Machine Learning Algorithms to Predict Mortality of Neonates on Mechanical Intubation for Respiratory Failure

Abstract

Background: Early identification of critically ill neonates with poor outcomes can optimize therapeutic strategies. We aimed to examine whether machine learning (ML) methods can improve mortality prediction for neonatal intensive care unit (NICU) patients on intubation for respiratory failure. Methods: A total of 1734 neonates with respiratory failure were randomly divided into training (70%, n = 1214) and test (30%, n = 520) sets. The primary outcome was the probability of NICU mortality. The areas under the receiver operating characteristic curves (AUCs) of several ML algorithms were compared with those of the conventional neonatal illness severity scoring systems including the NTISS and SNAPPE-II. Results: For NICU mortality, the random forest (RF) model showed the highest AUC (0.939 (0.921–0.958)) for the prediction of neonates with respiratory failure, and the bagged classification and regression tree model demonstrated the next best results (0.915 (0.891–0.939)). The AUCs of both models were significantly better than the traditional NTISS (0.836 (0.800–0.871)) and SNAPPE-II scores (0.805 (0.766–0.843)). The superior performances were confirmed by higher accuracy and F1 score and better calibration, and the superior and net benefit was confirmed by decision curve analysis. In addition, Shapley additive explanation (SHAP) values were utilized to explain the RF prediction model. Conclusions: Machine learning algorithms increase the accuracy and predictive ability for mortality of neonates with respiratory failure compared with conventional neonatal illness severity scores. The RF model is suitable for clinical use in the NICU, and clinicians can gain insights and have better communication with families in advance.

The past, current, and future of neonatal intensive care units with artificial intelligence: a systematic review

“…Early studies employed ANN and fuzzy linguistic models and achieved an AUC of 85–95% and accuracy of 90% 62 , 68 . New studies in a large preterm populations and extremely low birthweight infants found an AUC of 68.9–93.3% 65 , 71 . There are some shortcomings in these studies; for example, none of them used vital parameters to represent dynamic changes, and hence, there was no improvement in clinical practice in neonatology.…”

“…Only Apgar scores could be changed. Hsu et al, 2020 71 RF KNN ANN XGBoost Elastic-net To predict mortality of neonates when they were on mechanical intubation 1734 neonates 70% training 30% test Mortality scores Patient demographics Lab results Blood gas analysis Respirator parameters Cardiac inotrop agents from onset of respiratory failure to 48 hours 93.9% (AUC) RF has achieved the highest prediction of mortality + Employed several ML and statistics + Explained the feature analysis and importance into analysis - Two center study - Algorithmic bias - Inability to real time prediction Stocker et al, 2022 75 RF To predict blood culture test positivity according to the all variables, all variables without biomarkers, only biomarkers, only risk factors, and only clinical signs 1710 neonates from 17 centers Secondary analysis of NeoPInS data Biomarkers(4 variables) Risk factors (4 variables) Clinical signs(6 variables) Other variables(14) All variables (28) They included to RF analysis to predict culture positive early onset sepsis Only biomarkers 73.3% (AUC) All variables 83.4% (AUC) Biomarkers are the most important contributor + CRP and WBC are the most important variables in the model + Decrease the overtreatment + Multi-center data - Overfitting of the model due to the discrepancy with currently known clinical practice - Seemed not evaluated the clinical signs and risk factors which are really important in daily practice Temple et al, 2016 229 supervised ML and NLP To identify patients that will be medically ready for discharge in the subsequent 2–10 days. 4693 patients (103,206 patient-days 178 NLP using a bag of words (BOW) surgical diagnoses, pulmonary hypertension, retinopathy of prematurity, and psychosocial issues 63.3% (AUC) 67.7% (AUC) 75.2% (AUC) 83.7% (AUC) + Could potentially avoid over 900 (0.9%) hospital days …”

Risk factors for death caused by early onset sepsis in neonates: a retrospective cohort study

“…5) An abnormal procalcitonin (PCT) level was defined as a level with a cut-off point with time according to the age adjustment requirement as in a previous study [ 12 ]. 6) Persistent pulmonary hypertension of the newborn (PPHN) was defined according to the clinical presentation of refractory hypoxemia and echocardiographic evidence with an estimated peak systolic pulmonary-artery pressure that was higher than 35 mmHg or more than two-thirds of the systemic systolic pressure as indicated by a tricuspid regurgitant jet, a right-to-left ductus arteriosus shunt, or a right-to-left atrial-level shunt [ 13 ] 0.7) Respiratory failure was diagnosed according to the consensus of Jen-Fu Hsu et al [ 14 ] and our institution: ① Clinical manifestations of respiratory distress, and ② requirement respiratory support using a noninvasive or invasive ventilator to maintain a target arterial blood gas: pH value > 7.25, PaO 2 > 50 mmHg, PaCO 2 < 55 mmHg. 9) Diagnostic criteria of heart failure was meeting all the following clinical characteristics [ 15 ]: ① shortness of breath > 60 times/min; ② tachycardia > 160 times/min; ③ heart enlargement (X-ray showed cardiothoracic ratio > 0.6 or echocardiography proven); ④ pulmonary edema, and ⑤ liver enlargement > 3 cm or galloping rhythm.…”

Comparing different scoring systems for predicting mortality risk in preterm infants: a systematic review and network meta-analysis

“…A total of 5,804 studies were obtained in the initial search, of which 4,072 were found through the database search and 89 were obtained from other sources. Nineteen studies were finally included after screening according to the inclusion and exclusion criteria (7)(8)(9)(10)(11)(12)(13)(14)(15)(16)(17)(18)(19)(20)(21)(22)(23)(24)(25). The study selection process is illustrated in Figure 1.…”