254 RIP Hypercapnia

254.1 Summary

Hypercapnia and Sleep-Disordered Breathing: Research(es) In Progress
Overview
Research Motivation: Why OSA & CO2?
Principle Components of EHR data for Identifying Hypercapnic Respiratory Failure
Chung Y, Garden FL, Marks GB, Vedam H. Population Prevalence of Hypercapnic Respiratory Failure from Any Cause. Am J Respir Crit Care Med. 2022 Apr 15;205(8):966-967.
Multicausality
Chung Y, Garden FL, Marks GB, Vedam H. Causes of hypercapnic respiratory failure and associated in-hospital mortality. Respirology. 2022
Giulia Cavalot, Vera Dounaevskaia, Fernando Vieira, Thomas Piraino, Remi Coudroy, Orla Smith, David A. Hall, Karen E. A. Burns & Laurent Brochard (2021) One-Year Readmission Following Undifferentiated Acute Hypercapnic Respiratory Failure, COPD: Journal of Chronic Obstructive Pulmonary Disease, 18:6, 602-611
Vonderbank S, Gibis N, Schulz A, Boyko M, Erbuth A, Gürleyen H, Bastian A. Hypercapnia at Hospital Admission as a Predictor of Mortality. Open access emergency medicine : OAEM 2020; 12: 173-180.
Possible use
Admitted (ICU or floor) with one of the following diagnostic codes:
Why not just use PaCO2 > 45 mmHg?

