Machine Learning Approaches to Predict In-Hospital Mortality among Neonates with Clinically Suspected Sepsis in the Neonatal Intensive Care Unit

Abstract

Background: preterm and critically ill neonates often experience clinically suspected sepsis during their prolonged hospitalization in the neonatal intensive care unit (NICU), which can be the initial sign of final adverse outcomes. Therefore, we aimed to utilize machine learning approaches to predict neonatal in-hospital mortality through data-driven learning. Methods: a total of 1095 neonates who experienced clinically suspected sepsis in a tertiary-level NICU in Taiwan between August 2017 and July 2020 were enrolled. Clinically suspected sepsis was defined based on clinical features and laboratory criteria and the administration of empiric antibiotics by clinicians. The variables used for analysis included patient demographics, clinical features, laboratory data, and medications. The machine learning methods used included deep neural network (DNN), k-nearest neighbors, support vector machine, random forest, and extreme gradient boost. The performance of these models was evaluated using the area under the receiver operating characteristic curve (AUC). Results: the final in-hospital mortality of this cohort was 8.2% (90 neonates died). A total of 765 (69.8%) and 330 (30.2%) patients were randomly assigned to the training and test sets, respectively. Regarding the efficacy of the single model that most accurately predicted the outcome, DNN exhibited the greatest AUC (0.923, 95% confidence interval [CI] 0.953–0.893) and the best accuracy (95.64%, 95% CI 96.76–94.52%), Cohen’s kappa coefficient value (0.74, 95% CI 0.79–0.69) and Matthews correlation coefficient value (0.75, 95% CI 0.80–0.70). The top three most influential variables in the DNN importance matrix plot were the requirement of ventilator support at the onset of suspected sepsis, the feeding conditions, and intravascular volume expansion. The model performance was indistinguishable between the training and test sets. Conclusions: the DNN model was successfully established to predict in-hospital mortality in neonates with clinically suspected sepsis, and the machine learning algorithm is applicable for clinicians to gain insights and have better communication with families in advance.

A Continuous Late-Onset Sepsis Prediction Algorithm for Preterm Infants using Multimodal Physiological Signals from a Patient Monitor

“…In future work, varying segment length [38], and more advanced features and algorithms such as network-based features or deep neural networks could be investigated for LOS prediction to further improve the LOS-related physiology conveyed in the multimodal signals. These may also better capture the temporal information relating to disease development [39]. Such methods have been successfully applied to several disease predictions such as ventricular fibrillation [40], sepsis in adults [41], and sleep arousal disorder [42].…”

On usage of artificial intelligence for predicting mortality during and post-pregnancy: a systematic review of literature

“…Table 3 presents the metrics used in the primary works. Sensitivity and accuracy appears in 11 works [29,32,30,20,21,14,22,31,26,12,33,27,24], specificity in nine [32,20,21,14,22,31,26,12,27], F1-score, and precision in seven [32,30,14,22,31,26,33,23,24].…”

“…Machine learning was the most common modeling technique found among the primary works of this SLR, being proposed by 16 of the 18 works [29,32,30,21,10,14,22,25,31,26,12,33,15,27,28,24]. Deep learning models in turn were proposed by four works [20,33,23,15].…”

“…When analyzing Figure 5, one can note that deep learning models appeared from 2019 on wards, showing that there may be a large field of search in relation to these models. Regarding the neonatal mortality works, ten machine learning models were proposed [29,21,10,22,25,31,26,27,28,24], two deep learning models were proposed [20,23], and seven logistic regression models were proposed [10,25,31,26,23,28,24]. Even though infant mortality was the focus of more works than stillbirth, the total of proposed models was the same, seven for each.…”

“…Of the primary works of this SLR, 10 of them (more than half) did not applied any hyperparameter optimization. Of the 8 works that used it, 6 used a technique called Grid Search [29,30,14,25,27,24]. Grid Search is a traditional technique that uses an exhaustive search within a given limited search space [42].…”