DETERMINANTS OF LONG-TERM OUTCOMES AFTER STROKE : INSIGHTS FROM LINKED MICHIGAN COVERDELL ACUTE STROKE REGISTRY
Data needed to provide a comprehensive assessment of long-term recovery of stroke survivors is lacking for the Michigan Acute Stroke Registry, referred to as MiSP (Michigan Stroke Program), as it is for many other US-based stroke registries. The overall objective of this dissertation is to bridge this knowledge gap by linking stroke registry data with administrative claims data to obtain follow-up data on patient outcomes. The administrative data source is the Michigan Value Collaborative (MVC) a comprehensive, statewide, claims database that includes claims from Blue Cross Blue Shield of Michigan (BCBSM) (private and Medicare Advantage plans) and Medicare fee-for-service (FFS). The linked dataset enables us to report on a wide range of patient-centered outcome measures including mortality, stroke recurrence, readmission, and discharge to rehabilitation. However, this dissertation focuses on the detailed analysis of predicting all-cause hospital readmissions and estimating the comparative effectiveness of inpatient rehabilitation facility (IRF) versus skilled nursing facility (SNF) on functional recovery using home time (number of days alive and outside of inpatient care). In the first aim we generated a linked database by combining a 5-year retrospective cohort (2016-2020) of all acute stroke discharges entered into MiSP registry from 31 stroke certified hospitals (n=46,330) with MVC (n= 30,685) claims data using both deterministic and probabilistic matching techniques. We evaluated the accuracy, completeness, and representativeness of the linkage results using pre linkage qualitative and post linkage quantitative methods. We then generated descriptive data on 30-day, 90-day and 1-year outcome event rates including mortality, all-cause hospital readmissions, stroke recurrence, use of post-acute care services (i.e., IRF, SNF, and home health), use of out-patient visits, and home time. We showed that probabilistic linkage between MiSP acute stroke registry and MVC claims data using indirect identifiers produced slightly higher linkage rates compared to deterministic linkage and that our linkage is feasible and resulted in a valid linked dataset that has acceptable representation of Medicare FFS and BCBSM insured population in Michigan. In the second aim we developed 30-day and 1-year all-cause readmission prediction models using 3 different machine learning (ML) methods: LASSO logistic regression, XGBoost and ANN, and compared the predictive performance of these methods. After identifying the optimal performing model, we report the most important predictors. Our findings demonstrated that prediction of all cause readmission can be achieved with relatively modest accuracy (AUC range 0.67 - 0.68), that LASSO regression was able to predict readmission after stroke with similar accuracy to more advanced ML methods, and that clinical features of stroke (e.g., NIHSS, stroke etiology) were less important than the burden of existing comorbidities (e.g., chronic renal failure, atrial fibrillation, heart failure) or the hospitalization (e.g., admission duration, discharge destination) in predicting post-stroke readmission, especially over longer periods of time (1-year).The third aim was to estimate the comparative effectiveness of IRF versus SNF rehabilitation care to improve functional recovery 90 days and 1 year following discharge using inverse probability of treatment weighting analysis of differences in home time and mortality. This analysis was limited to Medicare FFS stroke patients. Our findings provided further evidence that discharge to IRF versus SNF was associated with longer home time and lower mortality over 1-year post discharge. However, our sensitivity analysis illustrated that home time is heavily impacted by rehabilitation length of stay especially over 90-days and hence future studies should avoid using home time and rely on more stable measures (less prone to bias) like mRS or successful community discharge (home for >30 consecutive days). In conclusion, we illustrated that data linkage can provide needed information to describe patient recovery up to 1 year after acute stroke discharge. Future studies should expand on the generalizability of the linkage by including data from more hospitals and claims data from Medicaid and other insurance providers.
Read
- In Collections
-
Electronic Theses & Dissertations
- Copyright Status
- In Copyright
- Material Type
-
Theses
- Authors
-
Hailat, Ra'ed S.
- Thesis Advisors
-
Reeves, Mathew J.
- Committee Members
-
Oostema, John A.
Thompson, Michael P.
de los Campos, Gustavo
- Date Published
-
2024
- Subjects
-
Epidemiology
- Program of Study
-
Epidemiology - Doctor of Philosophy
- Degree Level
-
Doctoral
- Language
-
English
- Pages
- 265 pages
- Permalink
- https://doi.org/doi:10.25335/mrg9-c465