254.2 Slide outline

254.2.1 Slide 1

Hypercapnia and Sleep-Disordered Breathing: Research(es) In Progress
Brian Locke MD
PCCM Fellow; T32; MSCI, etc.
RIP 10/20/22 ### Slide 2
Overview
Conceptual framing
Principle Components of EHR data for Identifying Hypercapnic Respiratory Failure: Brian Locke, Wayne Richards, Jeanette Brown, Krishna Sundar, Ram Gouripeddi
Other projects (rapid-fire, if time):
Test Characteristics of HCO3- with various causes of hypercapnia: (same group)
HCO3 Kinetics After OSA Dx and Treatment: Analisa Taylor, Brian Locke, Heather Howe, Kimberly Workman Krishna Sundar
Factors predicting Central Sleep Apnea CPAP Trial vs. Likelihood of Success: Brian Locke, Jeff Sellman, Jonathan McFarland, Franscisco Uribe, Kimberly Workman, Krishna Sundar
Burden of OSA Among Native Hawaiian Pacific Islanders: Brian Locke, Divya Sundar, Darin Riyujn
Effect of Weight Loss Interventions on OSA Severity, SRMA: Brian Locke, Tee Nonthasawadrsi, Ainhoa Gomez Lumbreras, Nui Chaiyakunapruk ### Slide 3
Research Motivation: Why OSA & CO2?
Epidemiologic Research on OSA and Hypercapnia is a quagmire
Assessing degree of ’exposure’ to OSA is inaccurate: control groups often contaminated; AHI is not a reliable surrogate for severity
Deleterious consequences of OSA on ventilatory control are extremely plausible and modifiable
Ventilatory control is malleable (breath hold divers, inspired CO2 studies)
Except perhaps sleepiness, much more direct consequence than other OSA sequalae
Why UT? Not as easily recognized at sea-level (yet; CO2 monitoring becoming commonplace..)
We have made progress reducing exposures for other lung disease (pollution, smoking), but we are not yet making progress on risks for ventilation control abnormalities (obesity, opiates)
Patients prone to ventilatory decompensation are not currently systematically identified or managed, and so frequently fall through the cracks. Identification and treatment of sleep disordered breathing and its risk factors is a promising approach to improve care but requires empiric support to justify implementation. ### Slide 4
Principle Components of EHR data for Identifying Hypercapnic Respiratory Failure
Overall research hypotheses: Epidemiologic research on hypercapnic respiratory failure has been limited and hard to interpret. Reasons include:
Reliance on EHR-based cohorts
Suboptimal cohort definitions
Conceptual ambiguity
EHR data can generate more useful findings if better patient identification methods (computable phenotypes) are used. ### Slide 5
Chung Y, Garden FL, Marks GB, Vedam H. Population Prevalence of Hypercapnic Respiratory Failure from Any Cause. Am J Respir Crit Care Med. 2022 Apr 15;205(8):966-967.
Australia, 1 regional hospital services the district
Identified by initial ABG (w/n 24h) PaCO2 over 45, excluded iatrogenic causes/sedation. N891 people, 1135 blood gasses (repeat hosp.)
Normalized rates of hypercapnia to population demographics
150 per 100,000 person/year.
Compared to Age 45-54:
RR 55-64: 2.1
RR 65-74: 6.2
RR 75-84: 15.7
RR 85-94: 26.2 ### Slide 6
Multicausality
Obesity
Untreated
Sleep Apnea
COPD
Loop diuretic
Opiates
Sufficient cause: a set of factors that will cause disease when present
Component cause: a factor that, if not present, no disease would occur
Necessary cause: in all sufficient sets, this component must be present
Hypothetical patient with hypercapnic respiratory failure ### Slide 7
Chung Y, Garden FL, Marks GB, Vedam H. Causes of hypercapnic respiratory failure and associated in-hospital mortality. Respirology. 2022
Further analysis of same data
Diagnostic codes: 52% Charlson Comorbidity 5+
Overall in-hosp mortality 12.8%, 55% with acute acidosis
5% dx’d obesity; 6% dx’d OSA – not credible
Prevalence-based estimates: NMD under-represented ### Slide 8
Giulia Cavalot, Vera Dounaevskaia, Fernando Vieira, Thomas Piraino, Remi Coudroy, Orla Smith, David A. Hall, Karen E. A. Burns & Laurent Brochard (2021) One-Year Readmission Following Undifferentiated Acute Hypercapnic Respiratory Failure, COPD: Journal of Chronic Obstructive Pulmonary Disease, 18:6, 602-611
All adult ED visits @ single Toronto ED w ABG <7.35/45+ OR VBG <7.34/50+ & CEDIS resp code for resp symptoms. Charts reviewed for 12 months.
Regression for risk of readmission: etiology not independently associated ### Slide 9
