193 Locke Msci

193.1 Summary

Building a Better Cohort of Patients with Hypercapnic Respiratory Failure
Overview
What is Hypercapnic Respiratory Failure?
Hypercapnia is common…
Age
Hypercapnia is associated with high - and likely increasing - morbidity, mortality, and cost
Patients with hypercapnia are identified unreliably and late in their illness
Hypothesis: We Can Identify Them & It Would Help?
Evidence Islands
Recognized Hypercapnia
All Hypercapnia
Disease Paradigm Doesn’t Mirror Practice

193.2 Slide outline

193.2.1 Slide 1

Building a Better Cohort of Patients with Hypercapnic Respiratory Failure
Brian Locke, MD
T32 Fellow. Division of Pulmonary, Critical Care, and Occupational Pulmonary Medicine
Wayne Richards MS, Joseph Finkelstein MD PhD MA, Jeanette Brown MD PhD, Ramkiran Gouripeddi MBBS MS, Krishna M Sundar MDDepartments of Internal Medicine and Biomedical Informatics
This research was supported by the National Institutes of Health under Ruth L. Kirschstein National Research Service Award 5T32HL105321 and the American Thoracic Society Academic Sleep and Pulmonary Integrated Research/Clinical Fellowship (ASPIRE) Program ### Slide 2
Overview
10 slides on the problem
10 slides on what we’ve done
10 slides on what we plan to do ### Slide 3
What is Hypercapnic Respiratory Failure?
𝐴𝑟𝑡𝑒𝑟𝑖𝑎𝑙 𝐶𝑂2 ∝𝑇𝑖𝑠𝑠𝑢𝑒 𝑃𝑟𝑜𝑑𝑢𝑐𝑡𝑖𝑜𝑛 𝑜𝑓 𝐶𝑂2𝑇𝑜𝑡𝑎𝑙 𝑉𝑒𝑛𝑡𝑖𝑙𝑎𝑡𝑖𝑜𝑛 −𝑊𝑎𝑠𝑡𝑒𝑑 𝑉𝑒𝑛𝑡𝑖𝑙𝑎𝑡𝑖𝑜𝑛
Diseases that cause hypercapnia (↑ Arterial CO2)
↑ Requirement
↓ Capacity
↓ Drive
COPD (inefficient ventilation)
Obesity (more CO2 production)
Neuro-musc. disease
COPD
Obesity
Opiates
Metabolic Alkalosis ### Slide 4
Hypercapnia is common…
Common contributors to Hypercapnic Respiratory Failure:
Advanced Age
Opiate Use
Obesity
Advanced Lung Dz
Multimorbidity
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: Decompensated Cirrhosis 94.9 (US, 2017) per 100,000 person-year ### Slide 5
Age
Group
Prevalence
Per 100,000 p.yr.
15-24
14 (9-22)
25-34
29 (22-39)
35-44
42 (33-54)
45-54
80 (66-96)
55-64
196 (71-225)
65-74
517 (463-577)
75-84
1160 (1046-1286)
85+
1712 (1481-1980) ### Slide 6
Hypercapnia is associated with high - and likely increasing - morbidity, mortality, and cost
Inpatient ICD code for Hypercapnic Respiratory Failure
23% 30d readmission rate
Same as CHF, more than MI
66% with recurrence
Decompensated Cirrhosis, for reference ### Slide 7
Patients with hypercapnia are identified unreliably and late in their illness
Published 2003; 3 hospitals, Denver
ABG on all inpatients BMI 35+
31% with unrecognized with hypercapnia ### Slide 8
Hypothesis: We Can Identify Them & It Would Help? ### Slide 9
Evidence Islands
Severe COPD
Obesity Hypovent
ALS
COPD: PaCO2 > 52 mmHg when stable, OSA noncontributory
OHS: PaCO2 > 45 mmHg; No obstruction, AM Co2 up by 7 or PSG
Restrictive (including NMD) ABG 45>, or FVC < 50% or low muscle strength. ### Slide 10
Recognized Hypercapnia
All Recognized Cases
(Multifactorial Causes,
Relapsing-Recurrent)
What portion have applicable evidence-based management?
What portion of patients with hypercapnic respiratory failure qualify for NIV?
Does treatment help the others?
???
Severe COPD
Obesity Hypovent
ALS ### Slide 11
All Hypercapnia
All Recognized Cases
(Multifactorial Causes,
Relapsing-Recurrent)
???
What portion of patients are missed?
Are the patients who are missed consequential?
Severe COPD
Obesity Hypovent
ALS
All Cases ### Slide 12
Disease Paradigm Doesn’t Mirror Practice
Disease-based
Syndrome-based
Understand the physiology
Mirrors Diagnostic Order (↑CO2 often recognized first)
Homogenous Sample for Clinical trials
Different Causes Managed Similarly
Quality Improvement (Presentation → Treatment) ### Slide 13
