164 Hypercapnia Update Jan 2023
164.1 Summary
- Hypercapnia Research GroupJan 6 Meeting
- Goal of project
- Hypothesis 2: We can construct methods that select a more representative sample of patients (would require expanded data)
- Hypothesis 3: validate how well hypercapnia definitions that can be based on features that can be created from other datasets (e.g. NIS) would perform and whether they are usable.
- Data pre-processing
- Data sanity checking
- Admission VS (skewed center) e.g. for BMI – version w missing data?
- Inpatient medication categories
- Procedures. No TTE, EMG/NCS, PSG. No crit care time?
- Chronic conditions: Heart failure high – rest not?
- for labs e.g. ( value27441 (ABG pH) )- missing values are currently 0. They should be 7.4 (should we mean center?). Also exclude <6.5 and above 7.8
- TODO: still need to sanity check acute diagnoses.
164.2 Slide outline
164.2.1 Slide 1
- Hypercapnia Research GroupJan 6 Meeting ### Slide 2
- Goal of project
- Hypothesis 1: Commonly used methods of identifying hypercapnia identify different patients, and those patients differ in ways expected to influence prognosis and distribution of effect modifiers to treatments.
- Question: how best to demonstrate this difference?
- Principal Components are Hard to Understand
- Stratification by a few key covarites (e.g. age, gender) by frequency of causative diagnoses. (or other features too?)
- idea, this improved stratification might allow better ‘conditional’ estimates knowing a patients comorbidities/clinical circumstance (as compared to more ‘marginal’ estimates that apply to the entire population. ### Slide 3
- Hypothesis 2: We can construct methods that select a more representative sample of patients (would require expanded data)
- In some sense, the fundamental problem of the research endeavor is that the reference standard diagnostic test (ABG) is frequently missing
- Modeling P(obtaining ABG) and P(Hypercapnia | ABG obtained) may be one avenue.
- Partial verification formulation: clinicians only seek a reference standard test in cases where some other features suggest hypercapnia is likely
- If you make certain assumptions… can you do better? IPW can extend things to work if data is assumed to be missing at random. To what extent is it?
- in order to know if that assumption is valid (enough), you’d need a gold standard reference… which would require prospective study on the population of interest (enriched population) ### Slide 4
- Hypothesis 3: validate how well hypercapnia definitions that can be based on features that can be created from other datasets (e.g. NIS) would perform and whether they are usable. ### Slide 5
- Data pre-processing
- Missing data: For hypercapnia icd codes, treatment modalities, and ABG results -> the missingness of the data is a feature of interest
- For other features: should we impute? ### Slide 6
- Data sanity checking
- Total
- ABG Group
- ICD Group
- Procedure Group
- N785,667
- N678,424
- N105,126
- N2,117
- Age
- 60 (16)
- 60 (17)
- 62 (16)
- 52 (19)
- Gender
- 46% (360,445)
- 45% (304,832)
- 52% (54,726)
- 42% (887)
- Asian
- 1% (9,673)
- 1% (8,420)
- 1% (1,245)
- 0% (8)
- Black
- 18% (138,026)
- 18% (119,911)
- 17% (18,026)
- 4% (89)
- Native American / Alaska Native
- 0% (3,545)
- 0% (3,041)
- 0% (504)
- 0% (0)
- Native Hawaiian / Pacific Islander
- 0% (800)
- 0% (644)
- 0% (156)
- White
- 62% (489,151)
- 62% (418,632)
- 66% (68,948)
- 74% (1,571)
- Latino
- 6% (31,744)
- 6% (27,896)
- 7% (3,818)
- 2% (30)
- Location
- ex
- 10% (73,473)
- 11% (72,632)
- 1% (841)
- midwest
- 16% (123,411)
- 17% (114,279)
- 10% (9,132)
- northeast
- 23% (177,929)
- 22% (148,086)
- 34% (29,843)
- south
- 45% (348,616)
- 47% (316,478)
- 34% (30,021)
- 100% (2,117)
- west
- 6% (43,175)
- 4% (23,934)
- 22% (19,241)
- split out group membership into 3 binaries (yes/no ABG, yes/no ICD, yes/no Procedure)
- currently too few procedure. But, that may be a function of sequential evaluation of criteria? ### Slide 7
- Admission VS (skewed center) e.g. for BMI – version w missing data?
