Research at a Glance
Varying Coefficient Models
Varying coefficient modeling has proven instrumental in the analysis of administrative claims, disease registries, and electronic health records. The increasing volume and complexity of data have the potential to vastly improve our understanding of important real-world problems but pose profound challenges to existing methods and implementations. To address these challenges, we have developed a series of novel statistical and computational tools targeted at large-scale analytical needs (Wu et al., 2022; Luo et al., 2023; Wu et al., 2023). These endeavors synthesize state-of-the-art techniques in statistics, computational science, and optimization, and are motivated by problems in kidney disease, cancer survival, and the COVID-19 pandemic.
Provider profiling is the procedure of comparing the performance of health care providers based on patient-centered outcomes, identifying providers with unusual performance, and developing targeted strategies for performance improvement. This high-stakes procedure has been routinely implemented by the U.S. Center for Medicare and Medicaid Services and major healthcare systems in a variety of programs to promote cost-effective health services; providers with unsatisfactory performance could experience revenue reduction or license decertification as a penalty. Our research thus far (Wu et al., 2022a, 2022b; Wu et al., 2023) has provided valuable insights into a plethora of computational and statistical problems in the methodology and practice of provider profiling.
Competing Risk Models
Competing risks, where the occurrence of one type precludes the occurrence of another, are omnipresent in health care contexts. In cancer studies, the probability of cancer-induced death is the primary focus, while the possibility of non-cancer death needs to be considered as it affects the occurrence of non-cancer death. When rehospitalization within 30 days of discharge is of chief interest, the occurrence of death should be appropriately accounted for, since an early death prevents the observation of subsequent hospital readmissions. Within the framework of competing risks, we have developed computational and statistical methods for etiological studies of breast cancer and kidney dialysis facility monitoring (Wu et al., 2022a, 2022b).
Computing and Optimization
The implementation of varying coefficient, provider profiling, and competing risk models on large-scale disease registries and administrative claims often poses significant computational challenges. To address these challenges, I have leveraged cutting-edge techniques in high-performance computing and convex optimization. For example, a proximal Newton-type method has been used to ameliorate numerical instability resulting from the inclusion of extremely distributed binary covariates in the time-varying coefficient model for competing risks, and a shared-memory parallel computing scheme has been devised to boost computational efficiency in the presence of millions of subjects (Wu et al., 2022a); a divide-and-conquer algorithm has been used to profile thousands of dialysis facilities based on emergency department visits (Wu et al., 2022b).
Kidney Disease and Transplant
I have worked with clinicians on various projects on clinical nephrology, including acute kidney injury (AKI) requiring dialysis versus non-AKI incident dialysis (Dahlerus et al., 2021), evaluating vascular surgeons based on arteriovenous fistula success (Shahinian et al., 2020), COVID-19 in-hospital and post-discharge outcomes of dialysis patients (Wu et al., 2022; Ding et al., 2022), and the survival benefit of repeat kidney transplantation (Sandal et al., 2023a, 2023b).
Alzheimer's and Dementia
My health services research on Alzheimer's disease and related dementias has been focused on leveraging natural language processing to extract information about caregiving and social determinants of health from electronic health records, and on using machine learning to predict preventable adverse events, e.g., unplanned hospital readmissions (Mahmoudi et al., 2022a, 2022b; Wu et al., 2023). In addition, I am interested in harnessing deep learning techniques to profile health care providers, with the goal of alleviating racial and ethnic disparities in the care delivery of Alzheimer's disease and related dementias.