Non-invasive ventilation is not the only option
Obesity and OSA are contributors to many cases
MDCRC 6200 Systematic Review and Meta-analysis
All RCTs: Pharmacologic and Surgical Weight Loss
Expected effect of Tirzepatide: AHI -13.3 events/hr [-8.9 to -17.7]; 44% improvement.
Decrease of 0.45 (95% CI: 0.18-0.73 events/hour) events per hour for each 1% body weight loss
The Association of Weight Loss from Antiobesity Medications or Bariatric Surgery and Apnea-Hypopnea Index in Obstructive Sleep Apnea: A Systematic Review and Meta-Regression. Brian W. Locke MD; Ainhoa Gomez Lumbreras MD, PhD; Chia Jie Tan PhD; Teerawat Nonthasawadsri PharmD; Sajesh Veettil MSPharm, PhD; Chanthawat Patikorn PharmD; Nathorn Chaiyakaunapruk PharmD, PhD ### Slide 14
Impact of Doing this Research
Patients with hypercapnic respiratory failure are not well served by our scientific knowledge base
Identified cases are at high risk, but often we don’t know what to do
We unreliably diagnose hypercapnia (even when best management practices known)
Better identification of patients with hypercapnia will allow:
Improved care for patients with understood best management
Improve the rigor of research on multifactorial hypercapnia
Allow for hypercapnic respiratory failure as a study outcome
Benchmark current health system performance at diagnosis and treatment ### Slide 15
Approach to the Solution
Hypothesis: Health Record Data contains enough information to helpfully stratify patients with hypercapnic respiratory failure.
Consider:
Example OSA
STOP BANG
Subjective measures:
Loud snoring, tired/fatigued/sleepy, observed apneas? Report of HTN
Objective measures:
BMI, Age, Neck Circumference, Gender
High-risk increased perioperative complications ### Slide 16
Pulse Oximeter vs ???? For CO2.
Hypoxemic (Low O2) Respiratory Failure
Hypercapnic (High CO2) Respiratory Failure
Low arterial blood oxygen tension
High arterial blood carbon dioxide tension
Definitive diagnosis by Arterial Blood Gas
Reliably Identified by Pulse Oximetry
Limited non-invasive testing
Immediate Life Threat if Severe
Indicates potentially treatable excess risk
Applied widely since 1980s with improved recognition
No systematic Identification ### Slide 17
Role of HCO3- in Diagnosis of Hypercapnia
𝐻2𝑂+𝐶𝑂2 ⇋𝐻2𝐶𝑂3 ⇋ H+ + HCO3-
Boston Acid-Base “Rules”:
Acute: ↑ CO2 by 10 mmHg ↑ HCO3- by 1 mEq/L
Chronic: ↑ CO2 by 10 mmHg ↑ HCO3- by 4 mEq/L
Thus,
If CO2 has been ↑ enough, for long enough, and kidneys work well enough: cases of hypercapnia will have ↑ HCO3-
If there is no other cause of metabolic alkalosis, non-cases will not have ↑ HCO3- ### Slide 18
Does HCO3- work in other situations?
Target population: patients whom a clinicians should consider hypercapnia
Spectrum bias/effect
Data source: all encounters in TriNetX research network during 2022 who met:
Predisposing condition: COPD, CHF, Obesity…
Diagnostic Testing Sent: VBG, ABG
Related Diagnosis: Other Resp Failure
Related Management: Vent. Support
Inpatient, Outpatient, Emergency Visits
De-identified, patient-level data
Federated Health Record Data
80 US Healthcare Orgs (mostly AMCs)
Allegedly, Reconciled Inter-institution Linkage
NOT able to identify institution ### Slide 19
Add slide with what Wayne did ### Slide 20
Data Cleaning
Types of Missing Data:
For a given encounter, the institution must have submitted some data (any data) for each of the data elements requested.
Data validation measures:
Positive controls: if paralytic, expect intubation
Negative controls: <BMI 30 in OHS
True (patho)physiologic state
What the clinician recognizes
What the clinician documents
What information is structured
