,98.88%and 98.99 for double-triggering negative features.All of the above indicators of CNN-SG were significantly better than those of the conventional CNN model.Conclusion:The established deep learning algorithm model based on CNN can effectively identify faults and abnormalities of machine signs of non-invasive ventilator.
