Framework for Integrating Equity Into Machine Learning Models: A Case Study.

Predictive analytic models leveraging machine learning methods increasingly have become vital to health care organizations hoping to improve clinical outcomes and the efficiency of care delivery for all patients. Unfortunately, predictive models could harm populations that have experienced interpersonal, institutional, and structural biases. Models learn from historically collected data that could be biased. In addition, bias impacts a model's development, application, and interpretation. We present a strategy to evaluate for and mitigate biases in machine learning models that potentially could create harm. We recommend analyzing for disparities between less and more socially advantaged populations across model performance metrics (eg, accuracy, positive predictive value), patient outcomes, and resource allocation and then identify root causes of the disparities (eg, biased data, interpretation) and brainstorm solutions to address the disparities. This strategy follows the lifecycle of machine learning models in health care, namely, identifying the clinical problem, model design, data collection, model training, model validation, model deployment, and monitoring after deployment. To illustrate this approach, we use a hypothetical case of a health system developing and deploying a machine learning model to predict the risk of mortality in 6 months for patients admitted to the hospital to target a hospital's delivery of palliative care services to those with the highest mortality risk. The core ethical concepts of equity and transparency guide our proposed framework to help ensure the safe and effective use of predictive algorithms in health care to help everyone achieve their best possible health.

Hospital-Level Variation in Death for Critically Ill Patients with COVID-19.

RATIONALE: Variation in hospital mortality has been described for coronavirus disease 2019 (COVID-19), but the factors that explain these differences remain unclear. OBJECTIVE: Our objective was to utilize a large, nationally representative dataset of critically ill adults with COVID-19 to determine which factors explain mortality variability. METHODS: In this multicenter cohort study, we examined adults hospitalized in intensive care units with COVID-19 at 70 United States hospitals between March and June 2020. The primary outcome was 28-day mortality. We examined patient-level and hospital-level variables. Mixed-effects logistic regression was used to identify factors associated with interhospital variation. The median odds ratio (OR) was calculated to compare outcomes in higher- vs. lower-mortality hospitals. A gradient boosted machine algorithm was developed for individual-level mortality models. MEASUREMENTS AND MAIN RESULTS: A total of 4,019 patients were included, 1537 (38%) of whom died by 28 days. Mortality varied considerably across hospitals (0-82%). After adjustment for patient- and hospital-level domains, interhospital variation was attenuated (OR decline from 2.06 [95% CI, 1.73-2.37] to 1.22 [95% CI, 1.00-1.38]), with the greatest changes occurring with adjustment for acute physiology, socioeconomic status, and strain. For individual patients, the relative contribution of each domain to mortality risk was: acute physiology (49%), demographics and comorbidities (20%), socioeconomic status (12%), strain (9%), hospital quality (8%), and treatments (3%). CONCLUSION: There is considerable interhospital variation in mortality for critically ill patients with COVID-19, which is mostly explained by hospital-level socioeconomic status, strain, and acute physiologic differences. Individual mortality is driven mostly by patient-level factors. This article is open access and distributed under the terms of the Creative Commons Attribution Non-Commercial No Derivatives License 4.0 (http://creativecommons.org/licenses/by-nc-nd/4.0/).

Neighborhood Disadvantage and Hospital Quality Ratings in the Medicare Hospital Compare Program.

BACKGROUND: The Centers for Medicare and Medicaid Services provide nationwide hospital ratings that may influence reimbursement. These ratings do not account for the social risk of communities and may inadvertently penalize hospitals that service disadvantaged neighborhoods. OBJECTIVE: This study examines the relationship between neighborhood social risk factors (SRFs) and hospital ratings in Medicare's Hospital Compare Program. RESEARCH DESIGN: 2017 Medicare Hospital Compare ratings were linked with block group data from the 2015 American Community Survey to assess hospital ratings as a function of neighborhood SRFs. SUBJECTS: A total of 3608 Medicare-certified hospitals in 50 US states. MEASURES: Hospital summary scores and 7 quality group scores (100 percentile scale), including effectiveness of care, efficiency of care, hospital readmission, mortality, patient experience, safety of care, and timeliness of care. RESULTS: Lower hospital summary scores were associated with caring for neighborhoods with higher social risk, including a reduction in hospital score for every 10% of residents who reported dual-eligibility for Medicare/Medicaid [-3.3%; 95% confidence interval (CI), -4.7 to -2.0], no high-school diploma (-0.8%; 95% CI, -1.5 to -0.1), unemployment (-1.2%; 95% CI, -1.9 to -0.4), black race (-1.2%; 95% CI, -1.7 to -0.8), and high travel times to work (-2.5%; 95% CI, -3.3 to -1.6). Associations between neighborhood SRFs and hospital ratings were largest in the timeliness of care, patient experience, and hospital readmission groups; and smallest in the safety, efficiency, and effectiveness of care groups. CONCLUSIONS: Hospitals serving communities with higher social risk may have lower ratings because of neighborhood factors. Failing to account for neighborhood social risk in hospital rating systems may reinforce hidden disincentives to care for medically underserved areas in the United States.

