Predicting low cognitive ability at age 5 years using perinatal data and machine learning.

BACKGROUND: There are no early, accurate, scalable methods for identifying infants at high risk of poor cognitive outcomes in childhood. We aim to develop an explainable predictive model, using machine learning and population-based cohort data, for this purpose. METHODS: Data were from 8858 participants in the Growing Up in Ireland cohort, a nationally representative study of infants and their primary caregivers (PCGs). Maternal, infant, and socioeconomic characteristics were collected at 9-months and cognitive ability measured at age 5 years. Data preprocessing, synthetic minority oversampling, and feature selection were performed prior to training a variety of machine learning models using ten-fold cross validated grid search to tune hyperparameters. Final models were tested on an unseen test set. RESULTS: A random forest (RF) model containing 15 participant-reported features in the first year of infant life, achieved an area under the receiver operating characteristic curve (AUROC) of 0.77 for predicting low cognitive ability at age 5. This model could detect 72% of infants with low cognitive ability, with a specificity of 66%. CONCLUSIONS: Model performance would need to be improved before consideration as a population-level screening tool. However, this is a first step towards early, individual, risk stratification to allow targeted childhood screening. IMPACT: This study is among the first to investigate whether machine learning methods can be used at a population-level to predict which infants are at high risk of low cognitive ability in childhood. A random forest model using 15 features which could be easily collected in the perinatal period achieved an AUROC of 0.77 for predicting low cognitive ability. Improved predictive performance would be required to implement this model at a population level but this may be a first step towards early, individual, risk stratification.

Improvement in the Prediction of Neonatal Hypoxic-Ischemic Encephalopathy with the Integration of Umbilical Cord Metabolites and Current Clinical Makers.

OBJECTIVE: To validate our previously identified candidate metabolites, and to assess the ability of these metabolites to predict hypoxic-ischemic encephalopathy (HIE) both individually and combined with clinical data. STUDY DESIGN: Term neonates with signs of perinatal asphyxia, with and without HIE, and matched controls were recruited prospectively at birth from 2 large maternity units. Umbilical cord blood was collected for later batch metabolomic analysis by mass spectroscopy along with clinical details. The optimum selection of clinical and metabolites features with the ability to predict the development of HIE was determined using logistic regression modelling and machine learning techniques. Outcome of HIE was determined by clinical Sarnat grading and confirmed by electroencephalogram grade at 24 hours. RESULTS: Fifteen of 27 candidate metabolites showed significant alteration in infants with perinatal asphyxia or HIE when compared with matched controls. Metabolomic data predicted the development of HIE with an area under the curve of 0.67 (95% CI, 0.62-0.71). Lactic acid and alanine were the primary metabolite predictors for the development of HIE, and when combined with clinical data, gave an area under the curve of 0.96 (95% CI, 0.92-0.95). CONCLUSIONS: By combining clinical and metabolic data, accurate identification of infants who will develop HIE is possible shortly after birth, allowing early initiation of therapeutic hypothermia.

Validation of a machine learning algorithm for identifying infants at risk of hypoxic ischaemic encephalopathy in a large unseen data set.

OBJECTIVE: To validate a hypoxic ischaemic encephalopathy (HIE) prediction algorithm to identify infants at risk of HIE immediately after birth using readily available clinical data. DESIGN: Secondary review of electronic health record data of term deliveries from January 2017 to December 2021. SETTING: A tertiary maternity hospital. PATIENTS: Infants >36 weeks' gestation with the following clinical variables available: Apgar Score at 1 min and 5 min, postnatal pH, base deficit, and lactate values taken within 1 hour of birth INTERVENTIONS: Previously trained open-source logistic regression and random forest (RF) prediction algorithms were used to calculate a probability index (PI) for each infant for the occurrence of HIE. MAIN OUTCOME: Validation of a machine learning algorithm to identify infants at risk of HIE in the immediate postnatal period. RESULTS: 1081 had a complete data set available within 1 hour of birth: 76 (6.95%) with HIE and 1005 non-HIE. Of the 76 infants with HIE, 37 were classified as mild, 29 moderate and 10 severe. The best overall accuracy was seen with the RF model. Median (IQR) PI in the HIE group was 0.70 (0.53-0.86) vs 0.05 (0.02-0.15), (p<0.001) in the non-HIE group. The area under the receiver operating characteristics curve for prediction of HIE=0.926 (0.893-0.959, p<0.001). Using a PI cut-off to optimise sensitivity of 0.30, 936 of the 1081 (86.5%) infants were correctly classified. CONCLUSION: In a large unseen data set an open-source algorithm could identify infants at risk of HIE in the immediate postnatal period. This may aid focused clinical examination, transfer to tertiary care (if necessary) and timely intervention.