176 June 2023 RIP

176.1 Summary

Project Potpourri
Roadmap
Does PA size predict mortality?
PA diameter
Unadjusted
Confounding can cause:
PA:AA
10 years of follow-up too short in young people?
Conclusions
Total
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
Wisconsin sleep cohort

176.2 Slide outline

176.2.1 Slide 1

Project Potpourri
B Locke
RIP
Supported by the National Institutes of Health under Ruth L. Kirschstein National Research Service Award 5T32HL105321 from the NIH and ATS ASPIRE program ### Slide 2
Roadmap
PA size and Mortality
SRMA of Medical & Surgical Weight Loss for OSA
Central Sleep Apnea CPAP prescribing ()
Computable Phenotype for Hypercapnia ### Slide 3
Does PA size predict mortality?
Does PA size add information beyond demographics?
Does PA size add information beyond what the ED might know?
PA diameter
PA:AA
Men
Women
Threshold
29 mm
27 mm
0.9 ### Slide 4
PA diameter
PA:AA
Men
Women
Threshold
29 mm
27 mm
0.9
Old: It is uncommon to see values more extreme than this in healthy patients (classification)
New: Values more extreme than this identify someone at risk of having or developing a problem (prediction)
What makes a finding abnormal?
Ideal: Values more extreme than this identify someone who benefits from intervention (prediction) ### Slide 5
Unadjusted
Adjusted for Age + Sex ### Slide 6
Confounding can cause:
A spurious relationship
A spurious non-relationship
Total
Normal PA:AA (<0.9)
Increased PA:AA (0.9+)
N990
N651
N339
Age (by decade)
<30 years
15% (147)
9% (58)
26% (89)
30-40 years
17% (165)
14% (89)
22% (76)
40-50 years
17% (169)
19% (121)
14% (48)
50-60 years
16% (159)
18% (118)
12% (41)
60-70 years
15% (145)
18% (114)
9% (31)
70+ years
21% (205)
23% (151)
16% (54)
Normal PAd
Enlarged PAd
N714
N276
18% (128)
7% (19)
19% (135)
11% (30)
18% (126)
16% (43)
16% (116)
13% (96)
18% (49)
16% (113)
33% (92) ### Slide 7
PA:AA
PAd ### Slide 8
10 years of follow-up too short in young people? ### Slide 9
Conclusions
When comparing people of the same age/sex, both PAd and PA:AA predict a higher mortality risk
PAd has a threshold; PA:AA is (log)linearly worse at higher ratio
PA:AA is strongly confounded by age → young people seem to have larger PA:AA without apparent excess risk when mildly elevated
Threshold for abnormal PA:AA should be higher in younger adults
(Yet) Unresolved Questions:
Is excess risk mediated by pulmonary vascular mechanisms?
Other outcomes besides death? (ER, Hospitalization) Who gets further workup and how do they do?
Can you use this information to target intervention? (e.g. PH referral)
Can data be combined for better prediction?
OR of TR Jet Vel. 2.8+ w/n 90d if PA enlarged
Unadjusted
Adjusted: Age, Sex ### Slide 10
Total
Normal AA
Enlarged AA
N990
N732
N258
Age (by decade)
<30 years
15% (147)
19% (142)
2% (5)
30-40 years
17% (165)
20% (149)
6% (16)
40-50 years
17% (169)
19% (141)
11% (28)
50-60 years
16% (159)
14% (103)
22% (56)
60-70 years
15% (145)
12% (91)
21% (54)
70+ years
21% (205)
14% (106)
38% (99) ### Slide 11
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 M.D.; Ainhoa Gomez Lumbreras M.D., Ph.D.; Chia Jie Tan PhD.; Teerawat Nonthasawadsri PharmD; Sajesh Veettil MSPharm, PhD; Chanthawat Patikorn PharmD; Nathorn Chaiyakaunapruk PharmD, PhD
Under Review: JAMA NO ### Slide 12
Wisconsin sleep cohort
A 10% weight loss predicted a 26% (95% CI, 18%-34%) decrease in the AHI.
74 RCTs. PAP improves:
ESS
Sleep QOL
Overall HrQOL
There is not evidence PAP improves:
