Electronic health record biobank cohort recapitulates an association between the MUC5B promoter polymorphism and ARDS in critically ill adults.

Background: Large population-based DNA biobanks linked to electronic health records (EHRs) may provide advantages over traditional study designs for identifying genetic drivers of ARDS. Research Question: Can ARDS be identified in an EHR biobank, and can this approach validate a previously reported genetic risk factor for ARDS? Study Design and Methods: We analyzed two genotyped cohorts from one academic medical center: a prospective biomarker study of critically ill adults (VALID cohort), and hospitalized participants in a de-identified EHR biobank (BioVU). ARDS status was assessed by clinician-investigator review in VALID and an EHR-derived algorithm in BioVU (EHR-ARDS). We tested the association between the MUC5B promoter polymorphism (rs35705950) with development of ARDS/EHR-ARDS in each cohort. Results: In VALID, 2,795 patients were included, age was 55 [43, 66] (median [IQR]) years, and 718 (25.7%) developed ARDS. In BioVU, 9,025 hospitalized participants were included, age was 60 [48, 70] , and 1,056 (11.7%) developed EHR-ARDS. We observed a significant interaction between age and rs35705950 on ARDS risk in VALID: in older patients rs35705950 was associated with increased ARDS risk (OR: 1.44; 95%CI 1.08-1.92; p=0.012) whereas among younger patients this effect was attenuated (OR: 0.84; 95%CI: 0.62-1.14; p=0.26). In BioVU, rs35705950 was associated with increased risk for EHR-ARDS among all participants (OR: 1.20; 95%CI: 1.00-1.43, p=0.043) and this relationship did not vary by age. The polymorphism was also associated with more severe oxygenation impairment among BioVU participants who required mechanical ventilation. Interpretation: The MUC5B promoter polymorphism was associated with ARDS in two cohorts of at-risk hospitalized adults. Although age-related effect modification was observed only in the prospective biomarker cohort, the EHR cohort identified a consistent association between MUC5B and ARDS risk regardless of age and a novel association with oxygenation impairment. Our study highlights the potential for EHR biobanks to enable precision-medicine ARDS studies.

Causes and attributable fraction of death from ARDS in inflammatory phenotypes of sepsis.

BACKGROUND: Hypoinflammatory and hyperinflammatory phenotypes have been identified in both Acute Respiratory Distress Syndrome (ARDS) and sepsis. Attributable mortality of ARDS in each phenotype of sepsis is yet to be determined. We aimed to estimate the population attributable fraction of death from ARDS (PAFARDS) in hypoinflammatory and hyperinflammatory sepsis, and to determine the primary cause of death within each phenotype. METHODS: We studied 1737 patients with sepsis from two prospective cohorts. Patients were previously assigned to the hyperinflammatory or hypoinflammatory phenotype using latent class analysis. The PAFARDS in patients with sepsis was estimated separately in the hypo and hyperinflammatory phenotypes. Organ dysfunction, severe comorbidities, and withdrawal of life support were abstracted from the medical record in a subset of patients from the EARLI cohort who died (n = 130/179). Primary cause of death was defined as the organ system that most directly contributed to death or withdrawal of life support. RESULTS: The PAFARDS was 19% (95%CI 10,28%) in hypoinflammatory sepsis and, 14% (95%CI 6,20%) in hyperinflammatory sepsis. Cause of death differed between the two phenotypes (p < 0.001). Respiratory failure was the most common cause of death in hypoinflammatory sepsis, whereas circulatory shock was the most common cause in hyperinflammatory sepsis. Death with severe underlying comorbidities was more frequent in hypoinflammatory sepsis (81% vs. 67%, p = 0.004). CONCLUSIONS: The PAFARDS is modest in both phenotypes whereas primary cause of death among patients with sepsis differed substantially by phenotype. This study identifies challenges in powering future clinical trials to detect changes in mortality outcomes among patients with sepsis and ARDS.

Large language models facilitate the generation of electronic health record phenotyping algorithms.