Association Between Room Location and Adverse Outcomes in Hospitalized Patients.

OBJECTIVE: To investigate whether a patient's proximity to the nurse's station or ward entrance at time of admission was associated with increased risk of adverse outcomes. METHOD: We conducted a retrospective cohort study of consecutive adult inpatients to 13 medical-surgical wards at an academic hospital from 2009 to 2013. Proximity of admission room to the nurse's station and to the ward entrance was measured using Euclidean distances. Outcomes of interest include development of critical illness (defined as cardiac arrests or transfer to an intensive care unit), inhospital mortality, and increase in length of stay (LOS). RESULTS: Of the 83,635 admissions, 4,129 developed critical illness and 1,316 died. The median LOS was 3 days. After adjusting for admission severity of illness, ward, shift, and year, we found no relationship between proximity at admission to nurse's station our outcomes. However, patients admitted to end of the ward had higher risk of developing critical illness (odds ratio [ OR] = 1.15, 95% confidence interval [CI] = [1.08, 1.23]), mortality ( OR = 1.16, 95% CI [1.03, 1.33]), and a higher LOS (13-hr increase, 95% CI [10, 15] hours) compared to patients admitted closer to the ward entrance. Similar results were observed in sensitivity analyses adjusting for isolation room patients and considering patients without room transfers in the first 48 hr. CONCLUSIONS: Our study suggests that being away from the nurse's station did not increase the risk of these adverse events in ward patients, but being farther from the ward entrance was associated with increase in risk of adverse outcomes. Patient safety can be improved by recognizing this additional risk factor.

Comparison of variable selection methods for clinical predictive modeling.

OBJECTIVE: Modern machine learning-based modeling methods are increasingly applied to clinical problems. One such application is in variable selection methods for predictive modeling. However, there is limited research comparing the performance of classic and modern for variable selection in clinical datasets. MATERIALS AND METHODS: We analyzed the performance of eight different variable selection methods: four regression-based methods (stepwise backward selection using p-value and AIC, Least Absolute Shrinkage and Selection Operator, and Elastic Net) and four tree-based methods (Variable Selection Using Random Forest, Regularized Random Forests, Boruta, and Gradient Boosted Feature Selection). We used two clinical datasets of different sizes, a multicenter adult clinical deterioration cohort and a single center pediatric acute kidney injury cohort. Method evaluation included measures of parsimony, variable importance, and discrimination. RESULTS: In the large, multicenter dataset, the modern tree-based Variable Selection Using Random Forest and the Gradient Boosted Feature Selection methods achieved the best parsimony. In the smaller, single-center dataset, the classic regression-based stepwise backward selection using p-value and AIC methods achieved the best parsimony. In both datasets, variable selection tended to decrease the accuracy of the random forest models and increase the accuracy of logistic regression models. CONCLUSIONS: The performance of classic regression-based and modern tree-based variable selection methods is associated with the size of the clinical dataset used. Classic regression-based variable selection methods seem to achieve better parsimony in clinical prediction problems in smaller datasets while modern tree-based methods perform better in larger datasets.

The Effect of a Pediatric Intensive Care Severity-Tiered Pathway for Status Asthmaticus on Quality Measures and Outcomes.