Mortality
MI/CVA events
~70% w/ OSA are obese
5% (gen obese pop) lose >10%BW
Consider pharmacologic (BMI>27) and surgical (BMI>35) weight loss options & still overweight ### Slide 13
Proportion >10% BWL
Mean %BWL
Gastric Bypass
96-99%
25-30%
Tirzepitide
83.5%
20.9%
Semaglutide
69.3%
10.8%
CMS National Coverage of Determination (NCD) – cannot pay for anti-obesity medications. (since 1993)
Heritability of obesity: 40-70%
Heritability of LDL levels: 22-91%
If LDL levels were visible and having low LDL were desirable/high-status, would we have the same restriction on statins?
Prejudice→ ### Slide 14
PubMed, Embase Ovid, and Cochrane Central
Design: RCTs, no blinding req.
Intervention: surgical or medical weight loss interventions
Comparator: placebo, no treatment, or standard of care.
Outcome: Must report percentage weight loss and AHI difference from randomization to one month or longer follow-up
Analysis: Meta-regression ### Slide 15
1st Author
Year
Country
Centers
Intervention
N
Comparison
% Female by arm
Age by arm
Follow-up (mo)
Hypopnea Definition
Sleep Study
Overall risk of bias
Surgery
Bakker32
2017
USA
2
LGB
28
CPAP
21
43 : 43
50.7±9.2 : 46.3±10.5
18
AASM 1999
PSG
Some concerns
Dixon35
2012
Austrila
7
LAGB
30
Conventional Weight Loss Program
43 : 40
47.4±8.8 : 50.0± 8.2
24
Not specified, but required to be the same on reassessment
Feigel-Gullier34
2015
France
1
33
Intensive Nutritional Care
47 : 25
46.9±8.6:50.1±7.4
36
French Guidelines (as of 1999)
High Risk
Furlan33
2021
Brazil
Roux-en-Y Bypass
Usual Care
13
92 : 69
40[36-49]:48[39-53]
AASM 2012
HSAT
Pharmacologic
Blackman36
2016
USA & Canada
40
Liraglutide 3 mg
180
Placebo
179
28 : 28
48.6±9.9: 48.4±9.5
8
AASM 2007
Low Risk
Chen41
2020
China
Exenatide
22
Insulin
55 : 52
51.14±15.25
55.35±11.83
3
Not Reported
Eskandari37
2013
Sweden
Zonisamide
16
0 : 13
55±10:53±14
ROMANCE38,42
N/A
UK
Liraglutide 3 mg ±CPAP
66
Placebo ±CPAP
58 : 52
54.5[45-61]:53.5[49-57]
26
Tang39
2019
Dapagliflozin and Metformin
Glimepiride and Metformin
45 : 35
56.1±7.2:57.9±10
6
Chinese Medical Society
Winslow40
2010
Phentermine-ER/Topiramate
23
59 : 35
53.4±7.0:51.4±5.7 ### Slide 16
10 eligible trials (n843 patients) were included. Four (n200) assessed bariatric surgery and 6 (n643) assessed pharmacologic interventions over a median follow-up of 13 months (interquartile range 6 to 26 months). ### Slide 17
Decrease of 0.45 (95% CI: 0.18-0.73 events/hour) events per hour for each 1% body weight loss
Corresponds to a 1.84% AHI change (95% CI: 0.34-3.33%) ### Slide 18
%Change in AHI
Change in ESS ### Slide 19
Conclusions
Novelty: can predict how much AHI improvement to expect from anti-obesity interventions
Prior observational data: subject to confounding by the cause of weight loss
This data: still subject to study-level confounders (selection into study; distribution of endotypes)
Intervention
Estimated Weight Loss
Source
Expected Improvement: AHI (events/hr)
Expected Improvement: AHI (%)
Vertical sleeve gastrectomy
25%
Heymsfield and Wadden 2017 NEJM Review22
AHI -15.2 (95 CI -9.80 to -20.61 events/hr)
51.8% improvement [24.1 – 79.5%]
Tirzepatide
20.9%
SURPASS Trial23
AHI -13.3 events/hr [-8.9 to -17.7]
44.3% improvement [22.0 – 66.4%]
Semaglutide
10.8%
AGA 2022 guideline24
-8.74 events/hr (95 CI -6.22 to -11.2 events/hr)
25.7% improvement [14.4 – 37.0%]
Could inform practice / policy.
More robustly: could inform trial design.
Surmount-OSA (Eli Lilly): 90% power for 50% AHI reduction. ETA 1 year. ### Slide 20