OBJECTIVES: Phenotyping is a core task in observational health research utilizing electronic health records (EHRs). Developing an accurate algorithm demands substantial input from domain experts, involving extensive literature review and evidence synthesis. This burdensome process limits scalability and delays knowledge discovery. We investigate the potential for leveraging large language models (LLMs) to enhance the efficiency of EHR phenotyping by generating high-quality algorithm drafts. MATERIALS AND METHODS: We prompted four LLMs-GPT-4 and GPT-3.5 of ChatGPT, Claude 2, and Bard-in October 2023, asking them to generate executable phenotyping algorithms in the form of SQL queries adhering to a common data model (CDM) for three phenotypes (ie, type 2 diabetes mellitus, dementia, and hypothyroidism). Three phenotyping experts evaluated the returned algorithms across several critical metrics. We further implemented the top-rated algorithms and compared them against clinician-validated phenotyping algorithms from the Electronic Medical Records and Genomics (eMERGE) network. RESULTS: GPT-4 and GPT-3.5 exhibited significantly higher overall expert evaluation scores in instruction following, algorithmic logic, and SQL executability, when compared to Claude 2 and Bard. Although GPT-4 and GPT-3.5 effectively identified relevant clinical concepts, they exhibited immature capability in organizing phenotyping criteria with the proper logic, leading to phenotyping algorithms that were either excessively restrictive (with low recall) or overly broad (with low positive predictive values). CONCLUSION: GPT versions 3.5 and 4 are capable of drafting phenotyping algorithms by identifying relevant clinical criteria aligned with a CDM. However, expertise in informatics and clinical experience is still required to assess and further refine generated algorithms.

Development and External Validation of Models to Predict Persistent Hypoxemic Respiratory Failure for Clinical Trial Enrichment.

OBJECTIVES: Improving the efficiency of clinical trials in acute hypoxemic respiratory failure (HRF) depends on enrichment strategies that minimize enrollment of patients who quickly resolve with existing care and focus on patients at high risk for persistent HRF. We aimed to develop parsimonious models predicting risk of persistent HRF using routine data from ICU admission and select research immune biomarkers. DESIGN: Prospective cohorts for derivation ( n = 630) and external validation ( n = 511). SETTING: Medical and surgical ICUs at two U.S. medical centers. PATIENTS: Adults with acute HRF defined as new invasive mechanical ventilation (IMV) and hypoxemia on the first calendar day after ICU admission. INTERVENTIONS: None. MEASUREMENTS AND MAIN RESULTS: We evaluated discrimination, calibration, and practical utility of models predicting persistent HRF risk (defined as ongoing IMV and hypoxemia on the third calendar day after admission): 1) a clinical model with least absolute shrinkage and selection operator (LASSO) selecting Pa o2 /F io2 , vasopressors, mean arterial pressure, bicarbonate, and acute respiratory distress syndrome as predictors; 2) a model adding interleukin-6 (IL-6) to clinical predictors; and 3) a comparator model with Pa o2 /F io2 alone, representing an existing strategy for enrichment. Forty-nine percent and 69% of patients had persistent HRF in derivation and validation sets, respectively. In validation, both LASSO (area under the receiver operating characteristic curve, 0.68; 95% CI, 0.64-0.73) and LASSO + IL-6 (0.71; 95% CI, 0.66-0.76) models had better discrimination than Pa o2 /F io2 (0.64; 95% CI, 0.59-0.69). Both models underestimated risk in lower risk deciles, but exhibited better calibration at relevant risk thresholds. Evaluating practical utility, both LASSO and LASSO + IL-6 models exhibited greater net benefit in decision curve analysis, and greater sample size savings in enrichment analysis, compared with Pa o2 /F io2 . The added utility of LASSO + IL-6 model over LASSO was modest. CONCLUSIONS: Parsimonious, interpretable models that predict persistent HRF may improve enrichment of trials testing HRF-targeted therapies and warrant future validation.

Large Language Models Facilitate the Generation of Electronic Health Record Phenotyping Algorithms.