Background: Patients with status asthmaticus admitted to the hospital contribute tremendous costs to the healthcare system. Treatment protocols directed at improving care to pediatric inpatients have improved both clinical and financial outcomes; however, there are limited data demonstrating the impact of goal-driven therapy for patients with status asthmaticus in the pediatric intensive care unit (PICU). The purpose of this initiative was to standardize care of children with status asthmaticus in the PICU to improve clinical outcomes. Materials and Methods: A goal-directed clinical pathway using a validated pediatric asthma severity score was developed. A pre-post intervention assessed patients 2-17 years of age admitted to the PICU with status asthmaticus. The effect of the protocol was measured from October 2015 to October 2016. The primary outcome was the transition time interval from continuous nebulized bronchodilator therapy to every 2 hourly intermittent treatments. Secondary outcomes related to treatment options and clinical complications were also assessed. Results: Postintervention patient (n = 124) demographics were similar to those in the preintervention group (n = 130). Compared with baseline, there was a 4.9-h decrease in the mean transition time from continuous to intermittent bronchodilator use (31.9 versus 27.0, P = 0.033) following the intervention. The mean and median hospital length of stay (LOS) was significantly decreased in the postintervention group by 14.76 h (P = 0.011) and 16.68 h (P = 0.003). There were no readmissions to the PICU in the postintervention group. The use of rescue BiPAP (Bi-level positive airway pressure ventilation) was significantly higher in the preintervention group compared with the postintervention group (42% versus 28%, P = 0.036). Conclusions: A severity-tiered, goal-driven treatment pathway for pediatric patients admitted to the PICU with status asthmaticus reduced the time required to transition from continuous to intermittent inhaled bronchodilator use. As a result, overall hospital LOS was significantly decreased with no increase in clinical complications.

Documenting Penicillin Allergy: The Impact of Inconsistency.

BACKGROUND: Allergy documentation is frequently inconsistent and incomplete. The impact of this variability on subsequent treatment is not well described. OBJECTIVE: To determine how allergy documentation affects subsequent antibiotic choice. DESIGN: Retrospective, cohort study. PARTICIPANTS: 232,616 adult patients seen by 199 primary care providers (PCPs) between January 1, 2009 and January 1, 2014 at an academic medical system. MAIN MEASURES: Inter-physician variation in beta-lactam allergy documentation; antibiotic treatment following beta-lactam allergy documentation. KEY RESULTS: 15.6% of patients had a reported beta-lactam allergy. Of those patients, 39.8% had a specific allergen identified and 22.7% had allergic reaction characteristics documented. Variation between PCPs was greater than would be expected by chance (all p<0.001) in the percentage of their patients with a documented beta-lactam allergy (7.9% to 24.8%), identification of a specific allergen (e.g. amoxicillin as opposed to "penicillins") (24.0% to 58.2%) and documentation of the reaction characteristics (5.4% to 51.9%). After beta-lactam allergy documentation, patients were less likely to receive penicillins (Relative Risk [RR] 0.16 [95% Confidence Interval: 0.15-0.17]) and cephalosporins (RR 0.28 [95% CI 0.27-0.30]) and more likely to receive fluoroquinolones (RR 1.5 [95% CI 1.5-1.6]), clindamycin (RR 3.8 [95% CI 3.6-4.0]) and vancomycin (RR 5.0 [95% CI 4.3-5.8]). Among patients with beta-lactam allergy, rechallenge was more likely when a specific allergen was identified (RR 1.6 [95% CI 1.5-1.8]) and when reaction characteristics were documented (RR 2.0 [95% CI 1.8-2.2]). CONCLUSIONS: Provider documentation of beta-lactam allergy is highly variable, and details of the allergy are infrequently documented. Classification of a patient as beta-lactam allergic and incomplete documentation regarding the details of the allergy lead to beta-lactam avoidance and use of other antimicrobial agents, behaviors that may adversely impact care quality and cost.

Targeted analysis of whole genome sequence data to diagnose genetic cardiomyopathy.