What data is aggregated in TriNetX
Problem for our study
Missing data here is part of what we want to study ### Slide 21
Results:
Total
No ABG Obtained
ABG Obtained
N879,019
N675,620
N203,399
Encounter Characteristics
Ambulatory
39% (343,699)
50% (334,457)
5% (9,242)
Emergency
20% (175,792)
22% (146,066)
15% (29,726)
Inpatient
41% (359,528)
29% (195,097)
81% (164,431)
Critical Care Services Billed
13% (111,320)
6% (41,976)
34% (69,344)
Patient Characteristics
Age (years)
58 (±18)
57 (±18)
61 (±17)
Female
53% (465,854)
55% (373,158)
46% (92,696)
BMI kg/m2
34 (±9)
35 (±9)
29 (±7)
CHF
17% (145,615)
16% (106,492)
19% (39,123)
CKD
15% (128,047)
14% (94,048)
17% (33,999)
COPD
16% (144,127)
16% (106,788)
18% (37,339)
Neuromuscular Disease
4% (35,539)
4% (27,725)
4% (7,814)
OSA
21% (188,757)
24% (160,082)
14% (28,675)
Diuretics (outpatient)
38% (334,400)
38% (256,188)
38% (78,212)
Opiate (outpatient)
59% (518,493)
57% (387,725)
64% (130,768)
Serum Bicarbonate (mEq/L)
24.6 (±4.7)
25.0 (±4.1)
23.6 (±5.7)
Outcomes
Non-invasive Ventilation?
4% (32,542)
2% (15,352)
8% (17,190)
Invasive Mechanical Ventilation
6% (52,533)
1% (10,054)
21% (42,479)
CO2 > 45 mmHg?
32% (64,544) ### Slide 22
Statistical Analyses:
Receiver Operating Characteristic (ROC) Curve Analysis
ROC-Generalized Linear Modeling
Interval Likelihood Ratios (iLR)
How well does HCO3- separate patients with ↑CO2 from those without (discrimination)?
What characteristics independently associate with better or worse discrimination?
How much does a particular HCO3- value change the likelihood of ↑CO2 being present?
(Exposure) HCO3- and (Outcome) ABG obtained on 1st day of encounter. ### Slide 23
Results: ### Slide 24
Results: ### Slide 25
Results: ### Slide 26
Additional hypothesis: Potassium levels will improve accuracy.
Rationale:
False positives are caused by metabolic alkaloses
Most metabolic alkaloses cause hypokalemia (diuretics, vomiting, hypovol.)
Therefore, high bicarbonate with normal/high potassium is more likely a true positive ### Slide 27
Results: ### Slide 28
Conclusions
Serum bicarbonate is potentially diagnostically useful in a variety of previously unvalidated situations
Some factors correlate with better performance (e.g. Loop diuretic, COPD, CHF, female sex, older age) and others with worse (ER, CKD, Opiate use)
Loop diuretic performance contradicts prior hypothesis: while they increase false positives, they decrease false negatives more.
New Finding: Potassium levels and bicarbonate together provide stronger evidence for or against hypercapnia being present. ### Slide 29
Limitations:
Partial verification: not everyone gets an arterial blood gas to verify if they have hypercapnia or not.
Current Finding applies to type of patients who get ABGs
Probability of obtaining an ABG is variable and idiosyncratic
Inverse probability of blood-gas weighting? Missing data is tricky
‘Acute’, transient hypercapnia will be under-identified.
Will miss if inadequate time for kidneys to retain HCO3-
Calendar-day time matching
Is this a bad thing?
Is the data quality sufficient?
Restricted analyses to single-encounter data (cross-sectional) ### Slide 30
Future: More Predictors Better?
Traditional clinical test evaluation → bioinformatics
(priority from easy to understand → works better)
Approach: Cross-validation
Bias
Variance
Miss the signal (underfitting)
Fitting the noise(overfitting)
Too far in either direction worse performance with new data. ### Slide 31
Approach:
Same Dataset
Take structured data elements that are immediately available in HER
Age, Sex, Serum HCO3- , Serum K+, Serum Creatinine, BMI, Hgb, Venous BG