Vonderbank S, Gibis N, Schulz A, Boyko M, Erbuth A, Gürleyen H, Bastian A. Hypercapnia at Hospital Admission as a Predictor of Mortality. Open access emergency medicine : OAEM 2020; 12: 173-180.
All patients with dyspnea or pulm disease admitted to hospital received capillary blood gas (some screening with VBG). Stratified by pH > 7.35 or pH < 7.35
Hospital specializing in lung disease (unclear referral pattern): 6750 admissions, 2710 with dyspnea or lung dz, 1626 normocapnic, 588 hypercapnic.
32% 1yr mortality. Varied by type. Comparisons all seem very limited by collider bias. ### Slide 10
Possible use
https://jcsm.aasm.org/doi/full/10.5664/jcsm.9962
It seems like literature is coalescing on ABG – though with additional parameters.
need – clarify the interpretability of these? ### Slide 11
Admitted (ICU or floor) with one of the following diagnostic codes:
J96.02 (acute hypercapnic respiratory failure)
J96.22 (acute and chronic respiratory failure with hypercapnia)
J96.92 (respiratory failure unspecified with hypercapnia)
J96.12 (chronic respiratory failure with hypercapnia)
E66.2 (morbid obesity with hypoventilation)
30d Readmission rate: 23% (2/3 recurrence)
~ CHF. > than MI, AECOPD, PNa
Admitted to hospital (Floor or ICU) with ABG showing PaCO2 over 45 mmHg and pH 7.35-7.45
COPD (2/3)
No COPD (1/3)
AHI Median
[IQR]
31.9
[14.3, 45.6]
66.0
[48.0, 83.8]
AHI > 5 present
66%
94%
AHI > 15 present
51%
81%
1. Admitted to the ICU
1. PaCO2 greater than 47.25 mmHg
1. Procedure code for non-invasive ventilation or invasive mechanical ventilation initiation
Admissions with hypercapnia indicate high risk of morbidity; may be driven by treatable conditions ### Slide 12
Why not just use PaCO2 > 45 mmHg?
ABGs are infrequently obtained
People who get ABGs are different from people who don’t.
If you only include people who get confirmatory ABG, you will both miss many patients, and get an unrepresentative sample.
Study design 101:
To test interventions: homogenous group, high risk of the outcome, likely to respond similarly to intervention
Requiring PaCO2 > 45 (or 52) mmHg not a problem
To understand incidence/prevalence, and improve processes of care: sensitivity of your criteria matter
Classic risk factors, unambiguous presentation, and severe disease over-represented: ↑ P(ABG obtained)
Underestimate (total) burden, oversimplify population, exclude patients that might benefit
To understand prognosis: variations in P(ABG obtained) matters ### Slide 13
What is ‘hypercapnic respiratory failure’?
Problems with PaCO2 > 45 mmHg (sea-level, current def)
97.5 Percentile of ‘normal’?
“Hypercapnic respiratory failure is a syndrome”; PaCO2 Lab finding (hypercapnia). Gap?
Syndrome (Greek: ‘con-currence’) a set of signs and symptoms that together suggest a common cause.
If there is no agreed-upon reference standard test: perhaps ‘construct’ is a better term.
‘Ventilatory Failure’: can the respiratory system meet the body’s demand
Is a patient with OHS (‘submissive hypercapnia’) meeting the body’s demand?
2 instrumental uses: Classification: “Hyperlipidemia” vs Prediction: “10-y ASCVD risk”
Predicting future risk of ventilatory failure: ‘ventilatory frailty’ (prognosis)
Predicting differential treatment response (management)
Ideal Definition: Who is at risk for future/ongoing ventilatory failure? Who will respond to treatments? ### Slide 14
Why not just use PaCO2 > 45 mmHg?
Criteria for a good definition:
Reliable (identifies only & all) –
Bayes: P(Pre-test) ↑ : Required strength of evidence ↓ for same P(Post-test)
Feasible (usual data elements) – ABGs frequently not obtained; HCVR, MVV…
Valid (truly measures what we’re interested in)
Possible to have PaCO2 < 45 and not meet bodies demand (e.g. metabolic acidosis)
Possible to have ‘hypercapnic success’ (e.g. metabolic alkalosis, submissive hypercapnia)
People who have high risk of future/ongoing death, readmission, or activity limitation from ventilatory insufficiency have the criteria. People with low risk do not.
Consensus… ### Slide 15
Review: Common Methods of Identifying Hypercapnic Respiratory Failure Produce Meaningfully Different CohortsBrian Locke, Krishna Sundar, Jeanette Brown, Ramikiran Gouripeddi
Hypothesis 1: The methods used in prior studies of hypercapnic respiratory failure (billing code-, procedural code-, and blood-gas-based criteria) identify different patients.
Outcome: Relative Sensitivity; Positive Predictive Agreement