Predictors of Initial CPAP Prescription and Subsequent Course in Patients with Central Sleep Apneas.
Brian W. Locke MD, Jeffrey Sellman MD, Jonathan McFarland DO, Francisco Uribe MD, Kimberly Workman RPSGT,
Krishna M. Sundar, MD ### Slide 21
Causes of Central Sleep Apnea
High loop gain from high controller gain
High loop gain from high plant gain
Failure of rhythm generation
CPAP often lowers controller gain: addressing (intermittent) hypoxemia
CPAP often lowers plant gain: improving hypercapnia
Individual Events:
Proportion of Events:
Pure
Central
Pure Obstructive ### Slide 22
Modeled clinicians implied prediction (independent predictors of decision to prescribe CPAP)
Modeled independent predictors of outcomes
Research Question: do clinicians accurately predict who will respond favorably to a trial of CPAP?
Operationalization: are predictors of receiving (or not receiving) a trial of CPAP also the predictors of failing a trial of CPAP when it is tried? ### Slide 23
TODO: No text extracted from this slide. ### Slide 24
Rational decision-making when you can’t predict perfectly:
Outcome →
CPAP adequate if trialed
CPAP not adequate if trialed
Prediction ↓
CPAP trialed
True Positive
False Positive
CPAP not trialed
False Negative
True Negative
??
Harms of the two types of errors
“False Positive “ (trialed CPAP, didn’t work)
“False Negative” (didn’t trial CPAP, would have worked)
Delayed effective treatment
Patients less likely to adhere to PAP if it has previously failed them
More expensive
Requires more expertise
May be harmful in some settings (SERVE-HF)
Optimal “threshold” (minimum predicted probability of CPAP failing where it Is still worthwhile to try) depends on how bad False Positive is relative to False negative.
Is it plausible that the optimal threshold varies by characteristics? Yes, for some… ### Slide 25
Other findings:
Inverse-probability of CPAP weighting adjustment leads to no major change ( more on this later)
However, if patients who received CPAP are different in unmeasured characteristics (HCSR, etc.) then groups aren’t validly comparable.
Total
Diagnosed by PSG
Diagnosed by HSAT
p-value
N588
N426
N162
…
Neurologic Causes
19% (114)
21% (89)
15% (25)
0.13
Cardiac Causes
27% (158)
27% (113)
28% (45)
0.76
Medication Etiology
11% (67)
14% (58)
6% (9)
0.006
Primary CSA
1.5% (9)
1.2% (5)
2.5% (4)
0.25
Treatment-Emergent
35% (207)
37% (157)
31% (50)
0.17
OSA-Associated Centrals
14% (82)
10% (43)
24% (39)
<0.001
Results
No CPAP Rx
27% (159)
31% (133)
16% (26)
Rx CPAP, Adequate
38% (226)
34% (145)
50% (81)
Rx CPAP, Inadequate
35% (203)
35% (148)
34% (55) ### Slide 26
Conclusions:
Providers decisions about whether to trial CPAP appear to correctly weight some features (e.g. Opiates, % CSA) but incorrectly weight others (Age, OSA-associated centrals)
Optimal rates of CPAP trial are TBD, but nudges might offer improvements.
Outcome assumes providers/patients continue therapy benefiting them…
When HSAT (unexpectedly) identifies CSA, CPAP is often trialed and often effective)
Baclofen- or Opiate-related centrals respond poorly to treatment, even when they constitute <50% events
Perhaps should be considered CSA, even when not meeting dx criteria ### Slide 27
Computable Phenotypes for Identifying Hypercapnic Respiratory Failure
Brian W Locke MD, Wayne Richards MSc, Jeanette K Brown MD PhD, Ramkiran Gouripeddi MBBS MSc, Krishna M Sundar MD
