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.

Prediction of 2-Year Cognitive Outcomes in Very Preterm Infants Using Machine Learning Methods.

Importance: Early intervention can improve cognitive outcomes for very preterm infants but is resource intensive. Identifying those who need early intervention most is important. Objective: To evaluate a model for use in very preterm infants to predict cognitive delay at 2 years of age using routinely available clinical and sociodemographic data. Design, Setting, and Participants: This prognostic study was based on the Swedish Neonatal Quality Register. Nationwide coverage of neonatal data was reached in 2011, and registration of follow-up data opened on January 1, 2015, with inclusion ending on September 31, 2022. A variety of machine learning models were trained and tested to predict cognitive delay. Surviving infants from neonatal units in Sweden with a gestational age younger than 32 weeks and complete data for the Bayley Scales of Infant and Toddler Development, Third Edition cognitive index or cognitive scale scores at 2 years of corrected age were assessed. Infants with major congenital anomalies were excluded. Exposures: A total of 90 variables (containing sociodemographic and clinical information on conditions, investigations, and treatments initiated during pregnancy, delivery, and neonatal unit admission) were examined for predictability. Main Outcomes and Measures: The main outcome was cognitive function at 2 years, categorized as screening positive for cognitive delay (cognitive index score <90) or exhibiting typical cognitive development (score ≥90). Results: A total of 1062 children (median [IQR] birth weight, 880 [720-1100] g; 566 [53.3%] male) were included in the modeling process, of whom 231 (21.8%) had cognitive delay. A logistic regression model containing 26 predictive features achieved an area under the receiver operating curve of 0.77 (95% CI, 0.71-0.83). The 5 most important features for cognitive delay were non-Scandinavian family language, prolonged duration of hospitalization, low birth weight, discharge to other destination than home, and the infant not receiving breastmilk on discharge. At discharge from the neonatal unit, the full model could correctly identify 605 of 650 infants who would have cognitive delay at 24 months (sensitivity, 0.93) and 1081 of 2350 who would not (specificity, 0.46). Conclusions and Relevance: The findings of this study suggest that predictive modeling in neonatal care could enable early and targeted intervention for very preterm infants most at risk for developing cognitive impairment.

Neonatal EEG graded for severity of background abnormalities in hypoxic-ischaemic encephalopathy.

This report describes a set of neonatal electroencephalogram (EEG) recordings graded according to the severity of abnormalities in the background pattern. The dataset consists of 169 hours of multichannel EEG from 53 neonates recorded in a neonatal intensive care unit. All neonates received a diagnosis of hypoxic-ischaemic encephalopathy (HIE), the most common cause of brain injury in full term infants. For each neonate, multiple 1-hour epochs of good quality EEG were selected and then graded for background abnormalities. The grading system assesses EEG attributes such as amplitude, continuity, sleep-wake cycling, symmetry and synchrony, and abnormal waveforms. Background severity was then categorised into 4 grades: normal or mildly abnormal EEG, moderately abnormal EEG, majorly abnormal EEG, and inactive EEG. The data can be used as a reference set of multi-channel EEG for neonates with HIE, for EEG training purposes, or for developing and evaluating automated grading algorithms.

Big data, machine learning, and population health: predicting cognitive outcomes in childhood.

The application of machine learning (ML) to address population health challenges has received much less attention than its application in the clinical setting. One such challenge is addressing disparities in early childhood cognitive development-a complex public health issue rooted in the social determinants of health, exacerbated by inequity, characterised by intergenerational transmission, and which will continue unabated without novel approaches to address it. Early life, the period of optimal neuroplasticity, presents a window of opportunity for early intervention to improve cognitive development. Unfortunately for many, this window will be missed, and intervention may never occur or occur only when overt signs of cognitive delay manifest. In this review, we explore the potential value of ML and big data analysis in the early identification of children at risk for poor cognitive outcome, an area where there is an apparent dearth of research. We compare and contrast traditional statistical methods with ML approaches, provide examples of how ML has been used to date in the field of neurodevelopmental disorders, and present a discussion of the opportunities and risks associated with its use at a population level. The review concludes by highlighting potential directions for future research in this area. IMPACT: To date, the application of machine learning to address population health challenges in paediatrics lags behind other clinical applications. This review provides an overview of the public health challenge we face in addressing disparities in childhood cognitive development and focuses on the cornerstone of early intervention. Recent advances in our ability to collect large volumes of data, and in analytic capabilities, provide a potential opportunity to improve current practices in this field. This review explores the potential role of machine learning and big data analysis in the early identification of children at risk for poor cognitive outcomes.

