PP-13 A systematic review of current prognostic tools used to estimate survival of people living with advanced dementia

Introduction The advanced stages of dementia are characterised by high complexity and symptom burden. Although dementia is a leading cause of death, its terminal nature is often under-recognised, delaying timely palliative care. Prognostic tools that estimate survival could support care planning and decision-making; however, there is limited consensus on which models are reliable for advanced dementia. Aims To identify, review, and assess the accuracy of prognostic models developed to estimate survival in people with advanced dementia. Methods A systematic search of Medline, Embase, CINAHL, PsycINFO and grey literature (from inception to April 2025) to identify studies that developed or validated survival prediction models for adults living with advanced dementia. Methodological risk of bias was appraised using the Prediction model study Risk of Bias Tool (PROBAST) and certainty of evidence was assessed using GRADE guidelines. Results Six studies were included, identifying six prognostic tools: the Palliative Support Dementia Model (PalS-DEM), Prognostic Model for Advanced Dementia (PRO-MADE), Advanced Dementia Prognostic Tool (ADEPT), Carey Index, Charlson Comorbidity Index (CCI), and MDS Risk Score. Studies varied substantially in settings, populations, prediction horizons, and methodological quality, preventing pooled or ranked performance estimates. Instead, we present a structured comparison of each tool, its predictors, and validation outcomes. Common predictors included age, sex, comorbidities, functional dependence, and nutritional status. Prediction horizons ranged from six months to three years. Conclusion Research within this field is early and inconsistent, therefore there is no prognostic tool can yet be recommended for routine use in advanced dementia. Impact Accurate survival estimates could empower clinicians, patients, and families to make care decisions aligned with patient preferences. Future research should prioritise multi-site validation, robust handling of missing data, full calibration reporting, and decision-focused evaluation. Our narrative summary supports the development of more generalisable models.