- Total
- ABG Group
- ICD Group
- Procedure Group
- N785,667
- N678,424
- N105,126
- N2,117
- Age
- 60 (16)
- 60 (17)
- 62 (16)
- 52 (19)
- Gender
- 46% (360,445)
- 45% (304,832)
- 52% (54,726)
- 42% (887)
- BMI
- 24 (5)
- 24 (6)
- 24 (4)
- Respiratory Rate
- 15 (3)
- 14 (0)
- Temperature (C)
- 37 (1)
- 37 (0)
- Systolic BP
- 123 (19)
- 123 (17)
- 122 (12)
- Diastolic BP
- 73 (12)
- 72 (12)
- 74 (11)
- 75 (7)
- Other features: [ ] spo2? o2? ### Slide 8
- Inpatient medication categories
- Total
- ABG Group
- ICD Group
- Procedure Group
- N785,667
- N678,424
- N105,126
- N2,117
- Age
- 60 (16)
- 60 (17)
- 62 (16)
- 52 (19)
- Gender
- 46% (360,445)
- 45% (304,832)
- 52% (54,726)
- 42% (887)
- PO Corticosteroid
- 41% (320,131)
- 41% (281,124)
- 37% (38,901)
- 5% (106)
- Narcan or other antidote
- 23% (183,224)
- 25% (167,285)
- 15% (15,896)
- 2% (43)
- Inhaled Treatments Rx’d
- 37% (293,266)
- 37% (252,748)
- 38% (40,403)
- 5% (115)
- Vasodilators
- 14% (107,701)
- 14% (97,077)
- 10% (10,554)
- 3% (70)
- Procedures. No TTE, EMG/NCS, PSG. No crit care time?
- Total
- ABG Group
- ICD Group
- Procedure Group
- N785,667
- N678,424
- N105,126
- N2,117
- Age
- 60 (16)
- 60 (17)
- 62 (16)
- 52 (19)
- Gender
- 46% (360,445)
- 45% (304,832)
- 52% (54,726)
- 42% (887)
- CPAP proc. code
- 7% (58,615)
- 6% (43,070)
- 15% (15,524)
- 1% (21)
- Aerosolized Treatment
- 15% (119,865)
- 14% (97,697)
- 21% (22,093)
- 4% (75)
- Inhaler Teaching
- 3% (22,402)
- 3% (19,943)
- 2% (2,401)
- 3% (58)
- 1 View CXR
- 23% (177,806)
- 22% (147,529)
- 29% (30,156)
- 6% (121)
- 2 View CXR
- 6% (46,579)
- 6% (39,661)
- 7% (6,903)
- 1% (15)
- CT Chest Non-Contrast
- 4% (32,650)
- 4% (27,317)
- 5% (5,325)
- 0% (8)
- CT Chest w Contrast
- 4% (30,349)
- 4% (26,256)
- 4% (4,072) ### Slide 10
- Chronic conditions: Heart failure high – rest not?
- Total
- ABG Group
- ICD Group
- Procedure Group
- N785,667
- N678,424
- N105,126
- N2,117
- Age
- 60 (16)
- 60 (17)
- 62 (16)
- 52 (19)
- Gender
- 46% (360,445)
- 45% (304,832)
- 52% (54,726)
- 42% (887)
- Asthma
- 0% (2,934)
- 0% (1,828)
- 1% (1,106)
- 0% (0)
- Cystic Fibrosis
- 0% (1,226)
- 0% (1,041)
- 0% (185)
- CHF
- 21% (166,238)
- 19% (130,533)
- 34% (35,641)
- 3% (64)
- CKD
- 1% (5,485)
- 1% (3,698)
- 2% (1,787)
- COPD
- 0% (1,680)
- 0% (1,232)
- 0% (448)
- Connective Tissue Disease
- 0% (1,112)
- 0% (813)
- 0% (299)
- Dementia
- 0% (2,076)
- 0% (1,394)
- 1% (682)
- Diabetes
- 1% (7,894)
- 1% (5,274)
- 2% (2,620)
- Nicotine dependence
- 1% (4,727)
- 0% (3,250)
- 1% (1,477)
- Neuromuscular Disease
- 0% (924)
- 0% (673)
- 0% (251)
- OSA
- 1% (5,675)
- 1% (3,523)
- 2% (2,152)
- Opiate Use Disorder
- 0% (936)
- 0% (680)
- 0% (256)
- Pulmonary Hypertension
- 0% (2,703)
- 0% (1,675)
- 1% (1,028)
- Peripheral Vascular Disease
- 0% (2,095)
- 0% (1,666)
- 0% (429)
- Stroke
- 0% (2,124)
- 0% (1,734)
- 0% (390) ### Slide 11
- for labs e.g. ( value27441 (ABG pH) )- missing values are currently 0. They should be 7.4 (should we mean center?). Also exclude <6.5 and above 7.8
- Total
- ABG Group
- ICD Group
- Procedure Group
- p-value
- N785,667
- N678,424
- N105,126
- N2,117
- Age
- 60 (16)
- 60 (17)
- 62 (16)
- 52 (19)
- <0.001
- Gender
- 46% (360,445)
- 45% (304,832)
- 52% (54,726)
- 42% (887)
- pH (ABG)
- 7.32 (0.11)
- 7.37 (0.12)
- HCO3 (ABG)
- 24.8 (6.3)
- 21.6 (5.8)
- pH (VBG)
- 7.35 (0.12)
- 7.33 (0.12)
- HCO3 (VBG)
- 25.5 (6.6)
- 25.4 (6.5)
- 26.6 (7.4)
- Serum Sodium
- 138 (5)
- 137 (5)
- Serum Hemoglobin
- 12.1 (2.5)
- 12.2 (2.6)
- 13.0 (2.5)
- // what is value 21600? ### Slide 12