Predicting Low Cognitive Ability at Age 5-Feature Selection Using Machine Learning Methods and Birth Cohort Data.

Objectives: In this study, we applied the random forest (RF) algorithm to birth-cohort data to train a model to predict low cognitive ability at 5 years of age and to identify the important predictive features. Methods: Data was from 1,070 participants in the Irish population-based BASELINE cohort. A RF model was trained to predict an intelligence quotient (IQ) score ≤90 at age 5 years using maternal, infant, and sociodemographic features. Feature importance was examined and internal validation performed using 10-fold cross validation repeated 5 times. Results The five most important predictive features were the total years of maternal schooling, infant Apgar score at 1 min, socioeconomic index, maternal BMI, and alcohol consumption in the first trimester. On internal validation a parsimonious RF model based on 11 features showed excellent predictive ability, correctly classifying 95% of participants. This provides a foundation suitable for external validation in an unseen cohort. Conclusion: Machine learning approaches to large existing datasets can provide accurate feature selection to improve risk prediction. Further validation of this model is required in cohorts representative of the general population.

Towards Deeper Neural Networks for Neonatal Seizure Detection.

Machine learning and more recently deep learning have become valuable tools in clinical decision making for neonatal seizure detection. This work proposes a deep neural network architecture which is capable of extracting information from long segments of EEG. Residual connections as well as data augmentation and a more robust optimizer are efficiently exploited to train a deeper architecture with an increased receptive field and longer EEG input. The proposed system is tested on a large clinical dataset of 4,570 hours of duration and benchmarked on a publicly available Helsinki dataset of 112 hours duration. The performance has improved from an AUC of 95.41% to an AUC of 97.73% when compared to a deep learning baseline.

Grading hypoxic-ischemic encephalopathy in neonatal EEG with convolutional neural networks and quadratic time-frequency distributions.

Objective.To develop an automated system to classify the severity of hypoxic-ischaemic encephalopathy injury (HIE) in neonates from the background electroencephalogram (EEG).Approach. By combining a quadratic time-frequency distribution (TFD) with a convolutional neural network, we develop a system that classifies 4 EEG grades of HIE. The network learns directly from the two-dimensional TFD through 3 independent layers with convolution in the time, frequency, and time-frequency directions. Computationally efficient algorithms make it feasible to transform each 5 min epoch to the time-frequency domain by controlling for oversampling to reduce both computation and computer memory. The system is developed on EEG recordings from 54 neonates. Then the system is validated on a large unseen dataset of 338 h of EEG recordings from 91 neonates obtained across multiple international centres.Main results.The proposed EEG HIE-grading system achieves a leave-one-subject-out testing accuracy of 88.9% and kappa of 0.84 on the development dataset. Accuracy for the large unseen test dataset is 69.5% (95% confidence interval, CI: 65.3%-73.6%) and kappa of 0.54, which is a significant (P<0.001) improvement over a state-of-the-art feature-based method with an accuracy of 56.8% (95% CI: 51.4%-61.7%) and kappa of 0.39. Performance of the proposed system was unaffected when the number of channels in testing was reduced from 8 to 2-accuracy for the large validation dataset remained at 69.5% (95% CI: 65.5%-74.0%).Significance.The proposed system outperforms the state-of-the-art machine learning algorithms for EEG grade classification on a large multi-centre unseen dataset, indicating the potential to assist clinical decision making for neonates with HIE.

Deep Learning for EEG Seizure Detection in Preterm Infants.

EEG is the gold standard for seizure detection in the newborn infant, but EEG interpretation in the preterm group is particularly challenging; trained experts are scarce and the task of interpreting EEG in real-time is arduous. Preterm infants are reported to have a higher incidence of seizures compared to term infants. Preterm EEG morphology differs from that of term infants, which implies that seizure detection algorithms trained on term EEG may not be appropriate. The task of developing preterm specific algorithms becomes extra-challenging given the limited amount of annotated preterm EEG data available. This paper explores novel deep learning (DL) architectures for the task of neonatal seizure detection in preterm infants. The study tests and compares several approaches to address the problem: training on data from full-term infants; training on data from preterm infants; training on age-specific preterm data and transfer learning. The system performance is assessed on a large database of continuous EEG recordings of 575[Formula: see text]h in duration. It is shown that the accuracy of a validated term-trained EEG seizure detection algorithm, based on a support vector machine classifier, when tested on preterm infants falls well short of the performance achieved for full-term infants. An AUC of 88.3% was obtained when tested on preterm EEG as compared to 96.6% obtained when tested on term EEG. When re-trained on preterm EEG, the performance marginally increases to 89.7%. An alternative DL approach shows a more stable trend when tested on the preterm cohort, starting with an AUC of 93.3% for the term-trained algorithm and reaching 95.0% by transfer learning from the term model using available preterm data. The proposed DL approach avoids time-consuming explicit feature engineering and leverages the existence of the term seizure detection model, resulting in accurate predictions with a minimum amount of annotated preterm data.

