199 Locke PGR
199.1 Summary
- Who has hypercapnia and how much should we care?
- What makes a problem worth researching?
- How should we study these patients?
- Shared Management
- Would GLP-1RA work in patients with obesity and hypercapnia regardless of cause?
- What are the most important (population) causes of Hypercapnia?
- What do we know about sleep?
- Analogy to decompensated cirrhosis: is hypercapnia decompensated ventilation?
- Beware of extrapolation: who counts is tricky
199.2 Slide outline
199.2.1 Slide 1
- Who has hypercapnia and how much should we care?
- BRIAN LOCKE PGR
- (Title Redacted)
- This work is supported by the American Thoracic Society ASPIRE Fellowship and the National Institutes of Health under the Ruth L. Kirschstein National Research Service Award 5T32HL105321
- Disclosure: I hold a financial stake and advise a local startup called Mountain Biometrics (focused on data processing of wearable sensor data) ### Slide 2
- What makes a problem worth researching?
- INT Framework (World Health Organization and other grant-makers)
- Importance of the problem:
- How many people (will be) affected by the problem, and how intensely?
- Neglected-ness
- How much effort is already going toward solving the problem? (is it going to be solved anyway? Will a bit more effort help?)
- Tractability
- Is the problem solvable? (What has prevented others from solving it?)
- Hypercapnic Respiratory Failure scores high in all 3 ### Slide 3
- How should we study these patients?
- Paradigm
- Disease-Based
- Syndrome-Based
- Pros
- The causative disease is usually an effect modifier for treatments → homogenous response better for efficacy trials.
- Funders often interested in specific diseases
- Hypercapnia is often recognized before the cause; outpatient workup, so initial management is empiric.
- Efficacious management shares features.
- Enables assessment and improvement of processes of care.
- Cons
- Difficult to extrapolate results to patients with multiple contributing conditions.
- Evidence only applies to definitely-evaluated patients
- May distract from better understanding the way different cases differ
- Lumping vs Splitting (e.g. Sepsis, ARDS) ### Slide 4
- Shared Management
- Annals ATS 2024 ### Slide 5
- Would GLP-1RA work in patients with obesity and hypercapnia regardless of cause?
- 10 RCTs (4 surgical; 6 med. n854)
- Study-level meta-regression
- 1% weight loss → 0.45 less AHI event/hr
- How often does obesity or OSA contribute to hypercapnic respiratory failure? ### Slide 6
- What are the most important (population) causes of Hypercapnia?
- 1 hospital serves population of Liverpool, AUS.
- Random sample of ED cases (PaCO2 ≥ 45 mmHg) and stratified sample of community non-cases.
- Assess: Med Use, BNP, Spiro, HSAT, SNIP
- Estimate Population Attributable Fraction: cases that would be prevented if factor eliminated.
- Fatal problems: 10% got HSAT; causal pathways
- (Spiro)
- (BNP)
- (HSAT)
- (SNIP) ### Slide 7
- What do we know about sleep? ### Slide 8
- Analogy to decompensated cirrhosis: is hypercapnia decompensated ventilation?
- (more properly, acute/chronic liver failure- as cirrhosis is the pathology, not the syndrome)
- Consider: Liver failure requires shared management regardless of disease cause.
- Population-standardized prevalence PaCO2 > 45 mmHg after excluding iatrogenic causes:
- 150 per 100,000 person/year
- Only inpatients, only those with blood gasses checked
- Comparison to decompensated cirrhosis: 94.9 per 100,000 person-year (2017)
- Inpatient ICD code for Hypercapnic Respiratory Failure
- 23% 30-day readmission rate (2 in 3 with recurrence of hypercapnia)
- Comparison to decompensated cirrhosis 12.9% in 30d, 21.2% in 90d in 2012 ### Slide 9
- Beware of extrapolation: who counts is tricky
- Data Source for several studies:
- Encounters: Jan 1 to Dec 31, 2022
- TriNetX research network: EHR data from 76 medical centers, roughly 115 million patients
- Enriched data set mirroring the spectrum of patients a good provider should consider that hypercapnia might be present
- One of: Predisposing condition (incl. BMI > 40), any resp failure, received ventilatory support, received ABG or VBG (regardless of result)
- Data quality check: each encounter must have 1 of each data type ### Slide 10
- Beware of extrapolation: who counts is tricky ### Slide 11
- Beware of extrapolation: who counts is tricky
- Structured, AI-assisted review identified 14 studies pertaining to the prognosis of any-cause hypercapnic respiratory failure.
