Measuring Implicit Bias in ICU Notes Using Word-Embedding Neural Network Models

BACKGROUND

Language in nonmedical data sets is known to transmit human-like biases when used in natural language processing (NLP) algorithms that can reinforce disparities. It is unclear if NLP algorithms of medical notes could lead to similar transmissions of biases.

RESEARCH QUESTION

Can we identify implicit bias in clinical notes, and are biases stable across time and geography?

STUDY DESIGN AND METHODS

To determine whether different racial and ethnic descriptors are similar contextually to stigmatizing language in ICU notes and whether these relationships are stable across time and geography, we identified notes on critically ill adults admitted to the University of California, San Francisco (UCSF), from 2012 through 2022 and to Beth Israel Deaconess Hospital (BIDMC) from 2001 through 2012. Because word meaning is derived largely from context, we trained unsupervised word-embedding algorithms to measure the similarity (cosine similarity) quantitatively of the context between a racial or ethnic descriptor (eg, African-American) and a stigmatizing target word (eg, nonco-operative) or group of words (violence, passivity, noncompliance, nonadherence).

RESULTS

In UCSF notes, Black descriptors were less likely to be similar contextually to violent words compared with White descriptors. Contrastingly, in BIDMC notes, Black descriptors were more likely to be similar contextually to violent words compared with White descriptors. The UCSF data set also showed that Black descriptors were more similar contextually to passivity and noncompliance words compared with Latinx descriptors.

INTERPRETATION

Implicit bias is identifiable in ICU notes. Racial and ethnic group descriptors carry different contextual relationships to stigmatizing words, depending on when and where notes were written. Because NLP models seem able to transmit implicit bias from training data, use of NLP algorithms in clinical prediction could reinforce disparities. Active debiasing strategies may be necessary to achieve algorithmic fairness when using language models in clinical research.

Negativity and Positivity in the ICU: Exploratory Development of Automated Sentiment Capture in the Electronic Health Record

OBJECTIVES

To develop proof-of-concept algorithms using alternative approaches to capture provider sentiment in ICU notes.

DESIGN

Retrospective observational cohort study.

SETTING

The Multiparameter Intelligent Monitoring of Intensive Care III (MIMIC-III) and the University of California, San Francisco (UCSF) deidentified notes databases.

PATIENTS

Adult (≥18 yr old) patients admitted to the ICU.

MEASUREMENTS AND MAIN RESULTS

We developed two sentiment models: 1) a keywords-based approach using a consensus-based clinical sentiment lexicon comprised of 72 positive and 103 negative phrases, including negations and 2) a Decoding-enhanced Bidirectional Encoder Representations from Transformers with disentangled attention-v3-based deep learning model (keywords-independent) trained on clinical sentiment labels. We applied the models to 198,944 notes across 52,997 ICU admissions in the MIMIC-III database. Analyses were replicated on an external sample of patients admitted to a UCSF ICU from 2018 to 2019. We also labeled sentiment in 1,493 note fragments and compared the predictive accuracy of our tools to three popular sentiment classifiers. Clinical sentiment terms were found in 99% of patient visits across 88% of notes. Our two sentiment tools were substantially more predictive (Spearman correlations of 0.62-0.84, p values < 0.00001) of labeled sentiment compared with general language algorithms (0.28-0.46).

CONCLUSION

Our exploratory healthcare-specific sentiment models can more accurately detect positivity and negativity in clinical notes compared with general sentiment tools not designed for clinical usage.

Trends in Geriatric Conditions Among Older Adults Admitted to US ICUs Between 1998 and 2015

BACKGROUND

Older adults are increasingly admitted to the ICU, and those with disabilities, dementia, frailty, and multimorbidity are vulnerable to adverse outcomes. Little is known about how pre-existing geriatric conditions have changed over time.

RESEARCH QUESTION

How have changes in disability, dementia, frailty, and multimorbidity in older adults admitted to the ICU changed from 1998 through 2015?

