University Way,
London E16 2RD
About Our Research
Necrotizing enterocolitis (NEC) is the leading cause of mortality and morbidity in preterm newborns. About 30-50% of babies who require surgery for NEC do not survive. Despite 60 years of research, our understanding of the causation of NEC has not improved significantly enough to change outcomes.
The lack of progress in outcomes for NEC is compounded by a lack of consensus in terms of case definition and data sets that are contaminated and as such limited or inaccurate in the insights they can offer.
Our project wishes to apply an Artificial Intelligence (AI) system to assist in formulating a case definition of NEC, explore the possibility of different subsets of NEC and develop a clinical decision tree that can facilitate an earlier diagnosis of NEC, which can significantly improve outcomes.
Machine learning can use current data such as including an expert opinion (even when they have differing opinions), medical imaging, lab results and clinical work-up to give more prognostic and diagnostic insights into NEC. Ensemble modelling is powerful enough to capture individual presentations (micro-model – clinical, radiology and biomarkers) while deep learning identifies patterns to identify sub-patient populations (macro-models). Using Real-World Data (RWD) current NEC case definitions and tools can be tested and refined. Hence AI automation can actively advance NEC research with a long-term solution with minimal investment. This can also be adapted as a quality control tool and to give immediate insights into the current state of a neonatal unit's performance.
Agreement from experts:
What AI can do:
1
“One key issue remains to be addressed to ensure progress in understanding NEC: the definition that should be used in the context of methodologically robust research studies. The difficulty in separating NEC from similar conditions (ANID) clinically, intraoperatively and histologically are well recognised. ”
2
“Variability internationally in NEC definitions and inclusion criteria detract from the ability to make meaningful comparisons and argue for establishing standard reporting criteria for NEC studies such as have been produced in other disease areas.”
“The machine learning (PDE model) of NEC introduced in this study incorporates key mechanisms of the disease and is able to reproduce expected physiological results. We believe this model can investigate specific issues related to NEC including identifying the relative roles of factors that influence NEC and the potential of various treatment and prevention strategies. ”
“Understanding the causes of disagreement among experts in clinical decision making has been a challenge for decades. Machine learning proved to be a handy objective framework to identify important features for experts and check whether the selected features reflect the agreements/disagreements. These findings may lead to improved diagnostic accuracy and standardization among clinicians, and pave the way for the application of this methodology to other problems which present inter-expert variability.”
3
“NEC leads to poorer neurodevelopmental outcomes. Understanding both the nature and prevalence of neurodevelopmental impairment among extremely preterm infants is important because it can lead to targeted interventions that in turn may lead to improved outcomes”.
“Very preterm children are at a high risk for neurodevelopmental impairments, but there is variability in the pattern and severity of outcome. Neonatal magnetic resonance imaging (MRI) combined with machine learning (CAD) enhances the capacity for early detection of brain injury and altered brain development. This in future will assist in the early recognition of high-risk infants who warrant surveillance and early intervention.”
4
“Recent success in NEC prevention and recognition of NECsubsets demonstrate that we must refine our global NEC definition if we are to progress in our understanding of the disease. The success of emerging biomarkers, applied diagnostics like ultrasound and quality initiatives that aim to reduce NEC and other ANIDS are highly dependent on differential case definitions. Each modality promises improved capacity to pinpoint the timing, accuracy, and potentially the sub-cohort categorization of NEC”
“Since the debate continuously existed in diagnostic and prognostic prediction Abdominal radiographs (AR) and Abdominal Ultrasound (AUS) in patients with NEC. We utilised objective analysis and independent feature extraction from Machine Learning (ML) to explore the performance of radiographic and sonographic parameters to predict prognosis of NEC (i.t.o AUROC using logistic models). We found that AUS (AUROC=0,86) was significantly better than AR (AUROC=0,75) for predicting the presence of NEC. Also, AUS detected NEC parameters earlier than AR.”
“Our results indicate a wide variation in the management of NEC, with significant differences between neonatologists and pediatric surgeons. A better appreciation of the relative significance and weighting that should be applied to the clinical features and investigations should reduce the variation in interpretation that appears to exist.”
​
5
“Understanding the causes of disagreement among experts in clinical decision making has been a challenge for decades. Machine learning proved to be a handy objective framework to identify important features for experts and check whether the selected features reflect the agreements/disagreements. These findings may lead to improved diagnostic accuracy and standardization among clinicians, and pave the way for the application of this methodology to other problems which present inter-expert variability.”