Predicting Individual-Level Probabilities of Dementia and Diabetes using Health Services Administrative Data
Year: 2015
Type: dissertation
Abstract: Health services administrative data (HSAD) provide a timely and readily available population-based data source for health research and planning. To perform disease-specific analyses, individuals with the disease must first be identified in the administrative data. In this validation study, a data modeling method (logistic regression) and two algorithmic modeling methods (random forests and gradient boosted machines) were compared in predicting individual-level probabilities of diabetes and dementia using HSAD. Prevalence for both diseases was calculated to demonstrate a population-level application and A1C testing rates were calculated to demonstrate an individual-level application. Diabetes prevalence and A1C testing rates were compared with the classification-based methods currently used for the Ontario Diabetes Database (ODD). The reference standard data were chart re-abstraction data for a sample of Toronto-areas physician for diabetes and the Ontario sample of the Canadian Study of Health and Aging for dementia. The three-year random forest models were the best-performing models for both diabetes (Nagelkerke’s R2=0.82, C statistic=0.98 and calibration slope=1.02) and dementia (Nagelkere’s R2=0.58, C statistic=0.92 and calibration slope=1.00). Based on the random forest probabilities, the prevalence of dementia in Ontario in 2008/09 was 8.5 percent for females and 5.7 percent for males. The prevalence of diabetes was 5.4 percent for females and 6.5 percent for males. A1C testing rates varied from 38 to 61 percent depending on the probability cutoff used. The probability-based diabetes prevalence was lower than the ODD-based rates and the A1C testing rates were higher than the ODD-rates. Although the random forest models are more complex than the pooling algorithms currently used for the ODD, they offer researchers the flexibility to construct cohorts with sensitivities and specificities specific to each purpose. Algorithmic modeling methods are well suited to research using health services administrative data and should be pursued as an alternative to data modeling methods.
Source: TSpace (University of Toronto)
Author: Munaza Chaudhry
Language: English
Subfield: Health Information Management
Field: Health Professions
Domain: Health Sciences
Open Access status: green
OpenAlex ID: W2592489759