Anticipating health and care trajectories from routinely collected social care records

Abstract

Predictive modelling in healthcare has advanced rapidly, yet social care systems, despite their central role in supporting vulnerable populations, remain underexplored in this domain. In this study, we apply machine learning to a large, pseudonymised dataset of social care records from 27,590 adults in Oxfordshire, encompassing around 90% of individuals receiving care in the region. We developed models to predict three outcomes of interest: future care plan needs, hospital admissions, and all-cause mortality, evaluated across multiple prediction horizons. Our results show that hospital admission and mortality can be predicted with meaningful discriminative performance (AUROC up to 0.893), while care plan needs are more variable and harder to predict. Post-hoc exploratory analyses revealed distinct risk signatures across outcomes, and temporal evaluation showed that care needs often emerge earlier than clinical deterioration. These findings demonstrate the feasibility of social care-based risk modelling and suggest new opportunities to improve anticipatory care planning, health management, and resource allocation using routinely collected non-clinical data.