- TODO: still need to sanity check acute diagnoses. ### Slide 13
- PCA output
- Knowing which data features are the major contributors to the principal components might guide which data elements should be included in future models?
- Is the PCA useful on its own?
- PCA: feature generation for other algorithms
- Which algorithms and for what purpose?
- K-means clustering -> what would this get us?
- Regression? ### Slide 14
- Partial verification formulation:
- ABG is reference standard, but how well can things like VBG work as imperfect standards?
- Latent class approach? ### Slide 15
- Expanded Data Set
- Target population: all patients in whom a clinician would consider that the patient has acute hypercapnic respiratory failure
- Analytic sample: widened criteria that includes both patients who had hypercapnic failure, and patients who might reasonably have been expected to. (at admission) ### Slide 16
- need robust truthfulness of data… integrity checks
- Aim: only patient admissions where all the types of data needed WOULD have been reported if they were obtained in clinical care.
- For diagnoses, only diagnosis that were known/made prior to that time. ### Slide 17
- Next steps:
- currently too few procedure. But, that may be a function of sequential evaluation of criteria?
- Problematic features: why no PSG EMG and NCS billed in the entire dataset? - restrictions on location of care?. Yet, TTE also isn’t but clearly that is done inpatient frequently.
- What pre-screening can we do to improve the completeness of the sample?
- require at least 1 result from all categories to be included?
- eliminate iatrogenic cases… somehow? ### Slide 18
- Structure to final paper:
- Currently used methods A, B, C, and D over-selected type and underselected type compared to
- Possible features from the data that could be used to identify, or phenotype, patients include
- For research that aims to enroll a representative sample of patients to estimates the prognosis, burden to the healthcare system, or improve processes of care should consider to identify these patients.
- A future prospective study evaluating the performance of these methods is warranted. ### Slide 19
- Stratifying causes
- Also from that - could I create —> key figure of comorbidities by age in hypercapnia.
- By age? by BMI ; stratified by gender ### Slide 20
- U of U Data
164.3 Learning objectives
- Hypercapnia Research GroupJan 6 Meeting
- Goal of project
- Hypothesis 2: We can construct methods that select a more representative sample of patients (would require expanded data)
- Hypothesis 3: validate how well hypercapnia definitions that can be based on features that can be created from other datasets (e.g. NIS) would perform and whether they are usable.
- Data pre-processing
164.4 Bottom line / summary
- Hypercapnia Research GroupJan 6 Meeting
- Goal of project
- Hypothesis 2: We can construct methods that select a more representative sample of patients (would require expanded data)
- Hypothesis 3: validate how well hypercapnia definitions that can be based on features that can be created from other datasets (e.g. NIS) would perform and whether they are usable.
- Data pre-processing
164.5 Approach
- TODO: Outline the initial assessment or decision point.
- TODO: Outline the next diagnostic or management step.
- TODO: Outline follow-up or escalation criteria.
164.6 Red flags / when to escalate
- TODO: List red flags that require urgent escalation.
164.7 Common pitfalls
- TODO: Capture common errors or missed steps.
164.8 References
TODO: Add landmark references or guideline citations.