Problems with retrospective EHR-based research: Missing Data, Time/Event determination (immortal time), Observational nature (residual confounding; multiple comparisons), Disease Identification (enrollment, covariates, outcomes)
Contrast: Prospective cohort with systematic assessment and verification of diagnostic labels… (Asthma, COPD, OHS, OSA)
Computable phenotype: an algorithms to identify specific types of patients
Simple: e.g. inclusion criteria to a study
Complex: use additional data elements (demographics, meds), natural-language processing, etc. ### Slide 28
Why not just include patients w PaCO2 > 45? (or an ICD code for hypercapnic respiratory failure?)
ABGs are variably obtained in clinical practice: less reliable in usual care than enrollment criteria
People who get ABGs are different from people who don’t
Practice varies widely by setting
Symptoms are variable, the diagnosis is missed in the majority of cases
(Nowbar 2004; Nowalk 2023 SLEEP Abstract)
Study Aim
Efficacy Trial
Determine Burden of Disease
Improve Processes of Care
Ideal Characteristic
Select a population with severe disease, homogeneous in characteristic that predicts treatment responsiveness
Capture all patients who suffer from the adverse consequences of hypercapnic respiratory failure
Identify all patients who will benefit from instituting certain management pathways
Example
Severe COPD, no other cause of CO2, severe air trapping, PaCO2 > 52 mmHg
More sensitive definition?
(If only use PaCO2, you will miss patients and skew toward people with classic/severe presentation)
What is the national re-admission rate?
How common are admission with hypercapnia?
Does PM 2.5 influence? ### Slide 29
Needs: initial decompensation → treatment
Presentation With Hypercapnia
Hypercapnia is common, highly morbid, and likely increasing in frequency (obesity, opiates, multimorbidity)
Correctly Recognized?
Causes Correctly Diagnosed?
Is there data to guide treatment?
Is it logistically possible?
Does it happen?
Inpt Hypercapnia is frequently missed in practice
Many patients do not receive definitive testing
We don’t know what to do for many (most?) patients with hypercapnia
Device qualification restrictions.
Requires coordination between Inpt  Outpt. Consider CHF (CMS f/u after discharge)
Marik, JICM 2012: 75% miss rate.
Nowbar, AJM 2004
OSA in 5-15% vs 50-80%
No good estimate? (NMD/Rest, COPD, OHS, OVS)
No good estimate? ### Slide 30
Proposed framework:Sufficient vs Component Causes (Rubin)
Obesity
Untreated
Sleep Apnea
COPD
Loop diuretic
Opiates
Sufficient cause: a set of factors that will cause disease when present
End-stage COPD with hyperinflation
Component cause: a factor that, if not present, no disease would occur
Untreated OSA, metabolic alkalosis, chronic opiate use
Hypothetical patient with hypercapnic respiratory failure ### Slide 31
COPD, Very Severe
PaCO2 52 mmHg
BMI 45 kg/m2
AHI 50 event/hr
PaCO2 50 mmHg
COPD, mod severe
AHI 50 events/hr
AHI 10 events/hr
Met. Alk loop diur
COPD
Trial Data
+Guidelines
Wait 2-4 wks
OHS
Observational Data
+Guideline
Start Immediately
Overlap Sydnrome
Undifferentiated
No tx data
Limited epi data
Swap any combination of:
Muscular weakness
Obesity
Opiate use
Lung disease
Etc.
???
Post-Admission for Hypercapnic Respiratory Failure Management ### Slide 32
Study
Definition Used
Aim
Meservey et al 2020. USA (Vermont)
DOI: 10.1007/s00408-019-00300-w
18 y/o+ admitted (either ICU or floor) with diagnostic code for hypercapnic respiratory failure