BACKGROUND: Cardiomyopathy is highly heritable but genetically diverse. At present, genetic testing for cardiomyopathy uses targeted sequencing to simultaneously assess the coding regions of >50 genes. New genes are routinely added to panels to improve the diagnostic yield. With the anticipated $1000 genome, it is expected that genetic testing will shift toward comprehensive genome sequencing accompanied by targeted gene analysis. Therefore, we assessed the reliability of whole genome sequencing and targeted analysis to identify cardiomyopathy variants in 11 subjects with cardiomyopathy. METHODS AND RESULTS: Whole genome sequencing with an average of 37× coverage was combined with targeted analysis focused on 204 genes linked to cardiomyopathy. Genetic variants were scored using multiple prediction algorithms combined with frequency data from public databases. This pipeline yielded 1 to 14 potentially pathogenic variants per individual. Variants were further analyzed using clinical criteria and segregation analysis, where available. Three of 3 previously identified primary mutations were detected by this analysis. In 6 subjects for whom the primary mutation was previously unknown, we identified mutations that segregated with disease, had clinical correlates, and had additional pathological correlation to provide evidence for causality. For 2 subjects with previously known primary mutations, we identified additional variants that may act as modifiers of disease severity. In total, we identified the likely pathological mutation in 9 of 11 (82%) subjects. CONCLUSIONS: These pilot data demonstrate that ≈30 to 40× coverage whole genome sequencing combined with targeted analysis is feasible and sensitive to identify rare variants in cardiomyopathy-associated genes.

Annexin A6 modifies muscular dystrophy by mediating sarcolemmal repair.

Many monogenic disorders, including the muscular dystrophies, display phenotypic variability despite the same disease-causing mutation. To identify genetic modifiers of muscular dystrophy and its associated cardiomyopathy, we used quantitative trait locus mapping and whole genome sequencing in a mouse model. This approach uncovered a modifier locus on chromosome 11 associated with sarcolemmal membrane damage and heart mass. Whole genome and RNA sequencing identified Anxa6, encoding annexin A6, as a modifier gene. A synonymous variant in exon 11 creates a cryptic splice donor, resulting in a truncated annexin A6 protein called ANXA6N32. Live cell imaging showed that annexin A6 orchestrates a repair zone and cap at the site of membrane disruption. In contrast, ANXA6N32 dramatically disrupted the annexin A6-rich cap and the associated repair zone, permitting membrane leak. Anxa6 is a modifier of muscular dystrophy and membrane repair after injury.

Abcc9 is required for the transition to oxidative metabolism in the newborn heart.

The newborn heart adapts to postnatal life by shifting from a fetal glycolytic metabolism to a mitochondrial oxidative metabolism. Abcc9, an ATP-binding cassette family member, increases expression concomitant with this metabolic shift. Abcc9 encodes a membrane-associated receptor that partners with a potassium channel to become the major potassium-sensitive ATP channel in the heart. Abcc9 also encodes a smaller protein enriched in the mitochondria. We now deleted exon 5 of Abcc9 to ablate expression of both plasma membrane and mitochondria-associated Abcc9-encoded proteins, and found that the myocardium failed to acquire normal mature metabolism, resulting in neonatal cardiomyopathy. Unlike wild-type neonatal cardiomyocytes, mitochondria from Ex5 cardiomyocytes were unresponsive to the KATP agonist diazoxide, consistent with loss of KATP activity. When exposed to hydrogen peroxide to induce cell stress, Ex5 neonatal cardiomyocytes displayed a rapid collapse of mitochondria membrane potential, distinct from wild-type cardiomyocytes. Ex5 cardiomyocytes had reduced fatty acid oxidation, reduced oxygen consumption and reserve. Morphologically, Ex5 cardiac mitochondria exhibited an immature pattern with reduced cross-sectional area and intermitochondrial contacts. In the absence of Abcc9, the newborn heart fails to transition normally from fetal to mature myocardial metabolism.-Fahrenbach, J. P., Stoller, D., Kim, G., Aggarwal, N., Yerokun, B., Earley, J. U., Hadhazy, M., Shi, N.-Q., Makielski, J. C., McNally, E. M. Abcc9 is required for the transition to oxidative metabolism in the newborn heart.

The CO-Regulation Database (CORD): a tool to identify coordinately expressed genes.

