Machine learning for pediatric healthcare
Date
2025
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
Recent ML advances have shown significant promise in assisting clinical decision-making in healthcare. In particular, the integration of ML predictive models into clinical settings has the potential to transform patient care through more personalized, timely, and effective interventions. Despite extensive research on ML for health in adults, there is a research gap regarding its application within pediatric populations. Children are not just "small adults"; they differ from adults in many ways that could affect the performance of ML models. These differences underscore the importance of tailoring ML approaches to pediatric care. ☐ In this work, we focus on two major pediatric health issues—obesity and sleep apnea—and explore ML applications in these domains. First, we propose a multi-task deep learning model that predicts BMI trajectory and attrition patterns among weight management program participants using EHR data. Second, we target sleep apnea detection using polysomnography (PSG) signals in children. We present a transformer-based architecture to detect apnea using multimodal biosignals. Subsequently, we investigate the contribution of each signal for apnea detection and identify a subset of signals that achieves a performance comparable to the complete set. We also study the effect of missing and noisy signals (a phenomenon commonly observed in pediatric settings) in apnea detection and present a tailored ML architecture and training algorithm that yields a robust apnea detector model. ☐ In this talk, we additionally cover two recent projects in more detail. The first one relates to extending the obesity case study above, where we develop an end-to-end pipeline following FHIR (Fast Healthcare Interoperability Resources) standards to predict patients' risk of obesity at different future ages. We also use a rigorous uncertainty quantification method to measure the uncertainty of our prediction. Our ML pipeline targets maximizing interoperability across various EHR formats and designing a user (provider)-centric UI to enable effective integration of our tools into clinical practice. ☐ Recognizing the growing influence of LLMs in healthcare, in the second project, we aim to address the need for scalable evaluation of fairness in LLMs. We propose a greatly customized RAG-based pipeline to generate synthetic red-teaming patient scenarios (vignettes) to evaluate social biases in medical LLMs. We merge an auxiliary generator LLM with medical knowledge graphs and a post hoc fact checker module in our design to minimize hallucination patterns when using an LLM-based pipeline and to follow an evidence-based scenario generation approach.
Description
Keywords
Machine learning, Obesity, Electronic health records, Large language models