Objectives: Phenotyping is a core task in observational health research utilizing electronic health records (EHRs). Developing an accurate algorithm demands substantial input from domain experts, involving extensive literature review and evidence synthesis. This burdensome process limits scalability and delays knowledge discovery. We investigate the potential for leveraging large language models (LLMs) to enhance the efficiency of EHR phenotyping by generating high-quality algorithm drafts. Materials and Methods: We prompted four LLMs-GPT-4 and GPT-3.5 of ChatGPT, Claude 2, and Bard-in October 2023, asking them to generate executable phenotyping algorithms in the form of SQL queries adhering to a common data model (CDM) for three phenotypes (i.e., type 2 diabetes mellitus, dementia, and hypothyroidism). Three phenotyping experts evaluated the returned algorithms across several critical metrics. We further implemented the top-rated algorithms and compared them against clinician-validated phenotyping algorithms from the Electronic Medical Records and Genomics (eMERGE) network. Results: GPT-4 and GPT-3.5 exhibited significantly higher overall expert evaluation scores in instruction following, algorithmic logic, and SQL executability, when compared to Claude 2 and Bard. Although GPT-4 and GPT-3.5 effectively identified relevant clinical concepts, they exhibited immature capability in organizing phenotyping criteria with the proper logic, leading to phenotyping algorithms that were either excessively restrictive (with low recall) or overly broad (with low positive predictive values). Conclusion: GPT versions 3.5 and 4 are capable of drafting phenotyping algorithms by identifying relevant clinical criteria aligned with a CDM. However, expertise in informatics and clinical experience is still required to assess and further refine generated algorithms.

MICB Genomic Variant Is Associated with NKG2D-mediated Acute Lung Injury and Death.

Rationale: Acute lung injury (ALI) carries a high risk of mortality but has no established pharmacologic therapy. We previously found that experimental ALI occurs through natural killer (NK) cell NKG2D receptor activation and that the cognate human ligand, MICB, was associated with ALI after transplantation. Objectives: To investigate the association of a common missense variant, MICBG406A, with ALI. Methods: We assessed MICBG406A genotypes within two multicenter observational study cohorts at risk for ALI: primary graft dysfunction (N = 619) and acute respiratory distress syndrome (N = 1,376). Variant protein functional effects were determined in cultured and ex vivo human samples. Measurements and Main Results: Recipients of MICBG406A-homozygous allografts had an 11.1% absolute risk reduction (95% confidence interval [CI], 3.2-19.4%) for severe primary graft dysfunction after lung transplantation and reduced risk for allograft failure (hazard ratio, 0.36; 95% CI, 0.13-0.98). In participants with sepsis, we observed 39% reduced odds of moderately or severely impaired oxygenation among MICBG406A-homozygous individuals (95% CI, 0.43-0.86). BAL NK cells were less frequent and less mature in participants with MICBG406A. Expression of missense variant protein MICBD136N in cultured cells resulted in reduced surface MICB and reduced NKG2D ligation relative to wild-type MICB. Coculture of variant MICBD136N cells with NK cells resulted in less NKG2D activation and less susceptibility to NK cell killing relative to the wild-type cells. Conclusions: These data support a role for MICB signaling through the NKG2D receptor in mediating ALI, suggesting a novel therapeutic approach.

Identifying molecular phenotypes in sepsis: an analysis of two prospective observational cohorts and secondary analysis of two randomised controlled trials.

BACKGROUND: In sepsis and acute respiratory distress syndrome (ARDS), heterogeneity has contributed to difficulty identifying effective pharmacotherapies. In ARDS, two molecular phenotypes (hypoinflammatory and hyperinflammatory) have consistently been identified, with divergent outcomes and treatment responses. In this study, we sought to derive molecular phenotypes in critically ill adults with sepsis, determine their overlap with previous ARDS phenotypes, and evaluate whether they respond differently to treatment in completed sepsis trials. METHODS: We used clinical data and plasma biomarkers from two prospective sepsis cohorts, the Validating Acute Lung Injury biomarkers for Diagnosis (VALID) study (N=1140) and the Early Assessment of Renal and Lung Injury (EARLI) study (N=818), in latent class analysis (LCA) to identify the optimal number of classes in each cohort independently. We used validated models trained to classify ARDS phenotypes to evaluate concordance of sepsis and ARDS phenotypes. We applied these models retrospectively to the previously published Prospective Recombinant Human Activated Protein C Worldwide Evaluation in Severe Sepsis and Septic Shock (PROWESS-SHOCK) trial and Vasopressin and Septic Shock Trial (VASST) to assign phenotypes and evaluate heterogeneity of treatment effect. FINDINGS: A two-class model best fit both VALID and EARLI (p<0·0001). In VALID, 804 (70·5%) of the 1140 patients were classified as hypoinflammatory and 336 (29·5%) as hyperinflammatory; in EARLI, 530 (64·8%) of 818 were hypoinflammatory and 288 (35·2%) hyperinflammatory. We observed higher plasma pro-inflammatory cytokines, more vasopressor use, more bacteraemia, lower protein C, and higher mortality in the hyperinflammatory than in the hypoinflammatory phenotype (p<0·0001 for all). Classifier models indicated strong concordance between sepsis phenotypes and previously identified ARDS phenotypes (area under the curve 0·87-0·96, depending on the model). Findings were similar excluding participants with both sepsis and ARDS. In PROWESS-SHOCK, 1142 (68·0%) of 1680 patients had the hypoinflammatory phenotype and 538 (32·0%) had the hyperinflammatory phenotype, and response to activated protein C differed by phenotype (p=0·0043). In VASST, phenotype proportions were similar to other cohorts; however, no treatment interaction with the type of vasopressor was observed (p=0·72). INTERPRETATION: Molecular phenotypes previously identified in ARDS are also identifiable in multiple sepsis cohorts and respond differently to activated protein C. Molecular phenotypes could represent a treatable trait in critical illness beyond the patient's syndromic diagnosis. FUNDING: US National Institutes of Health.

