Comparison of the prediction of extremely low birth weight neonatal mortality by regression analysis and by neural networks☆
Introduction
Mortality in extremely low birthweight (ELBW) infants is disproportionately high compared to larger neonates, although advances in neonatal care such as the use of surfactant have decreased mortality [1]. Mortality and morbidity increase markedly with decreasing gestational age [2], [3] and risk factors such intraventricular hemorrhage, birthweight, low 5-min Apgar scores, male gender, absence of surfactant therapy [4], caucasian race, and outborn status [5] may predispose to mortality in this population. Various scoring systems such as the SNAP [6], SNAP-PE [7], and the CRIB [8] scores have been developed to assess initial neonatal risk. The prediction of mortality in individual neonates is of interest for identifying departures from expected outcomes and quality assurance [9]. Furthermore, decisions regarding parental counseling, place of delivery [10], resuscitation in the delivery room [11], [12], and resource use [13] may be more appropriate if the expected outcome is more accurately estimated. It has been shown that pediatricians who underestimate neonatal survival and freedom from handicap also intervene less often with mechanical ventilation, resuscitation, and other interventions as compared to those who accurately estimate outcomes [14].
The statistical technique most often used for predicting the relationship of a dichotomous dependent variable (such as mortality) to multiple variables is multiple logistic regression analysis. Logistic regression analysis has limitations in some clinical situations as the relationship between an independent and dependent variable may be non-linear (e.g. mortality increases at extremes of birth weight among neonates). Although non-linear relationships can be modeled by various modifications of regression models, the nature of the relationship is not always known a priori. Artificial neural networks, more often simply called “neural networks”, are non-parametric pattern recognition techniques that can recognize complex non-linear relationships or “hidden patterns” between independent and dependent variables [15]. Neural networks are algorithms patterned after the structure of the brain which can “learn” mathematical relationships between a series of independent (input) variables and the corresponding dependent (output or outcome) variables. This learning is achieved by supervised “training” with a training set consisting of inputs with known dependent outcomes. Networks consisting of “layers” of units (nodes) interconnected with lines (connection weights) are programmed to adjust their weights based on the mathematical relationships identified between the inputs and outcomes in the training set [16]. Once trained, the neural network can be used for prediction in a separate test or validation set. Neural networks have been used for the prediction of outcomes in critically ill patients [17] and have been shown to be superior to regression analysis in predicting intracranial hemorrhage in preterm infants [18]. In a study similar to the one presented here, neural networks have also recently been shown to be superior to logistic regression in the prediction of individual mortality using admission data in very low birth weight infants [19]. We devised this study to compare neural networks and regression analysis in the prediction of mortality before discharge in individual ELBW neonates using data readily available within the first 6 h of age.
Section snippets
Methods
A database was created from the University of Alabama at Birmingham Regional neonatal intensive care unit (NICU) computerized records. This study was approved by the Institutional Review Board. Records of 810 extremely low birth weight (ELBW) (birth weight <1000 g) neonates born from January 1990 to December 1996 (both inborn and outborn) were entered into the database and stripped of identifiers. In this time period, neonates >500 g were aggressively resuscitated unless major malformations
Demographics
The demographics of the training set and test set were similar (data also similar for validation set; not shown). The birthweight (mean±SE) in the training set was 749 g±8 vs. 738 g±12 in the test set (p=NS), and the corresponding gestational ages were 25.8 weeks±0.1 vs. 25.5 weeks±0.2 (p=NS). The male:female ratio was equal (50:50) in both groups, and 36% were of Caucasian origin in the training set vs. 38% in the test set (p=NS). The median Apgar score at 5 min in both groups was 8. At least
Discussion
This study addresses the prediction of ELBW mortality using standard statistical techniques such as regression, as compared to the newer software algorithms-based method of neural network analysis. Neural networks closely paralleled results obtained by logistic regression in this data set. The determinants of ELBW mortality and their contribution to mortality were also defined with these two techniques. Both methods seem to perform with comparable accuracy, as noted by the high AUC of the ROC
References (48)
- et al.
Perinatal outcomes of a large cohort of extremely low gestational age infants (twenty-three to twenty-eight completed weeks of gestation)
J. Pediatr.
(1994) - et al.
