[MA 2026 01] Creating an explainable model for the prediction of bronchopulmonary dysplasia in preterm infants.
Emma Children’s Hospital – Follow Me health care innovation program / Emma Neuroscience Group
Proposed by: Frank Bennis [f.c.bennis@amsterdamumc.nl]
Introduction
Emma Neuroscience Group is a group which investigates the impact of disease and treatment on the brain of children and young adults in the context of daily life.
Follow Me is an ambitious program that develops and implements structured outpatient care for all tertiary care patients at the Emma Children’s Hospital of Amsterdam UMC with the ambition to improve clinical follow-up, support data-driven health care evaluation, facilitate clinical research to improve clinical care, and educate the future generation of health care professionals.
Bronchopulmonary dysplasia (BPD) is a serious complication of preterm birth [1]. Systemic corticosteroids can reduce BPD risk but carry significant adverse effects, necessitating targeted treatment for high-risk infants [2]. Early prediction of BPD is therefore crucial, yet only 25% of affected patients have received corticosteroids in the past 15 years. To this end, a BPD prediction model has been developed based on vital sign data of the first 7 days after birth, combined with data available at birth [3]. As sample frequency is once per minute, an autoencoder was used for the vital signs to reduce the dimensionality, after which an LSTM was trained on the vital signs, and combined in the final layer with a neural network trained on the data available at birth. The resulting model showed high predictive performance (AUC=0.83) and has been implemented in EPIC and provides real time predictions of BPD in preterm born children admitted to the NICU.
Description of the SRP Project/Problem
Although the model showed high predictive performance, the model itself is not explainable. Non-explainable predictions models are known to receive low adoption rates in the clinical setting. Therefore, our ambition is to develop a new model that provides high predictive performance of BPD and provides information on the factors contributing to the prediction. In the envisioned project, data from 2016-2023 will be used to develop the new BPD prediction model, whilst maintaining explainability. First, we envision the development of classical ML models in which the temporal component is incorporated in tabular features, after which explainability can be incorporated with tools like SHAP. Second, we envision the use of classical convolutional neural networks, combining both temporal parameters (e.g. segments from the autoencoder) and static data, after which explainability can be incorporated with tools like GRAD-CAM. Third, we aim to explore the use of attention based models such as a temporal fusion transformer, which incorporates temporal and static parameters, as well as possibilities for explainability. In these steps, model complexity increases, whilst explainability decreases, but is maintained to a certain level. Although the exact definition of explainability and especially the needed level of explainability for healthcare machine learning models is not agreed upon, it is a vast improvement upon the current model, which does not have explainability at all. The same holds for the question if attention equals interpretability, which may not be a good proxy, but an improvement upon the current situation. These models are solely possibilities, and we encourage the student to think outside the box and to propose alternative methods for explainable methods.
Research questions
RQ1) Can we develop explainable ML models, from basic logistic regression to advanced temporal fusion transformer models, to predict bronchopulmonary dysplasia in preterm infants using static data at birth combined with the first seven days of data from SpO2 and FiO2?
Expected results
- A ML model that uses vital sign data and data available at birth with explainable predictions.
- Preprocessed data from 2016-2023.
- A thesis describing the methods used and resulting model
- We offer the possibility to co-author a publication for students that fulfil the ICMJE criteria for authorship (https://www.icmje.org/recommendations/browse/roles-and-responsibilities/defining-the-role-of-authors-and-contributors.html)
Time period
X November – June
X May - November
References
[1] Jobe AH, Bancalari E. Bronchopulmonary Dysplasia. Am J Respir Crit Care Med 2001;163(7):1723–9. Available from: https://www.atsjournals.org/doi/10.1164/ajrccm.163.7.2011060
[2] van de Loo M, van Kaam A, Offringa M, Doyle LW, Cooper C, Onland W. Corticosteroids for the prevention and treatment of bronchopulmonary dysplasia: an overview of systematic reviews. Cochrane Database of Systematic Reviews 2024;2024(4):CD013271.
[3] F. Bennis, W. Onland, J. van der Vorst, M. Hoogendoorn, G. Hutten, A. van Kaam, J. Oosterlaan, M.Königs, “Prediction of Bronchopulmonary Dysplasia During the First Seven Days After Birth Using Vital Sign Data With Machine Learning Algorithms” (Submitted)