De-black-boxing health AI: demonstrating reproducible machine learning computable phenotypes using the N3C-RECOVER Long COVID model in the All of Us data repository.

Machine learning (ML)-driven computable phenotypes are among the most challenging to share and reproduce. Despite this difficulty, the urgent public health considerations around Long COVID make it especially important to ensure the rigor and reproducibility of Long COVID phenotyping algorithms such that they can be made available to a broad audience of researchers. As part of the NIH Researching COVID to Enhance Recovery (RECOVER) Initiative, researchers with the National COVID Cohort Collaborative (N3C) devised and trained an ML-based phenotype to identify patients highly probable to have Long COVID. Supported by RECOVER, N3C and NIH's All of Us study partnered to reproduce the output of N3C's trained model in the All of Us data enclave, demonstrating model extensibility in multiple environments. This case study in ML-based phenotype reuse illustrates how open-source software best practices and cross-site collaboration can de-black-box phenotyping algorithms, prevent unnecessary rework, and promote open science in informatics.

The relationship between high density lipoprotein cholesterol and sepsis: A clinical and genetic approach.

Sepsis accounts for one in three hospital deaths. Higher concentrations of high-density lipoprotein cholesterol (HDL-C) are associated with apparent protection from sepsis, suggesting a potential therapeutic role for HDL-C or drugs, such as cholesteryl ester transport protein (CETP) inhibitors that increase HDL-C. However, these beneficial clinical associations might be due to confounding; genetic approaches can address this possibility. We identified 73,406 White adults admitted to Vanderbilt University Medical Center with infection; 11,612 had HDL-C levels, and 12,377 had genotype information from which we constructed polygenic risk scores (PRS) for HDL-C and the effect of CETP on HDL-C. We tested the associations between predictors (measured HDL-C, HDL-C PRS, CETP PRS, and rs1800777) and outcomes: sepsis, septic shock, respiratory failure, and in-hospital death. In unadjusted analyses, lower measured HDL-C concentrations were significantly associated with increased risk of sepsis (p = 2.4 × 10-23 ), septic shock (p = 4.1 × 10-12 ), respiratory failure (p = 2.8 × 10-8 ), and in-hospital death (p = 1.0 × 10-8 ). After adjustment (age, sex, electronic health record length, comorbidity score, LDL-C, triglycerides, and body mass index), these associations were markedly attenuated: sepsis (p = 2.6 × 10-3 ), septic shock (p = 8.1 × 10-3 ), respiratory failure (p = 0.11), and in-hospital death (p = 4.5 × 10-3 ). HDL-C PRS, CETP PRS, and rs1800777 significantly predicted HDL-C (p < 2 × 10-16 ), but none were associated with sepsis outcomes. Concordant findings were observed in 13,254 Black patients hospitalized with infections. Lower measured HDL-C levels were significantly associated with increased risk of sepsis and related outcomes in patients with infection, but a causal relationship is unlikely because no association was found between the HDL-C PRS or the CETP PRS and the risk of adverse sepsis outcomes.