Discordant pair analysis for sample efficient model evaluation.

We present a new technique for assessing the effectiveness of a classification algorithm using discordant pair analysis. This method utilizes a known performance baseline algorithm and a large unlabeled dataset with an assumed class distribution to obtain overall performance estimates by only assessing the subset of examples that the algorithms classify discordantly. Our approach offers an efficient way to evaluate the performance of an algorithm that minimizes the human adjudications needed while also maintaining precision in the evaluation and in some cases improving the evaluation quality by reducing human adjudication errors. This approach is a computationally efficient alternative to the traditional exhaustive method of performance evaluation and has the potential to improve the accuracy of performance estimates. Simulation studies show that the discordant pair method reduces the number of adjudications by over 90%, while maintaining the same level of sensitivity and specificity.

Impact of COVID-19 pandemic on the size of US transplant waiting lists.

BACKGROUND: More patients are waitlisted for solid organs than transplants are performed each year. The COVID-19 pandemic immediately increased waitlist mortality and decreased transplants and listings. METHODS: To calculate the number of candidate listings after the pandemic began and short-term changes that may affect waiting time, we conducted a Scientific Registry of Transplant Recipients surveillance study from January 1, 2012 to February 28, 2021. RESULTS: The number of candidates on the liver waitlist continued a steady decline that began before the pandemic. Numbers of candidates on the kidney, heart, and lung waitlists decreased dramatically. More than 3000 fewer candidates were awaiting a kidney transplant on March 7, 2021, than on March 8, 2020. Listings and removals decreased for each solid organ beginning in March 2020. The number of heart and lung listings returned to equal or above that of removals. Listings for kidney transplant, which is often less urgent than heart and lung transplant, remain below numbers of removals. Removals due to transplant decreased for all organs, while removals due to death increased for only kidneys. CONCLUSIONS: We found no evidence of the predicted surge in listings for solid organ transplant with a plateau or control of the pandemic.

Outcomes of living liver donor candidate evaluations in the Living Donor Collective pilot registry.

BACKGROUND: To gather information on long-term outcomes after living donation, the Scientific Registry of Transplant Recipients (SRTR) conducted a pilot on the feasibility of establishing a comprehensive donor candidate registry. METHODS: A convenience sample of 6 US living liver donor programs evaluated 398 consecutive donor candidates in 2018, ending with the March 12, 2020, COVID-19 emergency. RESULTS: For 333/398 (83.7%), the donor or program decided whether to donate; 166/333 (49.8%) were approved, and 167/333 (50.2%) were not or opted out. Approval rates varied by program, from 27.0% to 63.3% (median, 46%; intraquartile range, 37.3-51.1%). Of those approved, 90.4% were white, 57.2% were women, 83.1% were < 50 years, and 85.5% had more than a high school education. Of 167 candidates, 131 (78.4%) were not approved or opted out because of: medical risk (10.7%); chronic liver disease risk (11.5%); psychosocial reasons (5.3%); candidate declined (6.1%); anatomical reasons increasing recipient risk (26.0%); recipient-related reasons (33.6%); finances (1.5%); or other (5.3%). CONCLUSIONS: A comprehensive national registry is feasible and necessary to better understand candidate selection and long-term outcomes. As a result, the US Health Resources and Services Administration asked SRTR to expand the pilot to include all US living donor programs.

Outcomes of Living Kidney Donor Candidate Evaluations in the Living Donor Collective Pilot Registry.

BACKGROUND: Gaps in our knowledge of long-term outcomes affect decision making for potential living kidney donors. METHODS: The Scientific Registry of Transplant Recipients was asked to determine the feasibility of a candidate registry. RESULTS: Ten living kidney donor programs evaluated 2107 consecutive kidney donor candidates; 2099 of 2107 (99.6%) completed evaluations, 1578 of 2099 (75.2%) had a decision, and 790 of 1578 (50.1%) were approved to donate as of March 12, 2020. By logistic regression, candidates most likely to be approved were married or had attended college or technical school; those least likely to be approved had ≥1 of the following characteristics: Black race, history of cigarette smoking, and higher blood pressure, higher triglycerides, or higher urine albumin-to-creatinine ratios. Reasons for 617 candidates not being approved included medical issues other than chronic kidney disease risk (25.3%), chronic kidney disease risk (18.5%), candidate withdrawal (15.2%), recipient reason (13.6%), anatomical risk to the recipient (10.3%), noneconomic psychosocial (10.3%), economic (0.5%), and other reasons (6.4%). CONCLUSIONS: These results suggest that a comprehensive living donor registry is both feasible and necessary to assess long-term outcomes that may inform decision making for future living donor candidates. There may be socioeconomic barriers to donation that require more granular identification so that active measures can address inequities. Some candidates who did not donate may be suitable controls for discerning the appropriateness of acceptance decisions and the long-term outcomes attributable to donation. We anticipate that these issues will be better identified with modifications to the data collection and expansion of the registry to all centers over the next several years.

Mortality among solid organ waitlist candidates during COVID-19 in the United States.

