IndividualsThis was a retrospective research, registered at chictr.org (ChiCTR2300067597). This retrospective research was carried out in the ICU of the Affiliated Hospital of Xuzhou Medical University and was accepted by the ethics committee (accepted quantity: XYFY2022-KL464). Due to the retrospective and observational nature of the research, knowledgeable consent was waived. The Affiliated Hospital of Xuzhou Medical University, situated in Xuzhou City, Jiangsu Province, China, is a tertiary hospital. The medical establishment has two hospitals, the east and west hospitals, with 4150 beds.Patients with ARF who acquired HFNO had been screened for enrollment. Inclusion standards: (1) identified as ARF (outlined as oxygenation index ≤ 300 mmHg, the oxygenation index is the proportion of arterial partial stress of oxygen divided by the focus of impressed oxygen) and given HFNO; (2) the age between 18 and 89 years outdated. Exclusion standards: (1) ICU keep < 24 h; (2) a number of admissions to ICU; (3) Patients with incomplete medical information.Clinical information traitsThe traits of medical information in the research are proven in Table 1. The medical information included: (1) baseline traits and comorbidities; (2) important indicators, Glasgow Coma Scale (GCS) rating, medical variables on the first day of HFNO; (3) therapy measures (use of sedation, vasopressors, albumin, diuretics and glucocorticoids); (4) main end result. HFNO failure was outlined as both utility of invasive mechanical air flow or switching to the different research therapy modality.Table 1 Characteristics of patients in coaching and testing information set.Development of machine learning fashionsThe outcome-related characteristic screening course of was carried out utilizing least absolute shrinkage and choice operator (LASSO). For the growth of the fashions, the most related options chosen are employed. Non-zero attribute indicators chosen via LASSO evaluation had been put into the multivariate logistic regression evaluation to establish the unbiased danger components related with HFNO failure.We thought of seven differing kinds of fashions: assist vector machine (SVM), adaptive boosting (ADABOOST), logistic regression (LR), excessive gradient boosting (XGBOOST), stacking ensemble algorithms (STACK), random forest (RF), and naive bayes (NB); The STACK is algorithms that combine LR, SVM, NB, and RF. For the choice of hyper-parameters in fashions, we used 5 fold cross-validation for the choice of hyper-parameters, which additionally helped to successfully forestall the mannequin's over-fitting.Model validationFor validation of prediction mannequin, we divided the information randomly right into a coaching set and validation set in response to a 70–30 break up, after which used the resampling technique for the inner validation of the prediction mannequin in coaching set. Finally, we carried out the validation once more in the validation set. We present extra technical info on the strategies and parameter settings in the Supplementary materials Table 1.Model efficiency and explainabilityTo consider our fashions, we thought of three predictive metrics: space below receiver working attribute (AUROC) curve, Brier rating and space below precision recall curve (AUPRC). AUROC is bounded between 0.5 and 1.0, with increased values being higher. The Brier rating is the imply squared distinction between the predicted chance of HFNO failure and the precise end result (0 or 1 the place 1 signifies failure of HFNO). The Brier rating is bounded between 0 and 1, with decrease values being higher. We moreover in contrast the fashions by plotting their receiver operator attribute (ROC) curves, precision recall (PR) curves and calibration curve.We utilized the Shapley (SHAP) worth to elucidate options in the coaching set. The SHAP abstract, combining characteristic significance with characteristic results, was visualized with dot plots to current the distribution of SHAP The place on the y-axis was decided by the characteristic and that on the x-axis by the SHAP worth. The options are ranked by significance. Moreover, partial dependence plots (PDPs) had been created to visualise the common change in chance of HFNO failure for all values of a predictor whereas retaining all different predictors constant17.Sample measurement and statistical analysisPmsampsize package deal (https://search.r-project.org/CRAN/refmans/pmsampsize/html/pmsampsize.html) in R software program computes the minimal pattern measurement required for the growth of a brand new multivariable prediction mannequin utilizing the standards proposed by Riley et al.18. Riley et al. lay out a collection of standards the pattern measurement ought to meet. These intention to reduce the over-fitting and to make sure exact estimation of key parameters in the prediction mannequin. Following the parameters set in the pmsampsize package deal, we set the c-statistic to 0.80, the potential prediction parameter to eight, and the goal occasion incidence to be 14.1%. Minimum pattern measurement required for brand new mannequin growth based mostly on the above parameters inputs was 459, with 65 occasions. The pattern measurement in the coaching set satisfies the minimal pattern measurement requirement for the growth of a brand new multivariable prediction mannequin.The Kolmogorov–Smirnov check was used to check the regular distribution for measurement information. Normally distributed information had been expressed as means ± normal deviation, and the skewed distributed information was reported as medians with interquartile (twenty fifth–seventy fifth) percentiles. The two teams had been in contrast utilizing pupil t-test or Mann–Whitney U checks. Numeric information had been expressed as a proportion (%), utilizing χ2 or Fisher’s precise chance checks. R software program was used for all analyses (R Foundation for Statistical Computing, Vienna, Austria).Ethics assertionThe research was accepted by the Ethics Committee of the Affiliated Hospital of Xuzhou Medical University (accepted quantity: XYFY2022-KL464).The procedures had been adopted in accordance with the moral requirements of the Ethics Committee of the Affiliated Hospital of Xuzhou Medical University on human experimentation and with the Helsinki Declaration of 1975. Due to the retrospective and observational nature of the research, knowledgeable consent was waived by the Ethics Committee of the Affiliated Hospital of Xuzhou Medical University.
https://www.nature.com/articles/s41598-024-52061-z
Machine learning prediction of the failure of high-flow nasal oxygen therapy in patients with acute respiratory failure