STUDY DESIGN AND METHODS

Medicare-linked Health and Retirement Survey (HRS) data identifying patients 65 years of age and older admitted to an ICU between 1998 and 2015. ICU admission was the unit of analysis. Year of ICU admission was the exposure. Disability, dementia, frailty, and multimorbidity were identified based on responses to HRS surveys before ICU admission. Disability represented the need for assistance with ≥ 1 activity of daily living. Dementia used cognitive and functional measures. Frailty included deficits in ≥ 2 domains (physical, nutritive, cognitive, or sensory function). Multimorbidity represented ≥ 3 self-reported chronic diseases. Time trends in geriatric conditions were modeled as a function of year of ICU admission and were adjusted for age, sex, race or ethnicity, and proxy interview status.

RESULTS

Across 6,084 ICU patients, age at admission increased from 77.6 years (95% CI, 76.7-78.4 years) in 1998 to 78.7 years (95% CI, 77.5-79.8 years) in 2015 (P < .001 for trend). The adjusted proportion of ICU admissions with pre-existing disability rose from 15.5% (95% CI, 12.1%-18.8%) in 1998 to 24.0% (95% CI, 18.5%-29.6%) in 2015 (P = .001). Rates of dementia did not change significantly (P = .21). Frailty increased from 36.6% (95% CI, 30.9%-42.3%) in 1998 to 45.0% (95% CI, 39.7%-50.2%) in 2015 (P = .04); multimorbidity rose from 54.4% (95% CI, 49.2%-59.7%) in 1998 to 71.8% (95% CI, 66.3%-77.2%) in 2015 (P < .001).

INTERPRETATION

Rates of pre-existing disability, frailty, and multimorbidity in older adults admitted to ICUs increased over time. Geriatric principles need to be deeply integrated into the ICU setting.

Utilization and Delivery of Specialty Palliative Care in the ICU: Insights from the Palliative Care Quality Network

CONTEXT

Palliative care (PC) benefits critically ill patients but remains underutilized. Important to developing interventions to overcome barriers to PC in the ICU and address PC needs of ICU patients is to understand how, when, and for which patients PC is provided in the ICU.

OBJECTIVES

Compare characteristics of specialty PC consultations in the ICU to those on medical-surgical wards.

METHODS

Retrospective analysis of national Palliative Care Quality Network data for hospitalized patients receiving specialty PC consultation January 1, 2013 to December 31, 2019 in ICU or medical-surgical setting. 98 inpatient PC teams in 16 states contributed data. Measures and outcomes included patient characteristics, consultation features, process metrics and patient outcomes. Mixed effects multivariable logistic regression was used to compare ICU and medical-surgical units.

RESULTS

Of 102,597 patients 63,082 were in medical-surgical units and 39,515 ICU. ICU patients were younger and more likely to have non-cancer diagnoses (all P < 0.001). While fewer ICU patients were able to report symptoms, most patients in both groups reported improved symptoms. ICU patients were more likely to have consultation requests for GOC, comfort care, and withdrawal of interventions and less likely for pain and/or symptoms (OR-all P < 0.001). ICU patients were less often discharged alive.

CONCLUSION

ICU patients receiving PC consultation are more likely to have non-cancer diagnoses and less likely able to communicate. Although symptom management and GOC are standard parts of ICU care, specialty PC in the ICU is often engaged for these issues and results in improved symptoms, suggesting routine interventions and consultation targeting these needs could improve care.

Racial and Ethnic Differences in the Prevalence of Do-Not-Resuscitate Orders among Older Adults with Severe Traumatic Brain Injury

Background

Older adults suffering from traumatic brain injury (TBI) are subject to higher injury burden and mortality. Do Not Resuscitate (DNR) orders are used to provide care aligned with patient wishes, but they may not be equitably distributed across racial/ethnic groups. We examined racial/ethnic differences in the prevalence of DNR orders at hospital admission in older patients with severe TBI.

Methods

We conducted a retrospective cohort study using the National Trauma Databank (NTDB) between 2007 to 2016. We examined patients ≥ 65 years with severe TBI. For our primary aim, the exposure was race/ethnicity and outcome was the presence of a documented DNR at hospital admission. We conducted an exploratory analysis of hospital outcomes including hospital mortality, discharge to hospice, and healthcare utilization (intracranial pressure monitor placement, hospital LOS, and duration of mechanical ventilation).

