Construction And Validation Of A Predictive Model For Weaning From Mechanical Ventilation In Patients With Sepsis

Objective: Septic patients undergoing mechanical ventilation are susceptible and fragility due to the dual effects of an inflammatory insult to the lung and the external forces of mechanical ventilation,while being more sensitive to the timing of weaning from the machine.In response to the weaning of patients with sepsis,in this study,through the early identification of high-risk patients with sepsis and the evaluation before weaning,we construct the weaning model and accurately predict the probability of weaning failure in real time,so as to provide a reference for the weaning of patients with clinical sepsis.In addition,this study further explored the effect of sequential treatment after weaning in patients with sepsis,aiming to provide a reference for the selection of sequential treatment after weaning in patients with sepsis.Methods: Data for our study were obtained from three large open source intensive care databases and locally collected patient data.The three open source databases are the Intensive Care Medical Information Database Ⅳ(MIMIC-Ⅳ),the Philips e ICU Collaborative Research Database(e ICU-CRD,version 2.0),the Intensive Care Medical Information Database Ⅲ(version 1.4,MIMIC-Ⅲ)and local dataset.Among these datasets,we included patients who were diagnosed with sepsis on the first day of ICU admission and using mechanical ventilation,and the training and test sets of the model were constructed with the MIMIC-Ⅳ set,and the others were validation sets.Finally,the number of patients included in the training set,test set,MIMIC-Ⅲ verification set,e ICUCRD verification set and local verification set was 9571,1595,4821,6624 and 110,respectively.We then constructed a weaning model based on the demographics,laboratory indices,vital signs,and therapeutic indices of the enrolled patients.To make the model more functional,we divided the weaning model into two parts.In the first part,the identification of high-risk patients was achieved by a clustering model using data from the first day of included patients.In part Ⅱ,a prediction model for weaning was constructed using data from the 24 hours prior to weaning and the results of the model in part I.In this way,the model could provide early warning at the beginning of mechanical ventilation and before considering weaning.After constructing the weaning model,in order to make the model more applicable,we developed a the online weaning system based on the model to predict weaning failure in sepsis timely,which can assist clinical making rapid weaning decision.Finally,to further support the weaning of patients with sepsis,we also explored and compared the effect of sequential treatment after weaning of patients with sepsis on the rate of mortality and weaning failure among sepsis patients using a counterfactual model.Results: In the first part of the study,we derived three distinct phenotypes of mechanically ventilated patients with sepsis and named them phenotype I,Ⅱ and Ⅲ in Roman numerical order.Among the three phenotypes,phenotype I patients are high-risk patients and the main manifestations are respiratory and circulatory failure,which has the worst prognosis and the highest rate of weaning failure.The weaning model was constructed based on the results of Part I and 31 clinical common variables for the patients before weaning.The model showed good predictive performance with area under the receiver operating curve(AUROC)values of 0.94,0.82,0.68,0.81,and 0.89 in the training,test,e ICU-CRD,MIMIC-Ⅲ,and local validation sets,respectively.In addition,the model had good interpretability,with ventilation time,respiratory rate,phenotype Ⅰ status,BMI,base excess,inspired oxygen concentration,urine output,heart rate,and age being the most important variables in the model.Further analysis showed only a slight decrease in performance of the retired model built with these nine variables alone,but high stability in multiple validation sets.In the Part Ⅲ,we found that after matching,the odds ratio(OR)of 28-day mortality was 0.40(P = 0.001)and the OR of weaning failure was 0.52(P = 0.006)in patients using high flow nasal cannula(HFNC)oxygen compared with patients without sequential ventilatory support.Patients using non-invasive ventilation(NIV)had an OR of 0.44 for 28-day mortality(P = 0.004)and an OR of 1.30 for weaning from work(P = 0.278)compared to patients without sequential ventilatory support.Conclusions: In this study,we achieved accurate prediction of weaning for sepsis patients on mechanical ventilation through staged prediction.At the same time,we developed an online weaning system that can calculate the failure rate of weaning in real time based on the model,which enhances the practicality of this study.In addition,this study also investigated the impact of different sequential ventilation treatments on the prognosis of sepsis patients after weaning.The research results indicate that the use of both HFNC and NIV can improve the short-term prognosis of sepsis patients.The above results were validated by different data sources or methods,and the research results can be a reference for clinical practice.