In vitro evaluation of delays in the adjustment of the fraction of inspired oxygen during CPAP: effect of flow and volume.

BACKGROUND: Adjusting the fraction of inspired oxygen (FiO2) delivered to preterm infants to keep their oxygen saturation within target range remains challenging. Closed-loop automated FiO2 control increases the time infants spend within the assigned target range. The delay with which FiO2 adjustments at the ventilator result in a change in the inspired gas limits the performance of both manual and automated controls. OBJECTIVE: To evaluate the equilibration time (Teq) between FiO2 adjustments and changes in FiO2 reaching the patient. METHODS: In vitro determination of the delay in FiO2 adjustments at the ventilator at 5 and 8 L/min of gas flow and two different humidifier/ventilator circuit volumes (840 and 432 mL). RESULTS: Teq values were 31, 23, 20 and 17 s for the volume-flow combinations 840 mL+5 L/min, 840 mL+8 L/min, 432 mL+5 L/min and 432 mL+8 L/min, respectively. CONCLUSION: The identified delay seems clinically relevant and should be taken into account during manual and automatic control of FiO2.

Identifying tracé alternant activity in neonatal EEG using an inter-burst detection approach.

Electroencephalography (EEG) is an important clinical tool for reviewing sleep-wake cycling in neonates in intensive care. Tracé alternant (TA)-a characteristic pattern of EEG activity during quiet sleep in term neonates-is defined by alternating periods of short-duration, high-voltage activity (bursts) separated by lower-voltage activity (inter-bursts). This study presents a novel approach for detecting TA activity by first detecting the inter-bursts and then processing the temporal map of the bursts and inter-bursts. EEG recordings from 72 healthy term neonates were used to develop and evaluate performance of 1) an inter-burst detection method which is then used for 2) detection of TA activity. First, multiple amplitude and spectral features were combined using a support vector machine (SVM) to classify bursts from inter-bursts within TA activity, resulting in a median area under the operating characteristic curve (AUC) of 0.95 (95% confidence interval, CI: 0.93 to 0.98). Second, post-processing of the continuous SVM output, the confidence score, was used to produce a TA envelope. This envelope was used to detect TA activity within the continuous EEG with a median AUC of 0.84 (95% CI: 0.80 to 0.88). These results validate how an inter-burst detection approach combined with post processing can be used to classify TA activity. Detecting the presence or absence of TA will help quantify disruption of the clinically important sleep-wake cycle.

Grading the severity of hypoxic-ischemic encephalopathy in newborn EEG using a convolutional neural network.

Electroencephalography (EEG) is a valuable clinical tool for grading injury caused by lack of blood and oxygen to the brain during birth. This study presents a novel end-to-end architecture, using a deep convolutional neural network, that learns hierarchical representations within raw EEG data. The system classifies 4 grades of hypoxic-ischemic encephalopathy and is evaluated on a multi-channel EEG dataset of 63 hours from 54 newborns. The proposed method achieves a testing accuracy of 79.6% with one-step voting and 81.5% with two-step voting. These results show how a feature-free approach can be used to classify different grades of injury in newborn EEG with comparable accuracy to existing feature-based systems. Automated grading of newborn background EEG could help with the early identification of those infants in need of interventional therapies such as hypothermia.

Neonatal seizure detection from raw multi-channel EEG using a fully convolutional architecture.

A deep learning classifier for detecting seizures in neonates is proposed. This architecture is designed to detect seizure events from raw electroencephalogram (EEG) signals as opposed to the state-of-the-art hand engineered feature-based representation employed in traditional machine learning based solutions. The seizure detection system utilises only convolutional layers in order to process the multichannel time domain signal and is designed to exploit the large amount of weakly labelled data in the training stage. The system performance is assessed on a large database of continuous EEG recordings of 834h in duration; this is further validated on a held-out publicly available dataset and compared with two baseline SVM based systems. The developed system achieves a 56% relative improvement with respect to a feature-based state-of-the art baseline, reaching an AUC of 98.5%; this also compares favourably both in terms of performance and run-time. The effect of varying architectural parameters is thoroughly studied. The performance improvement is achieved through novel architecture design which allows more efficient usage of available training data and end-to-end optimisation from the front-end feature extraction to the back-end classification. The proposed architecture opens new avenues for the application of deep learning to neonatal EEG, where the performance becomes a function of the amount of training data with less dependency on the availability of precise clinical labels.

