Kwok, T'ng Chang
(2024)
Utilising routine healthcare data and artificial intelligence to support personalised neonatal care and research.
PhD thesis, University of Nottingham.
Abstract
Introduction
Bronchopulmonary dysplasia (BPD) is a unique condition of abnormal lung development in premature infants that leads to long-term respiratory disease. Treatments such as postnatal dexamethasone (PND) can prevent BPD but have significant side effects. A “one size fits all” treatment plan in preterm infants, based on a small number of perinatal characteristics, is no longer justifiable. A more personalised approach is needed to identify the right infants for the right treatment. Artificial intelligence (AI) could support this by analysing patterns within large healthcare datasets providing personalised risk stratification, aiding clinical decision-making and research recruitment.
Aim
To develop the PRemature Outcome Infant Risk (PRIOR) tool to predict BPD and death using routinely recorded electronic patient data and AI.
Method
The methodologies used broadly covered two key areas. The first area demonstrated the need for the PRIOR tool via epidemiological studies and an electronic survey. Routinely recorded healthcare data of premature infants born below 32 weeks of gestational age (GA) between 2010 to 2020 in England and Wales were extracted from the National Neonatal Research Database (NNRD). The trend for respiratory morbidity was assessed using the chi-squared test for trend. An electronic Preterm Outcome Dashboard (POD) was developed to visualise the respiratory outcomes. The association of the timing of commencing PND on respiratory morbidity was assessed using logistic regression (LR) and propensity score weighting. The perception of neonatal mortality and respiratory morbidity risk among neonatal healthcare professionals was assessed using 12 clinical scenarios in an electronic survey.
The second area described the workstreams needed to develop the PRIOR tool. A systematic review was performed to comprehensively assess BPD prediction models and identify relevant predictors for the PRIOR tool. As preliminary work, LR and three other AI approaches of feedforward neural network (NN), Long Short-Term Memory (LSTM) and adaptive neuro fuzzy inference system (ANFIS) were used to predict mortality before neonatal discharge using nine perinatal factors in infants born below 28 weeks GA. Finally, LR and random forest approaches were used to develop the PRIOR tool using 25 predictors. Model performances were assessed based on their discrimination (area under the receiver operating characteristic curve (AUROC)), calibration (calibration plot) and utility (decision curve analysis) measures.
Results
4,947,596 daily data from 84,440 infants were extracted from the NNRD. Over the 11 years, BPD rates (28% to 33%) and PND use (4% to 6%) increased with improving mortality (10% to 9%) (all p<0.0001). Late PND use beyond six weeks of age was associated with higher odds of developing severe BPD or death than its earlier use in the second or third weeks of life with an adjusted odds ratio (adjOR) of 1.68 (95% confidence interval 1.28 to 2.21) in the propensity score weighting approach. For the LR analysis, the adjORs for each week of PND use beyond six weeks of age ranged between 1.75 to 4.90 when compared to PND use at two weeks of age. The electronic survey from 54 neonatal healthcare professionals found a wide variation in the perceived mortality and respiratory morbidity risks in the scenarios of infants born below 28 weeks GA. When compared to the large cohort data, 65% of the 54 responses underestimated the BPD risk.
The systematic review identified 53 BPD prediction models using a median of five predictors. The four models developed as part of the preliminary work to predict mortality before discharge had similar performance with AUROC of 0.75 to 0.76, good calibration and superior net benefit across the threshold probabilities of 20% to 60%. The best model found for the prediction of respiratory outcomes in the PRIOR tool was the parsimonious LR model using the six predictors of GA at birth, birth weight centile, sex, respiratory support type, off invasive ventilation for at least 48 hours and inotrope use. The parsimonious LR model was the simplest model and most easily interpreted while achieving similar good model performances with AUROC of 0.86 to 0.91, good calibration and superior net benefit across the threshold probabilities of above 10%.
Discussion
A more objective approach to assessing the risk of poor respiratory outcomes is needed to identify high-risk infants for treatments such as PND in a timely manner. The PRIOR tool has demonstrated the potential to support personalised neonatal care and research by applying AI to routinely collected healthcare datasets. Further work is needed to explore the application of the PRIOR tool in clinical practice.
The email addresses for the public to request access to the thesis during the embargo period are tkwok@nhs.net and don.sharkey@nottingham.ac.uk.
Actions (Archive Staff Only)
 |
Edit View |
Tools
Tools
![[thumbnail of Corrections (clean version)]](https://eprints.nottingham.ac.uk/style/images/fileicons/application_pdf.png "Corrections (clean version)")