Developing an Explainable AI model for predicting adverse events post-cardiac surgery in pediatric Congenital Heart Disease (CHD) patients

Congenital heart disease (CHD) is the most common congenital disability, presenting in 1% of live births. Healthcare advances have significantly decreased CHD infant mortality, yielding over 90% survival into adulthood. However, pediatric CHD patients are at higher risk of clinical deterioration either due to their underlying pathology or due to the surgical procedure they need to undergo in order to correct their defect. In current clinical practice, clinical deterioration in hospitalized pediatric patients is detected, in most cases, when it has already occurred, leading to a high in and out-hospital morbidity, healthcare costs and caregiver burden. With proper continuous monitoring of the patients and the integration of many variables by means of artificial intelligence (AI) based algorithms, we could predict the patient's risk of deterioration before it occurs, providing critical information for immediate patient care.

This project focuses on the development of an explainable machine learning (ML) model aimed at predicting unexpected adverse events during hospital stays following cardiac surgery in pediatric patients with congenital heart diseases. The project will be performed in collaboration with Hospital Sant Joan de Déu. The ML model should prioritize explainability to ensure that healthcare professionals can understand and trust its predictions. The study involves data harmonisation and pre-processing from the perioperative period, feature selection, model training, and model validation. The model will be trained with multimodality data including: clinical, demographic, laboratory, vital signs, etc. The explainability will be reached by using representative learning with algorithms like Multi Kernel Learning (MKL) but other techniques such as SHAP values (SHapley Additive exPlanations) or LIME (Local Interpretable Model-agnostic Explanations) can be studied to provide clear insights into the factors influencing each prediction, thus facilitating better clinical decision-making and improving patient outcomes.

Supervisors: Adriana Modrego (UPF/HSJD) and Bart Bijnens (UPF)