AUTHOR=Abakasanga Emeka , Kousovista Rania , Cosma Georgina , Akbari Ashley , Zaccardi Francesco , Kaur Navjot , Fitt Danielle , Jun Gyuchan Thomas , Kiani Reza , Gangadharan Satheesh 

TITLE=Equitable hospital length of stay prediction for patients with learning disabilities and multiple long-term conditions using machine learning

JOURNAL=Frontiers in Digital Health

VOLUME=Volume 7 - 2025

YEAR=2025

URL=https://www.frontiersin.org/journals/digital-health/articles/10.3389/fdgth.2025.1538793

DOI=10.3389/fdgth.2025.1538793

ISSN=2673-253X

ABSTRACT=PurposeIndividuals with learning disabilities (LD) often face higher rates of premature mortality and prolonged hospital stays compared to the general population. Predicting the length of stay (LOS) for patients with LD and multiple long-term conditions (MLTCs) is critical for improving patient care and optimising medical resource allocation. However, there is limited research on the application of machine learning (ML) models to this population. Furthermore, approaches designed for the general population often lack generalisability and fairness, particularly when applied across sensitive groups within their cohort.MethodThis study analyses hospitalisations of 9,618 patients with LD in Wales using electronic health records (EHR) from the SAIL Databank. A Random Forest (RF) ML model was developed to predict hospital LOS, incorporating demographics, medication history, lifestyle factors, and 39 long-term conditions. To address fairness concerns, two bias mitigation techniques were applied: a post-processing threshold optimiser and an in-processing reductions method using an exponentiated gradient. These methods aimed to minimise performance discrepancies across ethnic groups while ensuring robust model performance.ResultsThe RF model outperformed other state-of-the-art models, achieving an area under the curve of 0.759 for males and 0.756 for females, a false negative rate of 0.224 for males and 0.229 for females, and a balanced accuracy of 0.690 for males and 0.689 for females. Bias mitigation algorithms reduced disparities in prediction performance across ethnic groups, with the threshold optimiser yielding the most notable improvements. Performance metrics, including false positive rate and balanced accuracy, showed significant enhancements in fairness for the male cohort.ConclusionThis study demonstrates the feasibility of applying ML models to predict LOS for patients with LD and MLTCs, while addressing fairness through bias mitigation techniques. The findings highlight the potential for equitable healthcare predictions using EHR data, paving the way for improved clinical decision-making and resource management.