Research

Receiver operating characteristics (ROC) curves are widely used graphical tools to evaluate the diagnostic capacity of biomarkers. For decades, numerous ROC curve methodologies (parametric, non-parametric, semi-parametric, and empirical) have been proposed. While useful, a common theme of those methodologies has been to model the biomarkers of the binary groups (say, healthy and diseased) separately and estimate the ROC curve as the function of CDFs of those groups. Eminent researchers in this area (Margaret Sullivan Pepe, Tianxi Cai) commented that these ways don't reflect the purpose of the ROC curves. On the other hand, other methodologies (for example, placement value (PV)-based method) have been proposed that model the relationship between the biomarker groups to estimate the ROC curve. This is in a way a more direct approach and the effect of covariates on the diagnostic capacity can be evaluated directly on the diagnostic accuracy measures (such as ROC, AUC, etc.).  I have been working in this area for the last several years. I have developed and applied novel statistical methodologies in maternal and neonatal health.

Few works:

• Comparing the performances of some parametric PV-based models

• A novel PV-based method to obtain gestational-age wise diagnostic accuracy measures in predicting large-for-gestational age (LGA)

• Imposing orders on ROC curves

• Imposing concavity on ROC curves

Gestational Diabetes Mellitus (GDM): Research has shown differences in fetal cardiac function between women with and without diabetes. I worked on a project that evaluates how various cardiac parameters predict abnormal glucose challenge test (GCT) results or gestational diabetes mellitus (GDM) diagnoses. Key findings include significant associations of the 1st and 2nd HbA1c, the 2nd mitral E/A ratio, and the change in mitral E/A ratios with GDM diagnosis. The 1st and 2nd HbA1c and the change in mitral E/A ratios were linked to GCT results >130 mg/dL. Combining these eight parameters in a logistic regression model enhanced predictive accuracy for GDM and GCT results >130 mg/dL compared to individual parameters..

• Loehr, F. W., Ghosal, S., Urban, A., Ennen, C., Prediction of Gestational Diabetes Mellitus using Fetal Cardiac Function Parameters: A Pilot Study. UltraCon 2023, American Institute of Ultrasound Medicine (AIUM). March 2023.

I have also briefly worked on spatiotemporal modeling. The common theme of the proposed spatiotemporal models that I have worked on is the use of spline to model spatial variability. It is extremely useful for cases where the number of spatial parameters is unusually large. Here are examples of some published and ongoing works:

• A hierarchical Bayesian model to assess the impact of socio-economic factors on highly overdispersed and zero-inflated number of medical doctors across the USA

• Assessing the antibiotic overuse across KY (in process)

• Vision impairment in VA (in process)

I also worked on few projects related to causal inference as part of my thesis particularly for ordinal outcome with number of treatments greater than 2. We extended some existing methodologies (both nonparametric and parametric) and compared the performances across variety of settings. The manuscripts are yet not published primarily due to unavailability of interesting dataset.