- Inclusion crit: Variety of ABG, VBG, ICD, & procedure codes used
- TriNetX Emulation: Limited overlap, ICD codes very insensitive. ### Slide 12
- Beware of extrapolation: who counts is tricky
- If you have hypercapnia, context influences the likelihood of Hypercap RF Label
- Cohorts identified by the various case definitions varied widely. E.g:
- 16% Black using Wilson to 26% using Calvo
- 43% receiving a COPD diagnosis code in Meservey to 23% using Adler, Thille, and Cavalot
- Case definitions requiring a procedure code for ventilatory assistance → more patients who received critical care services, and died
- Crux of the problem: Many patients who have hypercapnia don’t get an ABG. P(ABG) varies. ### Slide 13
- Epidemiologic Theory
- What is the effect of an exposures (E) on an outcomes (O)
- Often, a confounder (influences both E and O) distorts the relationship.
- If there is an association between E and O, there are 4 explanations:
- Chance
- Confounding
- Bias
- Selection
- Measurement
- A real effect
- Argument for cause is the disjunctive syllogism:
- “When you have eliminated the impossible, whatever remains, however improbable, must be the truth.”
- If you can’t reliably identify who has hypercapnia, you can’t study its impact on outcomes, control for it in studies of other things, or use it as an outcome. ### Slide 14
- How often is respiratory depression reported in clinical trials?
- Severe adverse drug events (SAE): death, hospitalization, prolongation of hospitalization, disability, incapacity, or situations where action is required to prevent one of the preceding events. [no attribution to intervention]
- ADEs must be submitted to clinicaltrials.gov.
- Shi & Du used BioBERT to extract arm-level data, including SAEs
- 6,429 trials (summarizing data from 4,863,170 participants) reported outcome & adverse even
- Additional 5,300 trials (4,097,827 participants) uploaded at least one adverse event.
- Categorized: plausible (e.g. ‘respiratory failure’) or definitive (e.g. ‘hypoventilation’)
- 3.8% of completed trials, and 2.6% of partially reported trials reported 1+ definitive event.
- Mostly NOT medications known to suppress resp. drive: opioids, benzos, gabapentinoids
- Dustin Anderson-Bell & Brian Locke
- CHEST 2024 abstract ### Slide 15
- How can we improve identification? Make better use of the data elements we have
- Venous Blood Gas
- ABG: PaCO2 ≥ 45 mmHg calendar day of admit
- VBG: VBG PCO2 ≥ 50 mmHg calendar day of admit
- VBGcorr (Farkas): If VBG SO2 < 90%, then PCO2 - 0.2 (diff. in A-V O2 sats)
- How much are VBGs currently used?
- Patients with an ICD for Hypercap. RF
- Admit Day:
- No VBG
- Had VBG
- No ABG
- 16.6%
- 13.2%
- Had ABG
- 39.9%
- 30.3%
- Association of evidence of hypercapnia with clinical course
- Test showing ↑ CO2
- ABG
- VBG
- VBG Corr
- OR of Hypercapnic RF ICD
- 6.5
- 5.6
- 6.1
- Absolute Increase (%)
- 10.3%
- 9.8%
- 10.2%
- OR of subsequent events during hospitalization
- Steroid
- 1.51
- 1.13
- 1.2
- Crit Care
- 1.83
- 1.74
- 1.86
- NIV or IMV
- 2.6
- 2.02
- 2.29 ### Slide 16
- How can we improve identification? Make better use of the data elements we have
- Paired Serum HCO3- (BMP) and PaCO2 (ABG) measurements
- Optimal HCO3- Threshold (Youden Index)
- Ambulatory
- ≥ 27 mEq/L
- Se: 60.06%; Sp: 83.55%
- Emergency
- ≥ 26 mEq/L
- Se: 59.19%; Sp: 75.37%
- Inpatient
- Se: 54.44% Sp: 80.97%
- Critical Care
- ≥ 25 mEq/L
- Se: 57.07% Sp: 76.30% ### Slide 17
- TODO: No text extracted from this slide. ### Slide 18
- TODO: No text extracted from this slide. ### Slide 19
- TODO: No text extracted from this slide. ### Slide 20
- A list of important answered questions in Hypercapnic Respiratory Failure
- What are the reasons providers do or don’t order blood gasses?