Modeling approach: Logistic regression with restricted cubic splines
Random-forest decision trees: more flexible, less interpretable.
Output: Predicted (log-odds) of Hypercapnia ### Slide 32
Preliminary Results
TBD ### Slide 33
How can this be put to clinical use?
Computable Phenotype: algorithms that use more information to label whether a patient has a disease.
How you decide who ”counts” as having hypercapnic respiratory failure influences the types of patients included..
Inaccurate disease status classification is a problem for existing EHR-based research
Use for: Exposure, Covariable Adjustment, Outcomes ### Slide 34
Probabilistic Modeling Likelihood of Hypercapnia
Validation & Tuning
Aim 1: Add Unstructured Data
Aim 2: Prospective Validation of PPV
Aim 3: Evaluation of Prognostic Significance
K23:5-yr
Refinement and Validation of a method for constructing cohorts of patients with hypercapnia – robust to variable diagnostic performance. ### Slide 35
Aim 1: Incorporate Unstructured Data
Natural Language Processing, Large Language Models
BioBERT: LLM (like ChatGPT) pre-trained on medical text (PubMed/PMC)
Train to recognize hypercapnia predictors using MIMIC4: 65,000 ICU admissions MGH/BWH
Nursing triage assessments, Chest XR
Apply with model to retrospective data from the University of Utah.
Reference standard: patients who had arterial blood gasses obtained.
Adjust model weights
Generate measures of model performance (discrimination and calibration) to inform other aims
OSA
Pulmonary Embolism
STOP BANG
WELLS Score
Subjective measures:
Loud snoring, tired/fatigued/sleepy, observed apneas? Report of HTN
Objective measures:
BMI, Age, Neck Circumference, Gender
Signs/Symptoms of DVT
PE is most likely
HR over 100
Immobilized/Surgery w/n 4 weeks
Prior PE,
Hemoptysis
Malignancy
High-risk increased perioperative complications
0-1 points <1.2% prevalence
2-3 points 16.2% prevalence
4+ points 37.5% prevalence
Prior efforts use data points you can’t get from structured fields
(can only take specific values)
1990: Run a big enough prospective study and ask everyone
2023: Get ChatGPT to do it from what people document ### Slide 36
Aim 2: Prospectively Validate: UUH Inpatients
Deferred consent (CO2 Narcosis)
Verify Positive Predictive Value (precision) in predicted high-risk patients with Transcutaneous CO2
Smaller sample stratified sampling of the whole population.
Prioritizes proof of concept, rather than delineate scope of the problem (initially)
Sample Size Calculation: stata diagsampsi
Rate of presentation to hospital: (TriNetX)
% Predicted High Risk: (10%)
% of eligible enrolled
25% ‘prevalence’ (Nowbar); model AUC; Se/Sp TcCO2 for true reference standard 85%/85% ### Slide 37
Prior work:
…
49: Correct 93%, Wrong 7%: weight: 1.98%
48: Correct 88%, Wrong 12%: weight: 2.39%
47: Correct 80%, Wrong 20%: weight: 2.47%
46: Correct 69%, Wrong 31%: weight: 2.92%
45: Correct 57%, Wrong 43%: weight: 3.11%
SENSITIVITY
SPECIFICITY
44: Correct 57%, Wrong 43% 3.56%
43: Correct 69%, Wrong 31% weight 3.73
42: Correct 80%, Wrong 20% weight 4.16%
41: Correct 88%, Wrong 12% weight 4.25%
40: Correct 93%, Wrong 7% weight 4.58%
….
Expected test characteristics:
sensitivity 89%
specificity 92%
if
agreement ~6 mmHg
same distribution of CO2
Tool
Health Record-Based Modeling
Transcutaneous CO2
Purpose
Refine Estimation of Pre-test Odds of Hypercapnia
Allow non-invasive CO2 assessment (patients unlikely to consent for research ABGs)
Limitation
Approach only externally validated in outpatient OHS assessment
Prior 95% agreement +/- 6 mmHg insufficient for unstratified clinical use ### Slide 38
Aim 3: Estimate the potential impact