Results

Compared to White patients, Black patients (OR 0.48, 95% CI 0.35-0.64), Hispanic patients (OR 0.54, 95% CI 0.40-0.70), and Asian patients (OR 0.63, 95% CI 0.44-0.90) had decreased odds of having a DNR order at hospital admission. Patients with DNRs had increased odds of hospital mortality (OR 2.16, 95% CI 1.94-2.42), discharge to hospice (OR 2.08, 95% CI 1.75-2.46), shorter hospital LOS (−2.07 days, 95% CI −3.07 to −1.08) and duration of mechanical ventilation (−1.09 days, 95% CI −1.52 to −0.67). There was no significant difference in the utilization of ICP monitoring (OR 0.94, 95% CI 0.78-1.12).

Conclusions

We found significant racial and ethnic differences in the utilization of DNR orders among older patients with severe TBI. Additionally. DNR orders at hospital admission were associated with increased in-hospital mortality, increased hospice utilization, and decreased healthcare utilization. Future studies should examine mechanisms underlying race-based differences in DNR utilization.

Trends in Palliative Care Consultations in Critically Ill Patient Populations, 2013-2019

CONTEXT

Critically ill patients have important palliative care (PC) needs in the intensive care unit (ICU), but specialty PC is often underutilized.

OBJECTIVE

To evaluate changes in utilization and reasons for PC consultation over time.

METHODS

Data from a national multi-site network of inpatient PC visits were used to identify patients age ≥18 years admitted to an ICU between 2013 and 2019. Year of ICU admission was the exposure. Primary diagnosis and reason for referral were identified by standardized process measures within the dataset at the time of referral. Trends in primary diagnosis and reason for referral were modeled as a function of year of ICU admission.

RESULTS

Across 39,515 ICU patients seen by a PC team, overall numbers of consultations from the ICU increased each year. Referrals for patients with cancer decreased from 17.6% (95% CI 13.7%-21.5%) to 14.3% (95% CI 13.2%-14.7%) and for patients with cardiovascular disease increased from 16.8% in (95% CI 16.8%-16.9%) to 18.8% (95% CI 18.8%-18.9%). Reasons for referrals were primarily for goals of care and advance care planning and increased from 74.0% (95% CI 70.0%-78.0%) in 2013 to 80.0% (95% CI 79.4%-80.0%) in 2019 (P < 0.0001 for all trends).

CONCLUSION

PC referrals in ICU patients with cancer are decreasing, while those for cardiovascular disease are increasing. Reasons for referrals in the ICU are commonly for goals of care; other reasons, like pain control are uncommon. Early goals of care conversations and further training in advance care planning should be emphasized in the ICU setting.

Exploring Factors Associated With Morbidity and Mortality in Patients With Do-Not-Resuscitate Orders: A National Surgical Quality Improvement Program Database Analysis Within Surgical Groups

BACKGROUND

Anesthesiologists caring for patients with do-not-resuscitate (DNR) orders may have ethical concerns because of their resuscitative wishes and may have clinical concerns because of their known increased risk of morbidity/mortality. Patient heterogeneity and/or emphasis on mortality outcomes make previous studies among patients with DNR orders difficult to interpret. We sought to explore factors associated with morbidity and mortality among patients with DNR orders, which were stratified by surgical subgroups.

METHODS

Exploratory retrospective cohort study in adult patients undergoing prespecified colorectal, vascular, and orthopedic surgeries was performed using the American College of Surgeons National Surgical Quality Improvement Program Participant Use File data from 2010 to 2013. Among patients with preoperative DNR orders (ie, active DNR order written in the patient's chart before surgery), factors associated with 30-day mortality, increased length of stay, and inpatient death were determined via penalized regression. Unadjusted and adjusted estimates for selected variables are presented.

RESULTS

After selection as above, 211,420 patients underwent prespecified procedures, and of those, 2755 (1.3%) had pre-existing DNR orders and met above selection to address morbidity/mortality aims. By specialty, of these patients with a preoperative DNR, 1149 underwent colorectal, 870 vascular, and 736 orthopedic surgery. Across groups, 36.2% were male and had a mean age 79.9 years (range 21-90). The 30-day mortality was 15.4%-27.2% and median length of stay was 6-12 days. Death at discharge was 7.0%, 13.1%, and 23.0% in orthopedics, vascular, and colorectal patients with a DNR, respectively. The strongest factors associated with increased odds of 30-day mortality were preoperative septic shock in colorectal patients, preoperative ascites in vascular patients, and any requirement of mechanical ventilation at admission in orthopedic patients.