- What are the most common causes of hypercapnia?
- What proportion of patients identified as inpatients are referred for and receive definitive evaluation (e.g. PFTs, Sleep Study, Echo)?
- What proportion of inpatients with hypercapnia are recognized?
- What happens to the patients who are not recognized?
- Is hypercapnia a predictive or causal agent of harm?
- Are VBGs a sufficient surrogate for ABGs?
- Should we be aggressively identifying hypercapnia (or will this lead to overdiagnosis)?
- What are the burdens of disease associated with any-cause hypercapnic respiratory failure?
- How should we manage multifactorial hypercapnic respiratory failure?
- … etc … etc … etc ..
- Scoping the research is challenging. Where to start? ### Slide 21
- K23: Bioinformatics Tools to Quantify the Disease Burden Associated with Hypercapnia
- Gaps:
- Identification of hypercapnia appears to be idiosyncratic, which is a problem for patient care and science.
- Admissions with hypercapnia are associated with substantial health burdens, but estimates are from single-centers, rely on local identification approaches, and may not extrapolate.
- It is unclear if the occurrence of hypercapnic respiratory failure is a general marker of illness, a specific marker of respiratory frailty, or a causal agent of harm. ### Slide 22
- K23: Bioinformatics Tools to Quantify the Disease Burden Associated with Hypercapnia
- My goal is to be an expert in chronic respiratory failure that uses clinical research informatics to enable embedded clinical studies of this population
- Career Development Objectives:
- Build experience characterizing care processes and variation
- Informatics deployment: Gain expertise and experience deploying informatics tools (LLMs and Risk Models) for use in clinical research
- Learn Clinical investigation: Learn to lead prospective studies ### Slide 23
- Aim 1: Characterize Current Diagnosis and Workup of Hypercapnic Respiratory Failure
- Problem: Many patients are identified as inpatients at decompensation but need outpatient workup and management. There is ~no guidance.
- Gap: There is no modern estimates of how well we diagnose or evaluate patients presenting with hypercapnic respiratory failure
- Hypotheses:
- There is substantial variation by institution, provider (training), and patient factor in who receives ABG, VBG, diagnosis, and workup
- Most patients are identified as having hypercapnia do not receive definitive workup (PFTs, PSG, pulmonary referral) ### Slide 24
- Aim 1: Characterize Current Diagnosis and Workup of Hypercapnic Respiratory Failure
- Solution:
- Linkage of Intermountain Health medical records and UT All Payor Claims Database, which contains information on follow-up
- Includes (outpatient) testing, referrals, readmissions, deaths
- All patients with hypercapnia identified by VBG, ABG, ICD code, and risk modeling (experimental aim: augmented with NLP)
- Multi-level logistic regression modeling to characterize factors influencing test ordering
- Summary of referral for testing, receipt of testing, readmission, death. ### Slide 25
- Aim 2: Augment and Locally Validate Hypercapnia Diagnostic Modeling
- Problem: Methods of identifying hypercapnia are haphazardly applied, inconsistent, and not patient-centered.
- Gap: Reliable, routine methods for identification of hypercapnia are needed.
- Hypothesis: The likelihood a patient has hypercapnia can be estimated from routinely obtained data elements, like labs and notes. ### Slide 26
- Diagnostic Risk Modeling
- Solution: Diagnostic Modeling (Diagnostic Risk Modeling ~ Clinical Decision Rule ~ Prediction Modeling, Diagnostic nomogram)
- Wells score is a logistic regression ### Slide 27
- Diagnostic Risk Modeling
- Estimate the likelihood of hypercapnia at admission in inpatients.
