Section on Medical Devices and Diagnostics: Statistical Methods in Medical and Health Research

Spontaneous reporting system databases are key resources for post-marketing surveillance, providing real-world evidence (RWE) on the adverse events (AEs) of regulated drugs or other medical products. Various statistical methods have been proposed for AE signal detection in these databases, flagging drug-specific AEs with disproportionately high observed counts compared to expected counts under independence. However, signal detection remains challenging for rare AEs or newer drugs, which receive small observed and expected counts and thus suffer from reduced statistical power. Principled information sharing on signal strengths across drugs/AEs is crucial in such cases to enhance signal detection. However, existing methods typically ignore complex between-drug associations on AE signal strengths, limiting their ability to detect signals. We propose novel local-global mixture Dirichlet process (DP) prior-based nonparametric Bayesian models to capture these associations, enabling principled information sharing between drugs while balancing flexibility and shrinkage for each drug, thereby enhancing statistical power. We develop efficient Markov chain Monte Carlo algorithms for implementation and employ a false discovery rate (FDR)-controlled, false negative rate (FNR)-optimized hypothesis testing framework for AE signal detection. Extensive simulations demonstrate our methods' superior sensitivity---often surpassing existing approaches by a twofold or greater margin---while strictly controlling the FDR. An application to FDA FAERS data on statin drugs further highlights our methods' effectiveness in real-world AE signal detection. Software implementing our methods is provided as supplementary material.  

Diagnostic tests play a critical role in detecting and monitoring diseases. However, the effects of test results on health outcomes are indirect through their downstream influence on treatment decisions, posing a challenge in evaluating their true effectiveness. In this work, we propose causal estimands and targeted maximum likelihood estimation (TMLE) to evaluate the effectiveness of medical tests from their explanatory and/or pragmatic utility in medical care. We further propose an adaptive experimental design to better evaluate medical tests and devices. This framework is generally applicable to evaluate the effectiveness of any device output - e.g., artificial intelligence (Al) enabled prediction score - whose effect on the primary outcome of interest is mediated by the consequent treatment decision. The performance of our estimators and designs is demonstrated through simulation studies. 

In clinical practice, missing disease status verification is common and can bias estimators of diagnostic test accuracy. In this paper, we propose verification bias-corrected interval estimation methods for Youden index of a continuous test under the missing-at-random (MAR) assumption. Based on four estimators (FI, MSI, IPW, and SPE) introduced by Alonzo and Pepe for handling verification bias, we develop multiple confidence intervals for the Youden index by applying bootstrap resampling and the method of variance estimates recovery (MOVER). Through extensive simulation and real data studies, we find SPE estimator performs better when paired with bootstrap method. Notably, bootstrap-SPE intervals show appealing doubly robustness to the model misspecification and perform adequately across almost all scenarios considered. In contrast, FI and MSI estimators perform better when paired with MOVER method. When the disease model is correctly specified, MOVER-FI intervals achieve optimal coverage probability. We also find that when the verification proportion is low, bootstrap methods provide more accurate estimates while MOVER methods offer higher precision. 

Reference regions are invaluable in the interpretation of results of biochemical and physiological tests of patients. When there are multiple biochemical analytes measured from each subject, a multivariate reference region (MRR) is needed. MRRs are more desirable than multiple univariate reference regions because the latter has less specificity against false positives and disregards the cross-correlations between variables. In the laboratory medicine literature, there are MRRs available under multivariate normality. However, almost all laboratory test results follow a non-normal distribution. While this is true, very few procedures to compute MRRs outside a multivariate normal setting are available. For this reason, we develop MRRs in a nonparametric setting. We consider two criteria in constructing MRRs: the prediction region and the tolerance region criteria. Moreover, to make the MRRs amenable for component-wise outlier detection, which ellipsoidal regions are not capable of, we use rectangular regions. The accuracies of the proposed procedures are evaluated through coverage probabilities and expected volumes. A solution to include covariates in the model is also proposed. 

We have developed a methodology for the estimation of the prediction interval associated with Deming regression. Deming regression is a common methodology of analysis to compare the linearly correlated measurements from two methods, X and Y, over an interval[a,b]. The methodology applies to X and Y measurements whose variance is constant, or linearly, and quadratically increasing with respect to their mean over the interval. The prediction interval invokes the key term, variance of Y, which is a function of the regression parameters and the error terms, and their variances and co-variances. For each measurement variance case, we calculated the coverage of the prediction interval on data generated by simulation. We found that for the linear case, the methodology needs an adjustment k, whose function incorporates a, b, and the prediction level. The methodology applied to X and Y paired data generated by simulation shows coverage rates within 2% of the prescribed 90% and 95% prediction intervals. We present this methodology to complement those found in CLSI EP09 and EP14 for method comparison and commutability studies, respectively. 

Chronic kidney disease significantly influences outcomes in total hip and knee arthroplasty, yet its impact on total ankle arthroplasty (TAA) remains understudies. This analysis leverages statistical models to investigate the relationship between CKD progression and postoperative outcomes in TAA.

Using the American College of Surgeons National Surgical Quality Improvement Program (NSQIP) database, we identified all primary TAA cases performed between 2006-2021 in US. Univariable and Multivariable regressions were developed to evaluate the association of CKD stages with length of hospital stay, unplanned return to the operating room, and total postoperative complications. Model diagnosis was performed to confirm the robustness of the results.

675 patients were with different CKD stages. Univariable analyses reveal significant differences in anesthesia type, diabetes, postoperative, dialysis, ASA class, BUN, etc. Significant associations were found between CKD stages and prolonged hospital stays, increased odds (OR=2.18) of unplanned return to room, higher rate of complications (OR=1.85), etc. Our methods facilitated a nuanced understanding of these complex relationships. 

Effective glycemic control in Type 1 Diabetes Mellitus (T1DM) depends on understanding the complex interplay between carbohydrate intake, insulin administration, and blood glucose levels. This study investigates the causal pathways linking meal timing to postprandial glucose levels, focusing on the direct effect of carbohydrate intake and the indirect effect mediated by bolus insulin. Using the OHIO T1DM dataset, which includes continuous glucose monitoring, insulin administration, and detailed meal records, we apply Causal Mediation Analysis (CMA) to quantify these effects. We estimate the Average Causal Mediation Effect (ACME) of bolus insulin and the direct effect of carbohydrate intake at both the individual level and across different times of the day. To account for potential confounders, we fit mediator and outcome models that incorporate pre-treatment measurements and employ Inverse Probability of Treatment Weighting (IPTW) to balance covariates across time-of-day categories (morning, afternoon, evening, late evening). Finally, we assess whether these causal effects vary by time of day, providing new insights into the temporal dynamics of T1DM management.