What features of patients admitted for hypercapnic respiratory failure predict readmission?
Chung et al, AUS, 2021 DOI: DOI: 10.1164/rccm.202108-1912LE
Identified by initial ABG PaCO2 over 45, excluded iatrogenic causes/sedation.
What is the population prevalence of hypercapnia from any cause? Also, subsequent publication evaluating comorbidity prevalence.
Vonderbank et al 2020. Germany
DOI: 10.2147%2FOAEM.S242075
All patients with dyspnea or pulm disease admitted to hospital received capillary blood gas (some screening with VBG) ; PaCO2 45 mmHg.
Is hypercapnia predictive of mortality at a specialty hospital?
Mini-Systematic Review: (Reviewed in singlicate)
EMBASE search for human cohort studies: “hypercapnic respiratory failure” “hypercarbic respiratory failure” “hypoventilation” “respiratory acidosis” and all EMTREE subtypes.
AS Review – AI-assisted systematic review
9455 → 10 studies evaluating prevalence, comorbidity profile, post-discharge prognosis discharge of patients with hypercapnic respiratory failure
van de Schoot, R. et al. An open source machine learning framework for efficient and transparent systematic reviews.
Nat Mach Intell
DOI:10.1038/s42256-020-00287-7 ### Slide 33
Prevalence of OSA
Study
Definition Used
Aim
Thille et al 2018 France
DOI: 10.1007/s00134-017-4998-3
ICU; pH<7.35 & PaCO2 45+; treated with NIV or IMV; survivors
What is the prevalence of undiagnosed OSA among survivors of Hypercapnic RF?
Ouanes-Besbes et al. 2021, Tunisia PubMed: 33171053
Admitted to ICU 2015-2018, PaCO2 > 45mmHg, pH < 7.35. No prior OSA syndrome diagnosis.
Adler et al. 2017, Switzerland
DOI: 10.1164/rccm.201608-1666OC
Recruited at ICU discharge after surviving: primary admission for Resp failure, PaCO2 > 47.5 mmHg and requiring NIV or IMV
What is the prevalence of comorbidities, including OSA, among survivors of Hypercapnic RF? ### Slide 34
Evaluation of Subgroup by Compensation Status
Study
Definition Used
Aim
Compensated Hypercapnia
Wilson et al., 2021. USA (Michigan)
PubMed: 33951397
18+, hospitalized w/ ABG showing CO2 over 45 and pH 7.35-7.45
Prognostic Significance of Compensated Hypercapnia
Bulbul et al. 2014. Turkey.
doi: 10.4103/1817-1737.128851
Hospitalized adults with PaCO2 over 45, no acidosis, and breathing room air
What comorbidities occur with inpatient, compensated hypoventilation
Marik 2016. USA (Virginia)
DOI: 10.1002/osp4.27
Age 18-90; BMI ≥ 40 kg m2; daytime PaCO2 > 45 mmHg, admission HCO3 > 28 mEq L−1.
Excluded: intrinsic lung disease; thoracic MSK/NMD; 20+ pack year smoking, COPD
What is the prevalence of undiagnosed Obesity Hypoventilation among hospitalized adults?
Respiratory Acidosis
Cavalot et al 2021, Toronto DOI: 10.1080/15412555.2021.1990240
ABG with pH < 7.35 and PaCO2 > 45 mmHg OR VBG pH 7.34 and PvVO2 > 50 mmHg & presence of respiratory symptoms (CEDIS codes).
Exclude CF, NM dz, ILD, thoracic dz, lung ca, CNS dz, Overdose, or trach
1 year re-admission rate for patients presenting with acute respiratory acidosis ### Slide 35
Methods: Data source
De-identified, patient-level data (not PHI, though recommended to be treated as such)
Federated data from medical center EHRs
80 Healthcare Organizations; mostly (but not exclusively) AMCs
80+ gb of data.
1.2 Billion ‘facts’ in medications alone.
Center for High Performance Computing protected environment
Data Available
Data unavailable
Demographics
Diagnosis
Medications
Procedures
Labs
Cancer registry (NAACCR)
Allergies
Death Records (in progress)
Some notes
Diagnostic reports