Introduction to neural networks
Lancet
(1995) Advantages and disadvantages of using artificial neural networks versus logistic regression for predicting medical outcomes
J. Clin. Epidemiol.
(1996)- et al.
Prediction of outcome in critically ill patients using artificial neural network synthesized by genetic algorithm
Lancet
(1996) - et al.
Obstetric determinants of neonatal survival: influence of willingness to perform cesarean delivery on survival of extremely low-birth-weight infants. National Institute of Child Health and Human Development Network of Maternal-Fetal Medicine Units
Am. J. Obstet. Gynecol.
(1997) - et al.
Adolescent pregnancy: understanding the impact of age and race on outcomes
J. Adolesc. Health
(1997) - et al.
Clinical chorioamnionitis and the prognosis for very low birth weight infants
Obstet. Gynecol.
(1998) - et al.
CRIB and SNAP: assessing the risk of death for preterm neonates
Lancet
(1994) Analysis of the clinical variables driving decision in an artificial neural network trained to identify the presence of myocardial infarction
Ann. Emerg. Med.
(1992)- et al.
New Ballard Score, expanded to include extremely premature infants
J. Pediatr.
(1991)
Design construction and evaluation of systems to predict risk in obstetrics
Int. J. Med. Inf.
Outcomes of extremely low birth weight infants
Pediatrics
The limit of viability-neonatal outcome of infants born at 22 to 25 weeks' gestation
N. Engl. J. Med.
Multivariate risks among extremely premature infants
J. Perinatol.
Changes in survival patterns of very low-birth-weight infants from 1980 to 1993
Arch. Pediatr. Adol. Med.
Score for neonatal acute physiology: a physiologic severity index for neonatal intensive care
Pediatrics
Birthweight and illness severity: independent predictors of neonatal severity
Pediatrics
The CRIB (clinical risk index for babies) score: a tool for assessing initial neonatal risk and comparing performance of neonatal intensive care units
Lancet
The use of CRIB (clinical risk index for babies) score in auditing the performance of one neonatal intensive care unit
Acta Paediatr.
The Swedish national prospective study on extremely low birthweight (ELBW) infants. Incidence, mortality, morbidity and survival in relation to level of care
Acta Paediatr.
Delivery room resuscitation decisions for extremely premature infants
Pediatrics
How aggressive should delivery room cardiopulmonary resuscitation be for extremely low birth weight infants?
Pediatrics
Viability, morbidity, and resource use among newborns of 501- to 800-g birth weight. National Institute of Child Health and Human Development Neonatal Research Network
JAMA, J. Am. Med. Assoc.
Estimation of outcome and restriction of interventions in neonates
Pediatrics
Cited by (37)
The rescue and relief plan based on the risk assessment of debris flow in Yunnan Province, China
2021, Natural Hazards ResearchEnvironmental factors prediction in preterm birth using comparison between logistic regression and decision tree methods: An exploratory analysis
2021, Social Sciences and Humanities OpenArtificial Neural Networks, and Evolutionary Algorithms as a systems biology approach to a data-base on fetal growth restriction
2013, Progress in Biophysics and Molecular BiologyCitation Excerpt :AAS allow to structure the software architecture across different levels of analysis and to explicitly reveal in the model the understanding of the rules with which agents interact. In recent years ANNs have been used successfully in medicine as previously mentioned (Penco et al., 2005; Baldassarre et al., 2004; Lahner et al., 2005; Mecocci et al., 2002; Lapuerta et al., 1998; Buchman et al., 1994; DiRusso et al., 2002), and have been applied to solve clinical problems in pediatrics including those related to preterm births (Mueller et al., 2004; Jaing et al., 2001; DeGroff et al., 2001; Namasivayam and Waldemar, 2001; Zernikow et al., 1998). As previously explained, first we fed to the system all the variables we considered in our data base (Table 2), then we identified, using the optimisation method, the most significant (Table 3), and subsequently worked with these.
Early prediction of poor outcome in extremely low birth weight infants by classification tree analysis
2006, Journal of Pediatrics
- ☆
Presented in part at the Society of Pediatric Research meeting, San Francisco, CA, May 1999.