Prediction of short-term health outcomes in preterm neonates from heart-rate variability and blood pressure using boosted decision trees.

BACKGROUND AND OBJECTIVE: Efficient management of low blood pressure (BP) in preterm neonates remains challenging with considerable variability in clinical practice. There is currently no clear consensus on what constitutes a limit for low BP that is a risk to the preterm brain. It is argued that a personalised approach rather than a population based threshold is more appropriate. This work aims to assist healthcare professionals in assessing preterm wellbeing during episodes of low BP in order to decide when and whether hypotension treatment should be initiated. In particular, the study investigates the relationship between heart rate variability (HRV) and BP in preterm infants and its relevance to a short-term health outcome. METHODS: The study is performed on a large clinically collected dataset of 831 h from 23 preterm infants of less than 32 weeks gestational age. The statistical predictive power of common HRV features is first assessed with respect to the outcome. A decision support system, based on boosted decision trees (XGboost), was developed to continuously estimate the probability of neonatal morbidity based on the feature vector of HRV characteristics and the mean arterial blood pressure. RESULTS: It is shown that the predictive power of the extracted features improves when observed during episodes of hypotension. A single best HRV feature achieves an AUC of 0.87. Combining multiple HRV features extracted during hypotensive episodes with the classifier achieves an AUC of 0.97, using a leave-one-patient-out performance assessment. Finally it is shown that good performance can even be achieved using continuous HRV recordings, rather than only focusing on hypotensive events - this had the benefit of not requiring invasive BP monitoring. CONCLUSIONS: The work presents a promising step towards the use of multimodal data in providing objective decision support for the prediction of short-term outcome in preterm infants with hypotensive episodes.

Suitability of an inter-burst detection method for grading hypoxic-ischemic encephalopathy in newborn EEG.

Electroencephalography (EEG) is an important clinical tool for grading injury caused by lack of oxygen or blood to the brain during birth. Characteristics of low-voltage waveforms, known as inter-bursts, are related to different grades of injury. This study assesses the suitability of an existing inter-burst detection method, developed from preterm infants born <; 30 weeks of gestational age, to detect inter-bursts in term infants. Different features from the temporal organisation of the inter-bursts are combined using a multi-layer perceptron (MLP) machine learning algorithm to classify four grades of injury in the EEG. We find that the best performing feature, percentage of inter-bursts, has an accuracy of 59.3%. Combining this with the maximum duration of inter-bursts in the MLP produces a testing accuracy of 77.8%, with similar performance to existing multi-feature methods. These results validate the use of the preterm detection method in term EEG and show how simple measures of the inter-burst interval can be used to classify different grades of injury.

Heart Rate Variability during Periods of Low Blood Pressure as a Predictor of Short-Term Outcome in Preterms.

Efficient management of low blood pressure (BP) in preterm neonates remains challenging with a considerable variability in clinical practice. The ability to assess preterm wellbeing during episodes of low BP will help to decide when and whether hypotension treatment should be initiated. This work aims to investigate the relationship between heart rate variability (HRV), BP and the short-term neurological outcome in preterm infants less than 32 weeks gestational age (GA). The predictive power of common HRV features with respect to the outcome is assessed and shown to improve when HRV is observed during episodes of low mean arterial pressure (MAP) - with a single best feature leading to an AUC of 0.87. Combining multiple features with a boosted decision tree classifier achieves an AUC of 0.97. The work presents a promising step towards the use of multimodal data in building an objective decision support tool for clinical prediction of short-term outcome in preterms who suffer episodes of low BP.

Investigating the Impact of CNN Depth on Neonatal Seizure Detection Performance.

This study presents a novel, deep, fully convolutional architecture which is optimized for the task of EEG-based neonatal seizure detection. Architectures of different depths were designed and tested; varying network depth impacts convolutional receptive fields and the corresponding learned feature complexity. Two deep convolutional networks are compared with a shallow SVMbased neonatal seizure detector, which relies on the extraction of hand-crafted features. On a large clinical dataset, of over 800 hours of multichannel unedited EEG, containing 1389 seizure events, the deep 11-layer architecture significantly outperforms the shallower architectures, improving the AUC90 from 82.6% to 86.8%. Combining the end-to-end deep architecture with the feature-based shallow SVM further improves the AUC90 to 87.6%. The fusion of classifiers of different depths gives greatly improved performance and reduced variability, making the combined classifier more clinically reliable.