CONCLUSIONS

In patients with DNR orders undergoing common surgical procedures, the association of characteristics with morbidity and mortality varies in both direction and magnitude. The DNR order itself should not be the defining measure of risk.

Geospatial Variations and Neighborhood Deprivation in Drug-Related Admissions and Overdoses

Drug overdoses are a national and global epidemic. However, while overdoses are inextricably linked to social, demographic, and geographical determinants, geospatial patterns of drug-related admissions and overdoses at the neighborhood level remain poorly studied. The objective of this paper is to investigate spatial distributions of patients admitted for drug-related admissions and overdoses from a large, urban, tertiary care center using electronic health record data. Additionally, these spatial distributions were adjusted for a validated socioeconomic index called the Area Deprivation Index (ADI). We showed spatial heterogeneity in patients admitted for opioid, amphetamine, and psychostimulant-related diagnoses and overdoses. While ADI was associated with drug-related admissions, it did not correct for spatial variations and could not account alone for this spatial heterogeneity.

Bleeding After Musculoskeletal Surgery in Hospitals That Switched From Hydroxyethyl Starch to Albumin Following a Food and Drug Administration Warning

BACKGROUND

While US Food and Drug Administration (FDA) black box warnings are common, their impact on perioperative outcomes is unclear. Hydroxyethyl starch (HES) is associated with increased bleeding and kidney injury in patients with sepsis, leading to an FDA black box warning in 2013. Among patients undergoing musculoskeletal surgery in a subset of hospitals where colloid use changed from HES to albumin following the FDA warning, we examined the rate of major perioperative bleeding post- versus pre-FDA warning.

METHODS

The authors of this article used a retrospective, quasi-experimental, repeated cross-sectional, interrupted time series study of patients undergoing musculoskeletal surgery in hospitals within the Premier Healthcare Database, in the year before and year after the 2013 FDA black box warning. We examined patients in 23 "switcher" hospitals (where the percentage of colloid recipients receiving HES exceeded 50% before the FDA warning and decreased by at least 25% in absolute terms after the FDA warning) and patients in 279 "nonswitcher" hospitals. Among patients having surgery in "switcher" and "nonswitcher" hospitals, we determined monthly rates of major perioperative bleeding during the 12 months after the FDA warning, compared to 12 months before the FDA warning. Among patients who received surgery in "switcher" hospitals, we conducted a propensity-weighted segmented regression analysis assessing differences-in-differences (DID), using patients in "nonswitcher" hospitals as a control group.

RESULTS

Among 3078 patients treated at "switcher" hospitals (1892 patients treated pre-FDA warning versus 1186 patients treated post-FDA warning), demographic and clinical characteristics were well-balanced. Two hundred fifty-one (13.3%) received albumin pre-FDA warning, and 900 (75.9%) received albumin post-FDA warning. Among patients undergoing surgery in "switcher" hospitals during the pre-FDA warning period, 282 of 1892 (14.9%) experienced major bleeding during the hospitalization, compared to 149 of 1186 (12.6%) following the warning. In segmented regression, the adjusted ratio of slopes for major perioperative bleeding post- versus pre-FDA warning was 0.98 (95% confidence interval [CI], 0.93-1.04). In the DID estimate using "nonswitcher" hospitals as a control group, the ratio of ratios was 0.93 (95% CI, 0.46-1.86), indicating no significant difference.

CONCLUSIONS

We identified a subset of hospitals where colloid use for musculoskeletal surgery changed following a 2013 FDA black box warning regarding HES use in sepsis. Among patients undergoing musculoskeletal surgery at these "switcher" hospitals, there was no significant decrease in the rate of major perioperative bleeding following the warning, possibly due to incomplete practice change. Evaluation of the impact of systemic changes in health care may contribute to the understanding of patient outcomes in perioperative medicine.

A clinical investigation into the benefits of using charge codes in perioperative and critical care epidemiology: A retrospective cohort database study

Context:

Epidemiologic studies in critical care routinely rely on the codes listed in International Classification of Diseases (ICD) manuals which are primarily intended for reimbursement of claims to payers. Standardized billing codes may minimize the measurement error when used in conjunction with ICD codes.