- Predictors: must be routinely available, obtained before/independently of blood gas
- 2 approaches: LASSO Logistic Regression; Random Forest
- Assess distribution of predictions, discrimination, calibration
- Partial verification: vs any blood gas (ABG > 45, VBG > 50); inverse propensity weighted ABG
- Predictors: age, sex, body mass index [BMI], components of basic blood chemistry testing [sodium, potassium, chloride, bicarbonate, urea, and creatinine], and hemoglobin. Selected a priori on clinical relevance and availability. ### Slide 28
- Partial Verification: you can only train on data with outcomes available. ### Slide 29
- Aim 2: Augment
- Problem: model accuracy is imperfect
- Gap: Symptoms, which greatly improve prediction accuracy, are in notes.
- Hypothesis: extracting this data using locally-deployed LLMs will improve the accuracy of predictions about the risk of hypercapnia ### Slide 30
- Aim 2: Augment Model
- Solution: Large language models enable transfer learning (more on this next week).
- Extract: Chief concern, confusion, dyspnea, etc. from nursing triage note.
- Use as predictor in existing model
- Assess vs limited annotated corpus, but mainly for impact on model. ### Slide 31
- Aim 2: Validate Model
- Problem: We’re interested in people who haven’t received blood gas verification. How well will it work in practice?
- Gap: Research consent for ABGs will be biased (and patients with hypercapnia are confused)
- Hypothesis: Risk prediction modeling can identify a group of patients highly likely to have hypercapnia. ### Slide 32
- Is TcCO2 usable for hypercapnia ascertainment?
- Conway et al.: 7021 paired measurements taken from 2817 participants in 73 studies
- mean bias: 0.085 mmHg
- within study SD(agreement): 3.51 mmHg
- In-silico simulation of paired TcCO2 measurement for each PaCO2 obtained in 2022 via TriNetX.
- Sensitivity
- Specificity
- Neg. Pred. Val.
- Pos. Pred. Val
- 86.8%
- 93.7%
- 85.7%
- 94.2% ### Slide 33
- Aim 2: Validate Model
- Solution: Validate model predictions using TcCO2 on newly admitted adults to IMED medical floors.
- Protocol:
- Admit to medicine – apply TcCO2 within 24 hours to patients with predicted high likelihood of hypercapnia
- N63 patients for power to approximate PaCO2 +/- 12.5 mmHg in patients predicted 50+% likely.
- Deferred (vs waived) consent ### Slide 34
- Outcomes of this work:
- Three novel contributions:
- Estimate current use (and variation) of diagnostic surrogates like venous blood gas in a large integrated health system.
- Improve estimates of the prognostic significance of any-cause hypercapnia
- Establish ABG & VBG independent methods of enrollment
- Potential weaknesses: reliance on diagnostic modeling and transcutaneous CO2 sensors. Broad Scope
- Strengths: extensive preliminary modeling, contingency plans,
- Development of an informaticist-investigator with technical expertise in rapidly progressing subfield of data-science tools, such as LLMs. ### Slide 35
- Acknowledgements:
- UU PCCM Collaborators: Krishna Sundar, Jeanette Brown
- UU PCCM Advisors: Rob Paine, Barb Jones
- UU DBMI Collaborators: Ram Gouripeddi, Wayne Richards
- IMED PCCM Advisors/Collaborators: Ithan Peltan, Sam Brown, ACR Research Group
199.3 Learning objectives
- Who has hypercapnia and how much should we care?
- What makes a problem worth researching?
- How should we study these patients?
- Shared Management
- Would GLP-1RA work in patients with obesity and hypercapnia regardless of cause?
199.4 Bottom line / summary
- Who has hypercapnia and how much should we care?
- What makes a problem worth researching?
- How should we study these patients?
- Shared Management
- Would GLP-1RA work in patients with obesity and hypercapnia regardless of cause?
199.5 Approach
- TODO: Outline the initial assessment or decision point.
- TODO: Outline the next diagnostic or management step.
- TODO: Outline follow-up or escalation criteria.
199.6 Red flags / when to escalate
- TODO: List red flags that require urgent escalation.
199.7 Common pitfalls
- TODO: Capture common errors or missed steps.
199.8 References
TODO: Add landmark references or guideline citations.