Overdiagnosis: people who meet the definition of disease, but do not benefit from it
PaCO2 is an imperfect reference standard for what we care about: future harm
Possible – though not likely – clinicians already identify the important cases
Need a ‘loyalty cohort’: patients whose rehospitalizations would be captured.
Veterans Affairs – or artificially constructed one from another data source.
Hypothesis: Patients predicted to have had hypercapnia that was not recognized have a similar or excess risk of subsequent readmission as those identified (by blood gas and/or ICD code)
Prior efforts use data points you can’t get from structured fields
(can only take specific values) ### Slide 39
OVERALL: Patients with Consequentially Missed Hypercapnia can be Identified Using Health Record Elements
OVERALL: There is a significant burden of unrecognized hypercapnia among hospitalized patients.
3 aims
Aim 1: Model Refinement and Local Optimization
Aim 2: Prospective Validation
Aim 3: Assess Potential Impact
Hypothesis
Fine-tuning and refining previously derived model will improve local performance
Patients with hypercapnia that are not recognized are common can be identified using health record elements with high positive-predictive value.
Inpatients who were predicted, but not confirmed, to have hypercapnic respiratory failure are at high risk of adverse outcomes.
Rationale
The model should be optimized for local performance
Must be verified that model identifies unrecognized hypercapnia
Need to confirm that the identified patients face elevated risk (ABG PaCO2 is an imperfect reference standard).
Approach
Optimizing model performance +/- use of NLP tools) applied to U of U data
Non-invasively sample (with TcCO2) patients predicted to have hypercapnia that teams have not recognized
Among patients with an admission for hypercapnia that is identified, how many had preceding admissions likely to have hypercapnia?
Design
Retrospective, U of U
Prospective, U of U
Retrospective, VA
Payoff: This work will evaluate the potential for improved recognition of hypercapnia to help a particularly high-risk, enlarging, and understudied group of patients.
Career Development: Prospective validations of bioinformatic tools require a unique skillset ### Slide 40
Summary
Considering hypercapnic respiratory failure as a syndrome offers insight into its attributable health burden
Current diagnosis is haphazard and suboptimal
This introduces problems for patient care and reliable research.
Better use of health record data can help address both problems.
Preliminary work has led to new clinical features, but more sophisiticated modeling is needed for impact.
Related Questions
Methods of handling of partial verification
Regression-based vs Random Forest (or other ML)
Clinical and modeling Incorporation of Venous BG
Socio-technical barriers to diagnosis
Longitudinal Cohort Creation
Is UT atypical? (elevation)
Sources of provider/unit/institution variation?
Brian.Locke@hsc.utah.edu ### Slide 41
TODO: No text extracted from this slide.

193.3 Learning objectives

Building a Better Cohort of Patients with Hypercapnic Respiratory Failure
Overview
What is Hypercapnic Respiratory Failure?
Hypercapnia is common…
Age

193.4 Bottom line / summary

Building a Better Cohort of Patients with Hypercapnic Respiratory Failure
Overview
What is Hypercapnic Respiratory Failure?
Hypercapnia is common…
Age

193 Locke Msci

193.1 Summary

193.2 Slide outline

193.2.1 Slide 1

193.3 Learning objectives

193.4 Bottom line / summary

193.5 Approach

193.6 Red flags / when to escalate

193.7 Common pitfalls

193.8 References

193.9 Slides and assets

193.10 Source materials