Hypothesis 2: The cohorts created by these differ methods differ in risk for outcomes of interest, which hampers interpretation of these studies.
Outcome: distribution of age, ethnicity, BMI, and frequency of coexisting diagnoses (OSA, opiate use disorder, COPD, CHF, and neuromuscular disease)
69 Million MRNs aggregated from 50 academic medical centers
Deidentified patient level data, including admissions, diagnoses, medications, procedures, and lab values.
Missing data? ### Slide 16
Preliminary Results
ICD Group:
Relative sensitivity (vs ABG): 19.8%
Positive Predictive Agreement (vs ABG): 47.0%
NIV Group:
Relative sensitivity (vs ABG): 15.2%
Positive Predictive Agreement (vs ABG): 45.2%
Common Methods of Identifying Hypercapnic Respiratory Failure Produce Meaningfully Different CohortsBrian Locke, Krishna Sundar, Jeanette Brown, Ramikiran Gouripeddi ### Slide 17
ABG
Group
ICD
NIV
Age
62±18
65±16
62±17
% Female
46%
51%
42%
% white
66%
71%
65%
% Black
18%
19%
17%
BMI
30.4±8.3
33.1±10.3
29.1±8.2
% with CHF
37%
30%
% with COPD
31%
14%
% Opiate UD
6%
3%
% Sleep Apnea
23%
24%
10%
Common Methods of Identifying Hypercapnic Respiratory Failure Produce Meaningfully Different CohortsBrian Locke, Krishna Sundar, Jeanette Brown, Ramikiran Gouripeddi
This is a vague description and doesn’t encapsulate data about how they decompensated at time of Hypercap. RF
How could we do better? ### Slide 18
Principle Components of EHR data for Identifying Hypercapnic Respiratory Failure: Study Aims
Currently, all research on hypercapnic respiratory failure focuses on isolated consideration of 3 types of data: Lab Results, Comorbid Conditions and Outcomes
There are features we know are of interest medically that are not captured in current definitions or descriptions
High vs low drive to breath
Patterns of comorbidities
Patterns of management
Gap: Is there enough data in EHRs to make good guesses about who has hypercapnic respiratory failure when PaCO2 not measured?
If you were going to propose ways to identify or characterize patients with hypercapnic respiratory failure in EHR data, what features would you use? ### Slide 19
Methods: Principal Component Analysis
Modeling and explained variability: linear regression
R2 (Percentage of variation explained) 0.01
R2 (Percentage of variation explained) 0.70
R2 (Percentage of variation explained) 0.98
ANOVA meaning ### Slide 20
Methods: Principal Component Analysis
Height vs Weight
Perfectly correlated: if you know height, you know the weight.
R2 100%
You only need to know 1 value to describe each data point ### Slide 21
Methods: Principal Component Analysis ### Slide 22
Methods: Principal Component Analysis
PC1: Skeletal size
PC2: Soft tissue size ### Slide 23
Methods: Principal Component Analysis
PC1: Skeletal size
PC2: Soft tissue size
(This is a story I made up, it is a claim)
PC1: Explains the most variability
PC2: Explains the next most
(This is a data transform. NO CLAIMS MADE) ### Slide 24
Methods: Principal Component Analysis
Dimension Reduction
‘High dimensional data’: many characteristics (100s) on many data points (100,000s)
Genetic analyses, ‘-omics’
Complex signals (e.g. PSG data)
EHR data
Problematic for 2 reasons:
Computationally hard to run analyses
Statistically, inadvisable to run analyses (overfitting)
Application: if you want to summarize the ways data points vary, but using the minimum number of characteristics
What variables are most informative in linear combination, for our purposes? (data description)
From data not physiology: no guarantee the pattern represents a ‘physiologic’ thing ### Slide 25
Methods: Principal Component Analysis
Hypothetical Data: Serum [HCO3-] vs PaCO2
PC1: Degree of (Acute) CO2 retention
PC2: Degree of Metabolic Compensation ### Slide 26
Aim: Identify the principal components among patients with Hypercapnic RF
Why?
If only a few data axes describe most the variability, computational phenotypes without a reference standard will be harder
Allows us to understand which types of features vary between patients with hypercapnic RF
Assess differences between currently used methods of identification
Propose features to use in better identification methods
Principal components would be used for inputs into, clustering algorithms / Classifiers (i.e. ‘machine learning’) ### Slide 27
Aim: Identify the principal components among patients with Hypercapnic RF
