Work

Modifiable risk factors in administrative health data: validity of smoking identification algorithms and patterns of social determinants of health documentation

Year: 2025
Type: article
Abstract: Introduction: Modifiable risk factors such as smoking and adverse psychosocial circumstances contribute significantly to chronic health conditions such as hypertension. Administrative health data (AHD), including hospital discharge abstracts and physician billing claims, can be used to identify these risk factors and monitor disease burden. However, since AHD is primarily intended for healthcare management, validating algorithms that identify modifiable risk factors is crucial to ensure accuracy when repurposing for research or surveillance. This research aimed to evaluate the validity of using AHD to ascertain modifiable risk factors. The specific objectives were to: (1) synthesize evidence on the validity of electronic health data for identifying smoking status, (2) develop and assess model-based algorithms (MBAs) to identify smoking status using AHD, and (3) examine patterns in the documentation of social determinants of health (SDOH) in AHD to assess their face validity. Methods: Four studies were conducted. The first was a systematic review and meta-analysis that assessed the validity of electronic medical records (EMR) and AHD for capturing smoking status in observational studies. The remaining three were retrospective cohort studies using AHD and clinical registry data from the Manitoba Centre for Health Policy. The second study evaluated the validity of Least Absolute Shrinkage and Selection Operator and Random Forest machine learning (ML) MBAs for identifying current smoking. The third study examined whether synthetic data-based MBA training data augmentation could enhance diagnostic algorithm performance, using three data generation techniques: Synthetic Minority Oversampling Technique (SMOTE), Conditional Tabular Generative Adversarial Networks (CTGAN), and Classification and Regression Trees. The fourth study assessed the documentation of SDOH to evaluate face validity. Results: ML MBAs, especially those using EMR, outperformed rule-based algorithms but often had low sensitivity. ML improved sensitivity, although the positive predictive value was lower. Data augmentation, particularly with SMOTE, enhanced specificity, and CTGAN improved sensitivity for underrepresented groups. SDOH documentation varied across demographics, with shifts linked to policy changes and events like COVID-19. Significance: This research validates smoking identification algorithms for AHD, demonstrates how ML and synthetic data can enhance performance metrics, and highlights variability in SDOH documentation, informing risk prediction models and public health surveillance.
Language: English
Open Access status: green
OpenAlex ID: W7137790858
FWCI:
Cites: 229
Cited by:
Related to: 7