Optimal machine learning methods for prediction of high-flow nasal cannula outcomes using image features from electrical impedance tomography.

Background: High-flow nasal cannula (HNFC) is able to provide ventilation support for patients with hypoxic respiratory failure. Early prediction of HFNC outcome is warranted, since failure of HFNC might delay intubation and increase mortality rate. Existing methods require a relatively long period to identify the failure (approximately 12 h) and electrical impedance tomography (EIT) may help identify the patient's respiratory drive during HFNC.

Objectives: This study aimed to investigate a proper machine-learning model to predict HFNC outcomes promptly by EIT image features.

Methods: The Z-score standardization method was adopted to normalize the samples from 43 patients who underwent HFNC and six EIT features were selected as model input variables through the random forest feature selection method. Machine-learning methods including discriminant, ensembles, k-nearest neighbour (KNN), artificial neural network (ANN), support vector machine (SVM), AdaBoost, xgboost, logistic, random forest, bernoulli bayes, gaussian bayes and gradient-boosted decision trees (GBDT) were used to build prediction models with the original data and balanced data proceeded by the synthetic minority oversampling technique.

Results: Prior to data balancing, an extremely low specificity (less than 33.33%) as well as a high accuracy in the validation data set were observed in all the methods. After data balancing, the specificity of KNN, xgboost, random forest, GBDT, bernoulli bayes and AdaBoost significantly reduced (p<0.05) while the area under curve did not improve considerably (p>0.05); and the accuracy and recall decreased significantly (p<0.05).

Conclusions: The xgboost method showed better overall performance for balanced EIT image features, which may be considered as the ideal machine learning method for early prediction of HFNC outcomes.