Coupling between mean blood pressure and EEG in preterm neonates is associated with reduced illness severity scores.

Hypotension or low blood pressure (BP) is a common problem in preterm neonates and has been associated with adverse short and long-term neurological outcomes. Deciding when and whether to treat hypotension relies on an understanding of the relationship between BP and brain functioning. This study aims to investigate the interaction (coupling) between BP and continuous multichannel unedited EEG recordings in preterm infants less than 32 weeks of gestational age. The EEG was represented by spectral power in four frequency sub-bands: 0.3-3 Hz, 3-8 Hz, 8-15 Hz and 15-30 Hz. BP was represented as mean arterial pressure (MAP). The level of coupling between the two physiological systems was estimated using linear and nonlinear methods such as correlation, coherence and mutual information. Causality of interaction was measured using transfer entropy. The illness severity was represented by the clinical risk index for babies (CRIB II score) and contrasted to the computed level of interaction. It is shown here that correlation and coherence, which are linear measures of the coupling between EEG and MAP, do not correlate with CRIB values, whereas adjusted mutual information, a nonlinear measure, is associated with CRIB scores (r = -0.57, p = 0.003). Mutual information is independent of the absolute values of MAP and EEG powers and quantifies the level of coupling between the short-term dynamics in both signals. The analysis indicated that the dominant causality is from changes in EEG producing changes in MAP. Transfer entropy (EEG to MAP) is associated with the CRIB score (0.3-3 Hz: r = 0.428, p = 0.033, 3-8 Hz: r = 0.44, p = 0.028, 8-15 Hz: r = 0.416, p = 0.038) and indicates that a higher level of directed coupling from brain activity to blood pressure is associated with increased illness in preterm infants. This is the first study to present the nonlinear measure of interaction between brain activity and blood pressure and to demonstrate its relation to the initial illness severity in the preterm infant. The obtained results allow us to hypothesise that the normal wellbeing of a preterm neonate can be characterised by a nonlinear coupling between brain activity and MAP, whereas the presence of weak coupling with distinctive directionality of information flow is associated with an increased mortality rate in preterms.

Modelling interactions between blood pressure and brain activity in preterm neonates.

Hypotension or low blood pressure (BP) is a common problem in preterm neonates and has been associated with adverse short and long-term outcomes. Deciding when and whether to treat hypotension relies on an understanding of the relations between blood pressure and brain function. This study aims to investigate the interaction between BP and multichannel EEG in preterm infants less than 32 weeks gestational age. The mutual information is chosen to model interaction. This measure is independent of absolute values of BP and electroencephalography (EEG) power and quantifies the level of coupling between the short-term dynamics in both signals. It is shown that while adverse health conditions as measured by higher clinical risk indices for babies (CRIB II) are accompanied by consistently lower blood pressure (r=0.43), no significant correlation was observed between CRIB scores and EEG spectral power. More importantly, the chosen measure of interaction between dynamics of EEG and BP was found to be more closely related to CRIB scores (r=0.49, p-value=0.012), with higher CRIB score associated with lower levels of interaction.

Toward a Personalized Real-Time Diagnosis in Neonatal Seizure Detection.

The problem of creating a personalized seizure detection algorithm for newborns is tackled in this paper. A probabilistic framework for semi-supervised adaptation of a generic patient-independent neonatal seizure detector is proposed. A system that is based on a combination of patient-adaptive (generative) and patient-independent (discriminative) classifiers is designed and evaluated on a large database of unedited continuous multichannel neonatal EEG recordings of over 800 h in duration. It is shown that an improvement in the detection of neonatal seizures over the course of long EEG recordings is achievable with on-the-fly incorporation of patient-specific EEG characteristics. In the clinical setting, the employment of the developed system will maintain a seizure detection rate at 70% while halving the number of false detections per hour, from 0.4 to 0.2 FD/h. This is the first study to propose the use of online adaptation without clinical labels, to build a personalized diagnostic system for the detection of neonatal seizures.

Exploring temporal information in neonatal seizures using a dynamic time warping based SVM kernel.

Seizure events in newborns change in frequency, morphology, and propagation. This contextual information is explored at the classifier level in the proposed patient-independent neonatal seizure detection system. The system is based on the combination of a static and a sequential SVM classifier. A Gaussian dynamic time warping based kernel is used in the sequential classifier. The system is validated on a large dataset of EEG recordings from 17 neonates. The obtained results show an increase in the detection rate at very low false detections per hour, particularly achieving a 12% improvement in the detection of short seizure events over the static RBF kernel based system.