BACKGROUND: Meta-analysis of gene expression array databases has the potential to reveal information about gene function. The identification of gene-gene interactions may be inferred from gene expression information but such meta-analysis is often limited to a single microarray platform. To address this limitation, we developed a gene-centered approach to analyze differential expression across thousands of gene expression experiments and created the CO-Regulation Database (CORD) to determine which genes are correlated with a queried gene. RESULTS: Using the GEO and ArrayExpress database, we analyzed over 120,000 group by group experiments from gene microarrays to determine the correlating genes for over 30,000 different genes or hypothesized genes. CORD output data is presented for sample queries with focus on genes with well-known interaction networks including p16 (CDKN2A), vimentin (VIM), MyoD (MYOD1). CDKN2A, VIM, and MYOD1 all displayed gene correlations consistent with known interacting genes. CONCLUSIONS: We developed a facile, web-enabled program to determine gene-gene correlations across different gene expression microarray platforms. Using well-characterized genes, we illustrate how CORD's identification of co-expressed genes contributes to a better understanding a gene's potential function. The website is found at http://cord-db.org.

Population-based variation in cardiomyopathy genes.

BACKGROUND: Hypertrophic cardiomyopathy and dilated cardiomyopathy arise from mutations in genes encoding sarcomere proteins including MYH7, MYBPC3, and TTN. Genetic diagnosis of cardiomyopathy relies on complete sequencing of the gene coding regions, and most pathogenic variation is rare. The 1000 Genomes Project is an ongoing consortium designed to deliver whole genome sequence information from an ethnically diverse population and, therefore, is a rich source to determine both common and rare genetic variants. METHODS AND RESULTS: We queried the 1000 Genomes Project database of 1092 individuals for exonic variants within 3 sarcomere genes MHY7, MYBPC3, and TTN. We focused our analysis on protein-altering variation, including nonsynonymous single nucleotide polymorphisms, insertion/deletion polymorphisms, or splice site altering variants. We identified known and predicted pathogenic variation in MYBPC3 and MYH7 at a higher frequency than what would be expected based on the known prevalence of cardiomyopathy. We also found substantial variation, including protein-disrupting sequences, in TTN. CONCLUSIONS: Cardiomyopathy is a genetically heterogeneous disorder caused by mutations in multiple genes. The frequency of predicted pathogenic protein-altering variation in cardiomyopathy genes suggests that many of these variants may be insufficient to cause disease on their own but may modify phenotype in a genetically susceptible host. This is suggested by the high prevalence of TTN insertion/deletions in the 1000 Genomes Project cohort. Given the possibility of additional genetic variants that modify the phenotype of a primary driver mutation, broad-based genetic testing should be employed.

TBX5 drives Scn5a expression to regulate cardiac conduction system function.

Cardiac conduction system (CCS) disease, which results in disrupted conduction and impaired cardiac rhythm, is common with significant morbidity and mortality. Current treatment options are limited, and rational efforts to develop cell-based and regenerative therapies require knowledge of the molecular networks that establish and maintain CCS function. Recent genome-wide association studies (GWAS) have identified numerous loci associated with adult human CCS function, including TBX5 and SCN5A. We hypothesized that TBX5, a critical developmental transcription factor, regulates transcriptional networks required for mature CCS function. We found that deletion of Tbx5 from the mature murine ventricular conduction system (VCS), including the AV bundle and bundle branches, resulted in severe VCS functional consequences, including loss of fast conduction, arrhythmias, and sudden death. Ventricular contractile function and the VCS fate map remained unchanged in VCS-specific Tbx5 knockouts. However, key mediators of fast conduction, including Nav1.5, which is encoded by Scn5a, and connexin 40 (Cx40), demonstrated Tbx5-dependent expression in the VCS. We identified a TBX5-responsive enhancer downstream of Scn5a sufficient to drive VCS expression in vivo, dependent on canonical T-box binding sites. Our results establish a direct molecular link between Tbx5 and Scn5a and elucidate a hierarchy between human GWAS loci that affects function of the mature VCS, establishing a paradigm for understanding the molecular pathology of CCS disease.

Impaired muscle growth and response to insulin-like growth factor 1 in dysferlin-mediated muscular dystrophy.

