RECAST-DEMENTIA: A REGISTER-BASED ECONOMIC COST AND SEQUENCE TRANSFORMER FOR DEMENTIA CARE COST FORECASTING USING SWEDISH REGISTER DATA
Author(s)
Yassine Turki, MSc1, Sandar Aye, MSc, PhD1, Anders Wimo, MD, PhD1, Ron Handels, PhD2, Xin Xia, MSc, PhD1, Yohannes Mengistu Balcha, MD, MSc1, Emil Aho, MSc1, Linus Jönsson, MSc, PhD, MD3.
1Karolinska Institutet, Stockholm, Sweden, 2Maastricht University, maastricht, Netherlands, 3Karolinska Insitutet, Stockholm, Sweden.
1Karolinska Institutet, Stockholm, Sweden, 2Maastricht University, maastricht, Netherlands, 3Karolinska Insitutet, Stockholm, Sweden.
OBJECTIVES: To determine whether a time-aware transformer model applied on linked Swedish registers improves forecasting of 12-month post-index dementia care costs beyond conventional regression-based models.
METHODS: National healthcare, prescription and social care records were linked and converted to dated event sequences. RECAST-Dementia applies a sequence transformer with a gamma cost head to generate one person level mean forecast from pre-index history alone. Internal validation included 81,978 individuals alive on 1 January 2024 with at least 1 year of pre-index data. The outcome was censored register-priced cost summed from index until 12 months, death, or end of observation. The comparator was a Gamma generalized linear model with binary indicators for pre-index event occurrence. Accuracy was summarised by mean absolute error, cohort mean bias with 95% confidence intervals, coefficient of determination, and Spearman rank correlation.
RESULTS: Mean observed post-index cost was 120.5 kSEK. RECAST-Dementia achieved mean absolute error 50.9 kSEK, mean bias −0.5 kSEK (95% CI −1.4 to 0.4), coefficient of determination 0.764, and Spearman correlation 0.712; mean predicted cost was 120.0 kSEK. The Gamma GLM achieved mean absolute error 119.9 kSEK, mean bias +24.1 kSEK (95% CI 21.3 to 26.9), and Spearman correlation 0.629. Observed costs were highly concentrated: the highest-spending 10% accounted for 74.4% of total spending.
CONCLUSIONS: Time-aware transformer modelling improved absolute and rank-based forecast accuracy relative to a pre-specified non-sequential Gamma comparator in this Swedish internal validation cohort. Results may inform budget forecasting and identification of costly trajectories, but require external validation before broader policy application and do not establish cost-effectiveness or causal treatment effects.
METHODS: National healthcare, prescription and social care records were linked and converted to dated event sequences. RECAST-Dementia applies a sequence transformer with a gamma cost head to generate one person level mean forecast from pre-index history alone. Internal validation included 81,978 individuals alive on 1 January 2024 with at least 1 year of pre-index data. The outcome was censored register-priced cost summed from index until 12 months, death, or end of observation. The comparator was a Gamma generalized linear model with binary indicators for pre-index event occurrence. Accuracy was summarised by mean absolute error, cohort mean bias with 95% confidence intervals, coefficient of determination, and Spearman rank correlation.
RESULTS: Mean observed post-index cost was 120.5 kSEK. RECAST-Dementia achieved mean absolute error 50.9 kSEK, mean bias −0.5 kSEK (95% CI −1.4 to 0.4), coefficient of determination 0.764, and Spearman correlation 0.712; mean predicted cost was 120.0 kSEK. The Gamma GLM achieved mean absolute error 119.9 kSEK, mean bias +24.1 kSEK (95% CI 21.3 to 26.9), and Spearman correlation 0.629. Observed costs were highly concentrated: the highest-spending 10% accounted for 74.4% of total spending.
CONCLUSIONS: Time-aware transformer modelling improved absolute and rank-based forecast accuracy relative to a pre-specified non-sequential Gamma comparator in this Swedish internal validation cohort. Results may inform budget forecasting and identification of costly trajectories, but require external validation before broader policy application and do not establish cost-effectiveness or causal treatment effects.
Conference/Value in Health Info
2026-11, ISPOR Europe 2026, Vienna, Austria
Value in Health, Volume 29, Issue 12S
Code
EE670
Topic
Economic Evaluation, Methodological & Statistical Research
Topic Subcategory
Cost/Cost of Illness/Resource Use Studies
Disease
Geriatrics, Neurological Disorders