New Approach based on Machine Learning for Short-Term Mortality Prediction in Neonatal Intensive Care Unit

Abstract

Mortality remains one of the most important outcomes to predict in Intensive Care Units (ICUs). In fact, the sooner mortality is predicted, the better critical decisions are made by doctors based on patient's illness severity. In this paper, a new approach based on Machine Learning (ML) techniques for short-term mortality prediction in Neonatal Intensive Care Unit (NICU) is proposed. This approach relies on many steps. At first, relevant features are selected from available data upon neonates' admission and from the time-series variables collected within the two first hours of stay in the NICU from the Medical Information Mart for Intensive Care III (MIMIC-III). After that, to predict mortality, many classifiers were tested which are Linear Discriminant Analysis (LDA), K-Nearest Neighbors (KNN), Classification and Regression Trees (CART), Logistic Regression (LR), Support Vector Machine (SVM), Naïve Bayes (NB) and Random Forest (RF). The experimental results showed that LDA was the best performing classifier with an accuracy equal to 0.947 and AUROC equal to 0.97 with 31 features. The third step of this approach is mortality time prediction using the Galaxy-Random Forest method achieving an f-score equal to 0.871. The proposed approach compared favorably in terms of time, accuracy and AUROC with existing scoring systems and ML techniques. It is the first work predicting neonates mortality based on ML techniques and time-series data after only two hours of admission to the NICU.

Machine Learning Models for Predicting Neonatal Mortality: A Systematic Review

“…After identifying articles through the various databases and checking articles’ references for additional studies, a total of 484 articles were screened. After screening and removal of duplicates, a total of 29 full-text articles were read and 11 articles met inclusion criteria [17-27]. Reasons for exclusion are described in the flow diagram in Figure 2.…”

“…However, this should not be confused with a LR ML approach. Of the studies examined, 4 studies (36%) compared AI-based models to machine learning LR models [18, 23, 26, 27]. Three studies (27%) compared ML to the traditional LR approach [17, 22, 23].…”

“…These predictors were incorporated in at least 7 studies (64%). The K-nearest neighbor and least discriminant models by Kefi et al [27], with 17 and 31, respective features, had the highest AUC at 97.0%. The model with the least number of features and highest AUC was a neural network model built by Ambalavanan et al [18], which had an AUC of 81.6% and used only 3 variables (prenatal steroids, non-Hispanic blacks, and gestational age).…”

“…While there is limited research on what specifically defines low versus high discriminative power, Steyerberg et al [37] claim that models with an AUC around 60% as having low discriminative power versus high discriminative models having an AUC closer to 95%. Based on these criteria, only 5 of the models in the review had AUCs of 95% or higher, these were from Kefi et al [27] and Zernikow et al [26] with a range of 12–31 features and were models derived from linear discriminant analysis, K-nearest neighbor, and artificial neural network.…”

Unstructured clinical notes within the 24 hours since admission predict short, mid & long-term mortality in adult ICU patients

“…Over the years, researchers have focused on different granularity of time range for in-hospital mortality prediction such as 6-hour mortality in ICU patients [22], 8 to 24-hour, 1-day and 2-day mortality in NICU patients [23], 24-hour mortality in acute myocardial infarction patients [24], 7-day mortality in acute heart failure patients [25], 28-day mortality in sepsisinduced coagulopathy patients [26], 90-Day survival in acutely ill patients [27], and so on.…”