DICOM image objects
Providers
Departments/clinics ### Slide 36
Methods: Data Processing
Mechanisms of Data Incompleteness:
E.g. Diagnoses
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: e.g if dx of Pna, expect abx Rx
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 37
Overlap between definitions ### Slide 38
COPD, Very Severe
PaCO2 52 mmHg
BMI 45 kg/m2
AHI 50 event/hr
PaCO2 50 mmHg
COPD, mod severe
AHI 50 events/hr
AHI 10 events/hr
Met. Alk loop diur
COPD
Trial Data
+Guidelines
Wait 2-4 wks
OHS
Observational Data
+Guideline
Start Immediately
Overlap Sydnrome
Undifferentiated
No tx data
Limited epi data
Swap any combination of:
Muscular weakness
Obesity
Opiate use
Lung disease
Etc.
???
Post-Admission for Hypercapnic Respiratory Failure Management
ICD Hypercap code and ABG CO2 > 45
ICD Hypercap code and only VBG w CO2 > 48
N19,093
N10,018
Age
63 (±15)
61 (±16)
Location
midwest
21% (3,919)
13% (1,331)
northeast
25% (4,705)
21% (2,143)
south
46% (8,806)
32% (3,233)
west
9% (1,663)
33% (3,311)
BMI
30 (±8)
32 (±9)
CHF
26% (4,913)
18% (1,789)
33% (6,300)
30% (2,974)
Neuromuscular Disease
3% (655)
5% (452)
OSA
25% (4,854)
34% (3,381) ### Slide 39
Intuition: use common data elements to accurately infer missing data elements
May be missing because they are not tracked in the dataset
E.g. National Inpatient Sample: ICD, procedures, and Demographics
May be missing because they were not obtained/recorded by the clinician
Desired Specs:
REALLY simple: Usable by researchers (available as ShinyApp on my github). No ML
Usable on different data-sets:
e.g. NHIS (ICD codes & demographics only), health record data (more data elements)
Probabilistic: Give either a categorical (v. low, low, high, v. high) or numerical prediction.
Some tasks warrant high specificity; others high sensitivity. Allows researcher discretion ### Slide 40
HCO3- ≥ 27 mEq/L
(& K ≥ 4.0 mEq/L)
HCO3- < 27 mEq/L
(or K < 4.0 mEq/L)
PaCO2 ≥ 45 mmHg
True Positive
False Negative
PaCO2 < 45 mmHg
False Positive
True Negative
ABG and BMP obtained on the same day
TriNetX: inception to now (~2018)
Categorize by level of care of encounter (Ambulatory, ED, Inpatient, ICU)
Categorize by whether they had received a diagnosis for CHF, COPD, OSA, Neuro, Opiate Use
n242,298 with paired ABG-BMP
Age
65±19 years
Female
47%
Race/Ethnicity
17% Black
62% White
22% (n53,3k)
HCO3- ≥ 27 mEq/L & K+ ≥ 4.0 mEq/L
15% (n36.3k)
23% (n55.7k)
Accuracy of ↑HCO3- only to predict ↑PaCO2
Spectrum Bias; Thresholds ### Slide 41
Adding potassium ≥ 4.0 mEq/L criterion
Accuracy of ↑HCO3- only to predict ↑PaCO2 ### Slide 42
SRMA: Use of HCO3- in a outpatient sleep study center: different characteristics. ### Slide 43
Can you combine different data-elements?
Aim: identify hypercapnia to a certain threshold probability
Bayes theorem:
LR represents informativeness of a test
(↑)Pretest (↓)LR (no 𝚫) Posttest
Regardless of specific threshold
To assess accuracy: need a “ground truth”: What to use when ABG not obtained?
Many conditions: ”clinician review”, not possible nor feasible
Proposal: Inverse probability of ABG weigting. ### Slide 44
35
55
P(Hypercap | Blue) 1/3
P(Hypercap | Red) 2/3
CO2
Inverse Probability (of Testing) Weighting to create a pseudopopulation approximating if everyone had an ABG
To simplify, imagine everyone is either not-hypercapnic (PaCO2 35) or hypercapnic (PaCO2 55) ### Slide 45
35
2/3
55
1/3
P(ABG | Blue) 2/3
P(ABG | Red) 1/3
P(Hypercap | Blue) 1/3
P(Hypercap | Red) 2/3
CO2