We examined the effects of COVID-19 on solid organ waiting list mortality in the United States and compared effects across patient demographics (e.g., race, age, and sex) and donation service areas. Three separate piecewise exponential survival models estimated for each solid organ the overall, demographic-specific, and donation service area-specific differences in the hazard of waitlist mortality before and after the national emergency declaration on March 13, 2020. Kidney waiting list mortality was higher after than before the national emergency (adjusted hazard ratio [aHR], 1.37; 95% CI, 1.23-1.52). The hazard of waitlist mortality was not significantly different before and after COVID-19 for liver (aHR, 0.94), pancreas (aHR, 1.01), lung (aHR, 1.00), and heart (aHR, 0.94). Kidney candidates had notable variability in differences across donation service areas (aHRs, New York City, 2.52; New Jersey, 1.84; and Michigan, 1.56). The only demographic group with increased waiting list mortality were Blacks versus Whites (aHR, 1.41; 95% CI, 1.07-1.86) for kidney candidates. The first 10 weeks after the declaration of a national emergency had a heterogeneous effect on waitlist mortality rate, varying by geography and ethnicity. This heterogeneity will complicate comparisons of transplant program performance during COVID-19.

The Centers for Medicare and Medicaid Services' proposed metrics for recertification of organ procurement organizations: Evaluation by the Scientific Registry of Transplant Recipients.

On December 23, 2019, the US Centers for Medicare and Medicaid Services proposed 2 new standards that organ procurement organizations (OPOs) must meet for recertification. An OPO's organ donation rate (deceased donors/potential donors) and organ transplant rate (organs transplanted/potential donors) must not fall significantly below the 75th percentile for rates among all OPOs. We examined how OPOs would have fared under the proposed performance standards in 2016-2017. Data on donors and transplants were from the Organ Procurement and Transplantation Network; donor potential was estimated from Detailed Multiple Cause of Death data collected by the Centers for Disease Control and Prevention. In 2017, 31 (53%) OPOs failed to meet the proposed donation rate standard, 36 (62%) failed to meet the proposed organ transplant rate standard, and 37 (64%) failed at least 1 standard. We found that adjusting for age, race, and Hispanic ethnicity altered the evaluation: 8 OPOs changed their pass/fail status for the donation rate and 5 for the proposed organ transplant rate standard. We conclude that the proposed new standards may result in over half of OPOs facing decertification, and risk adjustment suggests that underlying characteristics of deaths vary regionally such that decertification decisions may be affected.

Fast, fully Bayesian spatiotemporal inference for fMRI data.

We propose a spatial Bayesian variable selection method for detecting blood oxygenation level dependent activation in functional magnetic resonance imaging (fMRI) data. Typical fMRI experiments generate large datasets that exhibit complex spatial and temporal dependence. Fitting a full statistical model to such data can be so computationally burdensome that many practitioners resort to fitting oversimplified models, which can lead to lower quality inference. We develop a full statistical model that permits efficient computation. Our approach eases the computational burden in two ways. We partition the brain into 3D parcels, and fit our model to the parcels in parallel. Voxel-level activation within each parcel is modeled as regressions located on a lattice. Regressors represent the magnitude of change in blood oxygenation in response to a stimulus, while a latent indicator for each regressor represents whether the change is zero or non-zero. A sparse spatial generalized linear mixed model captures the spatial dependence among indicator variables within a parcel and for a given stimulus. The sparse SGLMM permits considerably more efficient computation than does the spatial model typically employed in fMRI. Through simulation we show that our parcellation scheme performs well in various realistic scenarios. Importantly, indicator variables on the boundary between parcels do not exhibit edge effects. We conclude by applying our methodology to data from a task-based fMRI experiment.

Impaired Bottom-Up Effective Connectivity Between Amygdala and Subgenual Anterior Cingulate Cortex in Unmedicated Adolescents with Major Depression: Results from a Dynamic Causal Modeling Analysis.

Major depressive disorder (MDD) is a significant contributor to lifetime disability and frequently emerges in adolescence, yet little is known about the neural mechanisms of MDD in adolescents. Dynamic causal modeling (DCM) analysis is an innovative tool that can shed light on neural network abnormalities. A DCM analysis was conducted to test several frontolimbic effective connectivity models in 27 adolescents with MDD and 21 healthy adolescents. The best neural model for each person was identified using Bayesian model selection. The findings revealed that the two adolescent groups fit similar optimal neural models. The best across-groups model was then used to infer upon both within-group and between-group tests of intrinsic and modulation parameters of the network connections. First, for model validation, within-group tests revealed robust evidence for bottom-up connectivity, but less evidence for strong top-down connectivity in both groups. Second, we tested for differences between groups on the validated parameters of the best model. This revealed that adolescents with MDD had significantly weaker bottom-up connectivity in one pathway, from amygdala to sgACC (p=0.008), than healthy controls. This study provides the first examination of effective connectivity using DCM within neural circuitry implicated in emotion processing in adolescents with MDD. These findings aid in advancing understanding the neurobiology of early-onset MDD during adolescence and have implications for future research investigating how effective connectivity changes across contexts, with development, over the course of the disease, and after intervention.

A Naive Bayes machine learning approach to risk prediction using censored, time-to-event data.

Predicting an individual's risk of experiencing a future clinical outcome is a statistical task with important consequences for both practicing clinicians and public health experts. Modern observational databases such as electronic health records provide an alternative to the longitudinal cohort studies traditionally used to construct risk models, bringing with them both opportunities and challenges. Large sample sizes and detailed covariate histories enable the use of sophisticated machine learning techniques to uncover complex associations and interactions, but observational databases are often 'messy', with high levels of missing data and incomplete patient follow-up. In this paper, we propose an adaptation of the well-known Naive Bayes machine learning approach to time-to-event outcomes subject to censoring. We compare the predictive performance of our method with the Cox proportional hazards model which is commonly used for risk prediction in healthcare populations, and illustrate its application to prediction of cardiovascular risk using an electronic health record dataset from a large Midwest integrated healthcare system.