Patients: Met ICD code, ABG, or Procedure code-based definition for hypercapnic respiratory failure during 1st admission.
Method: Generate and operationalize a list of candidate features
Demographics (Age, Sex, Ethnicity/Race)
Lab Test (ABG, VBG, BMP, Hgb, LVEF)
Diagnostic codes during the admission (e.g. Heart Failure I50., Status Asthmaticus (J46.), Guillain-Barre(G61.), etc. etc.
Procedures (NIV, IMV, Nerve Conduction Studies, Inhaler Teaching, CXR, Critical Care Services, etc.)
Prior Diagnoses (OSA G47.3, Dementia F01-9., Spinal Cord Injury G95)
Inpatient Medications (Corticosteroids, Naloxone, Bronchodilators, Diuretics)
Prior Outpatient Medications (Opiates, Diuretics)
Vital signs (first recorded RR-14, recorded BMI-22, etc.) ### Slide 28
Preliminary Results:
Dataset: 925,512 patients from 70 Healthcare organizations since 2017. 20 gb of data.
1.2 Billion ‘facts’ in medications alone.
Center for High Performance Computing protected environment
Limited ”test data” set with ~100 mb worth of data (sparse)
Test scripts prior to main run in coming weeks. ### Slide 29
Preliminary Results
Most common dx’s Pneumonia, Dyspnea NOS, AECOPD, Long term use of opiates
Most common procedures: Xray, CPAP, treatment with inhalers
Most common pre-existing: Dementia, OSA, Secondary Polycthemia
Inpatient medications: Steroids > Narcan > Vasodilators
Outpatient medications: Opiates > Loop diuretics > bronchodilators > muscle relaxers. ### Slide 30
Preliminary Results ### Slide 31
Next steps:
Further proofreading that our encoding/capture of data features is accurate.
Run the PCA on the full dataset and evaluate whether (or not) there are physiologic correlates to most important principal components
Evaluate whether the patients identified by the 3 epidemiologic methods in current use (ABG, procedures, dx-based) differ based on the principal components ### Slide 32
Issues: inferences from data
Research degrees of freedom:
What candidate features do we include? How do we represent them?
Do principal components correspond to helpful things? Could you test it?
If you did it on a slightly different data-set, would you find the same thing?
Would they persist through time? Do people you expect to have certain features have them?
Big Data: ~roughly 1/4 of all people with hypercapnic RF in the USA
Aus prevalence (150 per 100,000 per year 500,000 people with hypercapnia in USA per year)
“Which one should I trust more: a 1% survey with 60% response rate or a self-reported administrative dataset covering 80% of the population?” Xiao-Li Meng. Depends on 3 things
Data quality measure: how non-random is the data included
Data quantity measure: how much of the population of interest is in the dataset
Problem difficulty measure ### Slide 33
Thoughts?
Future Aims:
Include patients without hypercapnic respiratory failure and define the principle components that differ between patients with with w/o hypercapnic RF
Evaluate which features correlate with readmission risk or mortality risk
Generate a reference standard dataset to test performance of various enrollment criteria
Use the TriNetX database to develop a classifier that uses features available in the NCUPS National Readmission Database (comorbidities, diagnoses) to identify hypercapnic respiratory failure (by traditional criteria), then apply for estimate of nationwide readmission burden (rates, costs). ### Slide 34
Rapid Fire Round
Test Characteristics of HCO3- with various causes of hypercapnia: (same group)
HCO3 Kinetics After OSA Dx and Treatment: Analisa Taylor, Brian Locke, Heather Howe, Krishna Sundar
Factors predicting Central Sleep Apnea CPAP Trial vs. Likelihood of Success: Brian Locke, Jeff Sellman, Krishna Sundar
Burden of OSA Among Native Hawaiian Pacific Islanders: Brian Locke, Divya Sundar, Darin Riyujn
Effect of Weight Loss Interventions on OSA Severity, SRMA: Brian Locke, Tee Nonthasawadrsi, Ainhoa Gomez Lumbreras, Nui Chaiyakunapruk ### Slide 35
Test Characteristics of HCO3- with various causes of hypercapnia: Brian Locke, Jeanette Brown, Ram Gouripeddi, Krishna Sundar
Gap: [HCO3-] has been validated for use in stable outpatients with suspected OHS undergoing sleep apnea testing.
Hypothesis: [HCO3-] won’t work as well to exclude hypercapnia in other settings
Methods: TriNetX database. 1.9 Mil patients with ABG and BMP on same day. Classified as
