Predictive modelling of long-term adult social care linking multiple social care and multiple healthcare dataset
No Thumbnail Available
Authors
Dr. Tha Han
Steven Wambua
Anthony Wakhisi
Krishnarajah Nirantharakumar
Contact
Check for full-text access
Issue Date
06-May-26
Type
Conference Abstract
Language
Keywords
Neighbourhood health & place-based working , population health management, planning and prevention
Alternative Title
Abstract
Background
Havering Council and King's College London used a linked dataset of routine data to produce an accurate predictive modelling for service planning and risk identification for prevention of long-term social care need.
Objectives
To develop and validate prediction models to estimate the 1-year risk of a long-term adult social care package.
METHODS
Study design: Retrospective population-based prognostic modelling study using Cox proportional hazards regression and validation.
Population: Adults registered in linked primary care, hospital, prescribing, and adult social care datasets in Havering in 2021/22 to 2023/24.
Analysis: Cox proportional hazards models estimated adjusted hazard ratios (HRs). All predictors were measured using information prior to the index date for each cohort year. The primary outcome was initiation of a long-term adult social care package. Factors were entered into multivariable models to obtain adjusted estimates.
RESULTS
Among 206,682 adults, 1.1% initiated long-term care in a year; and in ≥55 years over subgroup (n=77,683), 2.5%. Among all adults, the strongest predictors are learning disability (adjusted HR 16.46), non-elective admission (HR 3.90 (3.43-4.44)), elective admission, multimorbidity, dementia, CNS drugs and having a carer. Increasing age increases risk and living in the least deprived areas was protective. Among adults ≥55 years, similar findings are found, yet with higher risk for non-elective admission (HR 5.12 (4.34-6.03)) plus males were of lower risk.
Optimism-adjusted C-statistic was 0.936 in development and 0.937 in validation. At thresholds approximating the observed 1-year event incidence, the model showed good classification performance, with sensitivity 65.6%, specificity 96.2%, and PPV 3.6% in the adult cohort, and sensitivity 57.3%, specificity 95.0%, and PPV 5.4% in the ≥55 cohort, with similar patterns observed in the corresponding validation cohorts.
CONCLUSIONS
Our prediction models using routinely collected linked data accurately estimated 1-year risk of initiating long-term adult social care. The models showed excellent discrimination, acceptable calibration, and clinical utility, supporting their potential to inform proactive care planning and resource allocation.
RECOMMENDATIONS
Further research should refine the models, as well as assess the impact of their applications. In addition, we encourage the DHSC to conduct similar analyses to derive an evidence-based funding formula and inform national care service.