Aims:

The aim was to examine the impact of using charge codes in addition to ICD codes for ascertaining two common procedures in surgical intensive care unit (ICU) settings: hemodialysis (HD) and red blood cell (RBC) transfusions.

Settings and Design:

This was a retrospective cohort study of Premier Inc. Database.

Subjects and Methods:

Elective surgical patients aged >18 years treated in the ICU postoperatively were included in this study. This includes the ascertainment of HD and RBC transfusions in the population using a standard “ICD code” versus an “either ICD code or charge code” approach.

Statistical Analysis Used:

Descriptive analysis using t-tests, Chi-square tests as appropriate was used.

Results:

A total of 40,357 patients were identified as having undergone elective surgery, followed by admission to an ICU across 520 US hospitals. The use of “ICD codes only” uniformly underestimated rates of HD or RBC transfusions when compared to “Charge Codes only” and “ICD Codes or Charge Codes” (% increase of 15.4%–45.6% and 50.8%–93.1%, respectively). Differences varied with specific surgical populations studied. Patients identified using the “ICD code” approach had more comorbidities, were more likely to be female, and more likely to be Medicare beneficiaries.

Conclusions:

Epidemiologic studies in critical care should consider using multiple independent data sources to improve ascertainment of common critical care interventions.

Aldehyde dehydrogenase 3a2 protects AML cells from oxidative death and the synthetic lethality of ferroptosis inducers

Metabolic alterations in cancer represent convergent effects of oncogenic mutations. We hypothesized that a metabolism-restricted genetic screen, comparing normal primary mouse hematopoietic cells and their malignant counterparts in an ex vivo system mimicking the bone marrow microenvironment, would define distinctive vulnerabilities in acute myeloid leukemia (AML). Leukemic cells, but not their normal myeloid counterparts, depended on the aldehyde dehydrogenase 3a2 (Aldh3a2) enzyme that oxidizes long-chain aliphatic aldehydes to prevent cellular oxidative damage. Aldehydes are by-products of increased oxidative phosphorylation and nucleotide synthesis in cancer and are generated from lipid peroxides underlying the non-caspase-dependent form of cell death, ferroptosis. Leukemic cell dependence on Aldh3a2 was seen across multiple mouse and human myeloid leukemias. Aldh3a2 inhibition was synthetically lethal with glutathione peroxidase-4 (GPX4) inhibition; GPX4 inhibition is a known trigger of ferroptosis that by itself minimally affects AML cells. Inhibiting Aldh3a2 provides a therapeutic opportunity and a unique synthetic lethality to exploit the distinctive metabolic state of malignant cells.

Delayed Neuromuscular Blockade Reversal With Sugammadex After Vecuronium, Desflurane, and Magnesium Administration: A Case Report

A variety of factors are known to prolong neuromuscular blockade, including several medications commonly used in anesthetic practice. We present a patient who underwent general anesthesia using desflurane, vecuronium, and magnesium infusion with delayed neuromuscular blockade reversal after sugammadex administration. A higher than anticipated total dose of sugammadex was required for adequate reversal, and quantitative neuromuscular monitoring was essential to ensuring complete neuromuscular recovery before extubation in this case.

Intrapulmonary Activated Factor VII for Hemoptysis Complicating Pulmonary Thromboendarterectomy

Massive hemoptysis represents a life-threatening disorder that has numerous different causes. The development of recombinant factor concentrates has allowed for novel treatments in this emergency setting. This report describes a patient with chronic thromboembolic pulmonary hypertension who underwent pulmonary thromboendarterectomy. The postoperative course was complicated by massive hemoptysis resulting in severe hypoxemia that required extracorporeal membrane oxygenation and multiple daily blood transfusions. After failure of conservative and interventional approaches, recombinant factor VII was administered by bronchial isolation. After treatment, the patient's hemoptysis dramatically resolved, with eventual hospital discharge and excellent function at follow-up. This case presents the use of intrapulmonary activated factor VII to control massive hemoptysis.