True Positive (BMP HCO3 27+, PaCO2 45+)
TN (BMP HCO3 <27, PaCO2 <45)
FP (BMP HCO3 27+, PaCO2 <45)
FN (BMP HCO3 <27, PaCO2 45+)
Conclusion: Most helpful excluding
COPD and CHF; diagnosing NMD
TODO: stratify by location
(ER, ICU, etc). ROC curves ### Slide 36
HCO3- Kinetics After OSA Dx and Treatment: Analisa Taylor, Brian Locke, Heather Howe, Kimberly Workman, Krishna Sundar
Gap: The ERS defines OHS on a spectrum: At-risk (OSA+Obesity), Nocturnal Hypoventilation (↑ HCO3-), then hypercapnia (↑PaCO2). Unclear validity
Hypothesis: Bicarbonate can identify patients with ERS Stage 1-4 OHS, and monitor treatment response
Methods: n52 obese pt undergoing SDB eval; obtained ABG as part of defunct study on sex hormones on OHS. 𝚫 HCO3-, 𝚫BMI, PAP compliance, rate of drugs influencing acid-base balance.
Results: n12 with HCO3- over 25 (elev. adj)
n 13 with PaCO2 over 42.
Se 100%, Sp 88.9%, AUROC 0.944
TODO: Compliance data, multi-level modeling ### Slide 37
Factors predicting Central Sleep Apnea CPAP Trial vs. Likelihood of Success: Brian Locke, Jeff Sellman, Jonathan McFarland, Franscisco Uribe, Kimberly Workman, Krishna Sundar
Gap: Guidelines recommend at least considering a trial of CPAP for all etiology of Central Sleep Apnea. Little is known about who gets a trial vs straight to ASV.
Hypothesis: Factors clinicians use to decide whether to forego CPAP will not correlate with factors predicting CPAP failure in those who get CPAP
Methods: Chart review of all w Dx code for CSA. Extract probable causes, demographics, %CSA
Conclusions: Age, diagnosis by HSAT, and CSA in the context of OSA are not concordant ### Slide 38
Burden of OSA Among Native Hawaiian/Pacific Islanders: Brian Locke, Divya Sundar, Darin Riyujn
Gap: NHPI patients have high rates of sleep symptoms of comorbidities known to predispose to OSA and its consequences. The prevalence or consequence of OSA has not been assessed.
Hypothesis: There will be a skew toward severe disease and high comorbidity burden among NHPI patients diagnosed at U of U reflecting barriers to diagnosis.
Methods: Chart review of NHPI patients dx’d between 2014-21 were reviewed.
Conclusions: Expected skew found, esp in young males. Adherence was low. Low rates of referral.
TODO: NHPI NHIS (or BRFSS) regression (for PHS 7000 MSCI course) on sleep symptoms and snoring/apneas – does OSA explain? ### Slide 39
Effect of Weight Loss Interventions on OSA Severity, SRMA: Brian Locke, Tee Nonthasawadrsi, Ainhoa Gomez Lumbreras, Nui Chaiyakunapruk
Gap: CPAP adherence is generally poor, limiting effectiveness. Guidelines recommend weight-loss, but this seldom occurs. Several anti-obesity meds (AOM) recently approved and cause more weight loss.
Hypothesis: %𝚫weight explains %𝚫AHI sufficient that new AOM can be inferred to be effective in OSA.
Methods: Systematic Review and Meta-Regression of RCTs of bariatric surgery or AOM vs placebo, no tx, or std care & report 𝚫weight and 𝚫AHI
Results: 4 RCTs of Bariatric Surgery, 8 RCTs of AOM (5 of GLP-1)
Conclusions: TBD (weekly semaglutide for obesity-related ventilatory dysfunction? Someday…) ### Slide 40
Influence of Ambient Air Pollution on FlowAHI in Patients with Sleep Apnea. Jay Kitt, Krishna Sundar, Brian Locke, Julio Facelli, Ram Gouripeddi
Gap: leave this one out
Hypothesis:
Methods
Results
Conclusions
TODO

254.3 Learning objectives

Hypercapnia and Sleep-Disordered Breathing: Research(es) In Progress
Overview
Research Motivation: Why OSA & CO2?
Principle Components of EHR data for Identifying Hypercapnic Respiratory Failure
Chung Y, Garden FL, Marks GB, Vedam H. Population Prevalence of Hypercapnic Respiratory Failure from Any Cause. Am J Respir Crit Care Med. 2022 Apr 15;205(8):966-967.

254.4 Bottom line / summary

Hypercapnia and Sleep-Disordered Breathing: Research(es) In Progress
Overview
Research Motivation: Why OSA & CO2?
Principle Components of EHR data for Identifying Hypercapnic Respiratory Failure
Chung Y, Garden FL, Marks GB, Vedam H. Population Prevalence of Hypercapnic Respiratory Failure from Any Cause. Am J Respir Crit Care Med. 2022 Apr 15;205(8):966-967.

254 RIP Hypercapnia

254.1 Summary

254.2 Slide outline

254.2.1 Slide 1

254.3 Learning objectives

254.4 Bottom line / summary

254.5 Approach

254.6 Red flags / when to escalate

254.7 Common pitfalls

254.8 References

254.9 Slides and assets

254.10 Source materials