Loss-of-function mutations in dysferlin cause muscular dystrophy, and dysferlin has been implicated in resealing membrane disruption in myofibers. Given the importance of membrane fusion in many aspects of muscle function, we studied the role of dysferlin in muscle growth. We found that dysferlin null myoblasts have a defect in myoblast-myotube fusion, resulting in smaller myotubes in culture. In vivo, dysferlin null muscle was found to have mislocalized nuclei and vacuolation. We found that myoblasts isolated from dysferlin null mice accumulate enlarged, lysosomal-associated membrane protein 2 (LAMP2)-positive lysosomes. Dysferlin null myoblasts accumulate transferrin-488, reflecting abnormal vesicular trafficking. Additionally, dysferlin null myoblasts display abnormal trafficking of the insulin-like growth factor (IGF) receptor, where the receptor is shuttled to LAMP2-positive lysosomes. We studied growth, in vivo, by infusing mice with the growth stimulant IGF1. Control IGF1-treated mice increased myofiber diameter by 30% as expected, whereas dysferlin null muscles had no response to IGF1, indicating a defect in myofiber growth. We also noted that dysferlin null fibroblasts also accumulate acidic vesicles, IGF receptor and transferrin, indicating that dysferlin is important for nonmuscle vesicular trafficking. These data implicate dysferlin in multiple membrane fusion events within the cell and suggest multiple pathways by which loss of dysferlin contributes to muscle disease.

Altered chromosomal positioning, compaction, and gene expression with a lamin A/C gene mutation.

BACKGROUND: Lamins A and C, encoded by the LMNA gene, are filamentous proteins that form the core scaffold of the nuclear lamina. Dominant LMNA gene mutations cause multiple human diseases including cardiac and skeletal myopathies. The nuclear lamina is thought to regulate gene expression by its direct interaction with chromatin. LMNA gene mutations may mediate disease by disrupting normal gene expression. METHODS/FINDINGS: To investigate the hypothesis that mutant lamin A/C changes the lamina's ability to interact with chromatin, we studied gene misexpression resulting from the cardiomyopathic LMNA E161K mutation and correlated this with changes in chromosome positioning. We identified clusters of misexpressed genes and examined the nuclear positioning of two such genomic clusters, each harboring genes relevant to striated muscle disease including LMO7 and MBNL2. Both gene clusters were found to be more centrally positioned in LMNA-mutant nuclei. Additionally, these loci were less compacted. In LMNA mutant heart and fibroblasts, we found that chromosome 13 had a disproportionately high fraction of misexpressed genes. Using three-dimensional fluorescence in situ hybridization we found that the entire territory of chromosome 13 was displaced towards the center of the nucleus in LMNA mutant fibroblasts. Additional cardiomyopathic LMNA gene mutations were also shown to have abnormal positioning of chromosome 13, although in the opposite direction. CONCLUSIONS: These data support a model in which LMNA mutations perturb the intranuclear positioning and compaction of chromosomal domains and provide a mechanism by which gene expression may be altered.

Cardiomyocyte sulfonylurea receptor 2-KATP channel mediates cardioprotection and ST segment elevation.

Sulfonylurea receptor-containing ATP-sensitive potassium (K(ATP)) channels have been implicated in cardioprotection, but the cell type and constitution of channels responsible for this protection have not been clear. Mice deleted for the first nucleotide binding region of sulfonylurea receptor 2 (SUR2) are referred to as SUR2 null since they lack full-length SUR2 and glibenclamide-responsive K(ATP) channels in cardiac, skeletal, and smooth muscle. As previously reported, SUR2 null mice develop electrocardiographic changes of ST segment elevation that were shown to correlate with coronary artery vasospasm. Here we restored expression of the cardiomyocyte SUR2-K(ATP) channel in SUR2 null mice by generating transgenic mice with ventricular cardiomyocyte-restricted expression of SUR2A. Introduction of the cardiomyocyte SUR2A transgene into the SUR2 null background restored functional cardiac K(ATP) channels. Hearts isolated from rescued mice, referred to as MLC2A, had significantly reduced infarct size (27 ± 3% of area at risk) compared with SUR2 null mice (36 ± 3% of area at risk). Compared with SUR2 null hearts, MLC2A hearts exhibited significantly improved cardiac function during the postischemia reperfusion period primarily because of preservation of low diastolic pressures. Additionally, restoration of cardiac SUR2-K(ATP) channels significantly reduced the degree and frequency of ST segment elevation episodes in MLC2A mice. Therefore, cardioprotective mechanisms both dependent and independent of SUR2-K(ATP) channels contribute to cardiac function.