Activity of single-agent bevacizumab in patients with metastatic renal cell carcinoma previously treated with vascular endothelial growth factor tyrosine kinase inhibitors

PURPOSE

The activity of systemic agents after progression when using vascular endothelial growth factor receptor (VEGFR) tyrosine kinase inhibition (TKI) and mammalian target of rapamycin (mTOR) inhibition in patients with metastatic renal cell carcinoma (mRCC) is poorly characterized. The anti-vascular endothelial growth factor (VEGF) monoclonal antibody bevacizumab has a broad US Food and Drug Administration label and National Comprehensive Cancer Network guideline level 2b recommendation in this setting; we thus explored our institutional experience in this population.

METHODS

We conducted a retrospective analysis of patients with mRCC who were treated with bevacizumab in the second- and/or third-line settings; the primary endpoint was progression-free survival (PFS). Overall response rates (ORR), overall survival (OS), and toxicity were analyzed.

RESULTS

Twenty-one patients were treated with bevacizumab: the median age was 63 years old; 80% were white and 14% were black; 80% had clear cell histology. All the patients had prior VEGFR TKI therapy; 43% had prior mTOR inhibitor; the median number of prior therapies was 3. The median PFS was 4.4 months (95% CI, 2.8-9.6 months), and the median OS was 19.4 months (95% CI, 9.9-NR months). ORR was 9.5%; 52% of subjects had stable disease as best response, and 52% had disease progression. For subjects treated with prior VEGF and mTOR inhibitors, median PFS and OS were 4.4 and 13.2 months, respectively. Grade 3 to 4 toxicities included fatigue (29%), dehydration (24%), failure to thrive (10%), constipation (10%), and muscle weakness (10%).

CONCLUSIONS

Single-agent bevacizumab has acceptable toxicity and moderate disease-stabilizing activity in selected patients with mRCC who have failed prior VEGFR TKI and mTOR inhibitors. These data support clinical benefit to continued ongoing VEGF inhibition. Further prospective studies of bevacizumab alone or with alternative targeted agents in previously treated populations with mRCC are warranted.

Activity of single-agent bevacizumab (B) in patients with metastatic renal cell carcinoma (RCC) previously treated with VEGF- and mTOR-based therapies

435

Background: Given a lack of clinical information on therapeutic efficacy of agents following progression after vascular endothelial growth factor (VEGF) tyrosine kinase inhibition (TKI) and mammalian target of rapamycin (mTOR) inhibition in metastatic renal cell carcinoma (mRCC), we investigated the activity of single agent bevacizumab (B) in this setting.

Methods: We conducted a retrospective analysis of single agent B-treated patients with mRCC in the second/third line setting, and identified 21 subjects who met inclusion criteria. The primary endpoint was progression-free survival (PFS). Baseline characteristics, survival, response efficacy outcomes, and toxicities were assessed and summarized.

Results: 21 patients (15 women/6 men) were treated with B at a dose of 5 mg/kg/week, dosed q2-3 weeks. Median age was 63, 80% were white, 14% black; 80% had clear cell histology. Median time from diagnosis to B therapy was 3 years (range 1-18); 100% had prior VEGF TKI therapy; 43% had prior mTOR inhibitor; 43% had prior IFN and 19% prior IL-2; median number of prior therapies was 3 (range 1-7); 100% were considered Motzer intermediate risk. Median PFS on B for all subjects was 4.4 mo (95% CI 2.8-9.6) and median OS was 19.4 mo (95% CI 9.9-NR) from start of B therapy. ORR was 2 CR/PR (9.5%), 11 SD (52%), 5 PD, 3 NE. For subjects treated with prior VEGF and mTOR inhibitors, median PFS and OS were 4.4/13.2 mo. Toxicities were as expected and severe adverse events included grade 3-4 fatigue (6), grade 3-4 dehydration (5), and grade 4 failure to thrive (2), grade 4 constipation (2), and grade 3 muscle weakness (2).

Conclusions: Single agent B therapy has acceptable toxicity and moderate disease stabilizing activity in selected patients with mRCC who have failed prior VEGF TKI and mTOR inhibitor therapy, and suggests a benefit to continued ongoing VEGF inhibition. Further prospective study of B alone, in combination with mTOR inhibition, or with alternative targeted agents is warranted.