P(ABG)
Inverse Probability (of Testing) Weighting to create a pseudopopulation approximating if everyone had an ABG
To simplify, imagine everyone is either not-hypercapnic (PaCO2 35) or hypercapnic (PaCO2 55)
Imagine we know (or perfectly model) the propensity of different types of people to get ABGs ### Slide 46
35
2/3
55
1/3
P(ABG | Blue) 2/3
P(ABG | Red) 1/3
P(Hypercap | Blue) 1/3
P(Hypercap | Red) 2/3
CO2
P(ABG)
Who actually gets an ABG?
(in real life, there would be some noise) ### Slide 47
35
2/3
55
1/3
P(ABG | Blue) 2/3
P(ABG | Red) 1/3
P(Hypercap | Blue) 1/3
P(Hypercap | Red) 2/3
CO2
P(ABG)
Unweighted Analysis
2 of 18 have documented CO2 > 45
Calculated Relative Risk: 1
Actual Relative Risk: 2 ### Slide 48
35
2/3
55
1/3
P(ABG | Blue) 2/3
P(ABG | Red) 1/3
P(Hypercap | Blue) 1/3
P(Hypercap | Red) 2/3
CO2
P(ABG)
Weighted Analysis
Weight Inverse of propensity
Blue 1.5
Red 3
Actual Relative: 2 ### Slide 49
‘Silver’ Standard: Inverse Probability (of ABG) Weighted PaCO2
Intuition behind the propensity weighted ‘pseudopopulation’:
If a patient has the finding, but was unlikely to get the test: there are probably many more out there that aren’t captured.
It is a valid approximation to the extent that:
Observations are independent and identically distributed 👍
Overlap: some possibility that each patients might have/not get an ABG 👍
Conditional independence: after account for all the covariates, there is no longer an association between the P(ABG) and P(PaCO2 > 45) 👎
To the extent we can capture the reasons for chance variation in the propensity score, it will improve estimation of the features associated with hypercapnia (if an ABG had been checked in everyone) ### Slide 50
LASSO Ordinal logistic regression(s)
Problem: many possible predictors: which ones do you choose?
Take a set of ~10-15 physio/sociologically plausible predictors
LASSO: most parsimonious set of predictors to allow for a categorical probabilistic assessment of hypercapnia (within 24h of admission) if it were checked.
Separate model for diagnostic codes only (expect lower discriminative ability)
Separate model including lab values (EHR data-set)
Very High Probability (90%+) of hypercapnic RF
Age
BMISex
Dx Hypercap. RF?
Dx Any Resp Failure?
COPD?
Etc..
HCO3
Cl-
VBG? PCO2; pH
ABG? PCO2; pH
…
(API & published regression equation) ### Slide 51
Will it work? Validation
Retrospective Validation in an independent data set (Acute, Chronic)
Collaborators EHR, or MIMIC etc.
Retrospective Validation at U of U with elevation correction
Prospective K23 Grant Application: obtain SennTec TcCO2 (95% LOA: +/- 7mmHg) on a random sample of patients admitted and predicted to various levels of PaCO2 probability. 3 Aims Addressed
Does the identification algorithm work?
What happens to the hypercapnic patients? (cohort, care processes)
How well do clinicians do at identification? (and does it matter?)
Grant Writing Fall 2023; Aim for Feb 2024 NIH Grant Cycle ### Slide 52
Questions? Comments

176.3 Learning objectives

Project Potpourri
Roadmap
Does PA size predict mortality?
PA diameter
Unadjusted

176.4 Bottom line / summary

Project Potpourri
Roadmap
Does PA size predict mortality?
PA diameter
Unadjusted

176 June 2023 RIP

176.1 Summary

176.2 Slide outline

176.2.1 Slide 1

176.3 Learning objectives

176.4 Bottom line / summary

176.5 Approach

176.6 Red flags / when to escalate

176.7 Common pitfalls

176.8 References

176.9 Slides and assets

176.10 Source materials