Alternative titles

• How to lose friends and infect people
• Epidemiology! For fun and for profit

• 

• 
  
  
  
  
  

About me

• University of Nebraska, Omaha
  • B.S. Computer Science and Mathematics (2009)
  • M.S. Mathematics, Data Science (May, 2018)
• Software Engineer (2004-present)
• Flight Operations, U.S. Army National Guard (2000-2009)
  

  

  

  

Agenda

1. Introduction and motivation
2. C. diff and renal failure
3. The data
4. Methodology
5. Results

Why are we here?

Pr(You)

• $Pr(\text{boy meets girl}) = \frac{1}{20000}$
• $Pr(\text{same boy knocks up same girl}) = \frac{1}{2000}$
• $Pr(\text{right sperm meets right egg})$ = 1 in 400 quadrillion
• $Pr(\text{lineage})$: 1 in $10^{45000}$
• $Pr(\text{you})$ = 1 in $10^{2685000}$

Data Science

Big Data

• Volume
  • How big is BIG? Gigabytes? Petabytes? Exabytes?
• Velocity
  • Twitter averages 5,700 tweets/second
    Kirkorian, R. New Tweets per second record, and how! (2013)
• Variety
  • Plain text, XML, JSON, video, audio, etc.
• And sometimes veracity
  • Questionable data quality

Finding a project

• 🚫 Union Pacific PTC data (2016-2017) - SCRAPPED
• ✅ HCUP project through Creighton

C. diff

• Clostridium difficile
• Gram-positive, anaerobic, rod-shaped, Endospore-forming bacterium
• Has surpassed MRSA as most common nosocomial (hospital-acquired) disease

Scientific Classification

Where it lives

• Intestinal tract of healthy people
  • 13.6% of children, 5.5% of adults are colonized
    Tetro, J. Clostridium difficile Is More Common Than You Think (2016)
  • 70% of infants < 10-12 months old carry C. diff
    Lamont, T. C. diff in 30 Minutes. (2017)
• Soil
• Water
• Feces of infected animals and humans
• Surfaces for up to 5 months

Signs / Symptoms

• Diarrhea
• Fever
• Nausea
• Abdominal pain
• Pseudomembranous colitis
• Toxic megacolon
• Perforation of the colon
• Sepsis

CDI risk

How CDI spreads

CDI treatments

• Antibiotics
  • Flagyl (metronidazole) - Cheap, no longer recommended
  • Vanco (vancomycin) - Expensive but effective
  • Dificid (fidaxomicin) - Most expensive, most effective
    Louie, T., et. al. Fidaxomicin versus Vancomycin for Clostridium difficile Infection. (2011)
• Fecal Microbiota Transplants (FMT)
  • via ng-tube or colonoscopy (similar outcomes)
    Postigo, R. and Kim, J.H. Colonoscopic versus nasogastric fecal transplantation for the treatment of Clostridium difficile infection: a review and pooled analysis. (2012)
  • via OpenBiome capsules
• Probiotics are not recommended

Renal (kidney) disease

• Acute kidney injury (AKI)
• Chronic kidney disease
  • Stages 1-4
• Stage 5 - End-stage renal disease (ESRD)
• Dialysis or transplant

Renal disease signs / symptoms

• Nausea
• Vomiting
• Loss of appetite
• Fatigue and weakness
• Sleep problems
• Changes in urine volume
• …much more

Renal disease treatments

• Treat the underlying cause

AKI Causes

• Decreased blood flow
• Direct damage to kidneys
• Urinary tract blockage

AKI Risks

• Hospitalization
• Advanced age
• Blood vessel blockage in arms/legs
• Diabetes
• High blood pressure
• Heart failure
• Kidney diseases
• Liver diseases
  Mayo Clinic

CKD Causes

• Type I/II diabetes
• High blood pressure
• …much more

CKD risk

• Cardiovascular disease
• Smoking
• Obesity
• Race
  • African-, Native-, or Asian-American
• Family history of kidney disease
• Abnormal kidney structure
• Older age
  Mayo Clinic

Measuring kidney function

• Glomerular Filtration Rate (GFR)
• MDRD

  $$GFR = 175 \times S_{cr} - 1.154 \times \text{Age}^{-0.203} \times 0.742 \cdot I(\text{F}) \times 1.212 \cdot I(\text{AA})$$

• CKD-EPI

  $$GFR = 141 \times min\bigg(\frac{S_{cr}}{\kappa}, 1\bigg)^{\alpha} \times max\bigg(\frac{S_{cr}}{\kappa}, 1\bigg)^{-1.209} \\ \times 0.993^{\text{Age}} \times 1.018 \cdot \text{I}(\text{F}) \times 1.159 \cdot \text{I}(\text{AA})$$

CKD Stages

End-stage renal disease (ESRD)

• When stage 5 is reached
• Dialysis or kidney transplant

Readmissions

• If hospital has “excess readmissions”, penalties are assessed
• 30-day risk standardized measure to calculate Payment Readjustment Factor (PRF)

  All-cause unplanned readmissions to the same or another applicable acute care hospital, occurring within 30 days - for any reason, regardless of principal diagnosis - from the index admission are counted in this measure. Some planned readmissions are not counted. HRRP

  $$\text{PRF} = 1 - min\bigg(0.03, \sum_{dx} \frac{\text{Payment}(dx) \cdot max\big((\text{ERR}(dx) - 1.0), 0\big)}{\text{All payments}}\bigg)$$

• Where $dx$ is one of six measure cohorts, incl heart failure, pneumonia, et. al.
• ERR is a hospital’s performance measure $dx$, and payment refers to base operating DRG payments.

A fun experiment

• Step 1: Pick a random percentage. e.g. 54%, 28%, 77%, etc.
• Step 2: Type that number into Google followed by “of Americans”
• Step 3: Follow rabbit hole for hours

Simple random sample

• Pólya urn model
• With (SRSWR) or Without Replacement (SRSWOR)
  • With replacement - makes use of i.i.d. assumption
  • Without replacement - not i.i.d. but still exchangeable
• Requires access to the entire population

Sampling design

Design effects

Design effects

• “deft”
• Similar to variance inflation factor (VIF)
• Effective sample size

$$D^2(\hat{\theta}) = \frac{SE(\hat{\theta})^2_{complex}}{SE(\hat{\theta})^2_{srs}} = \frac{var(\hat{\theta})_{complex}}{var(\hat{\theta})_{srs}}$$

$$n_{eff} = \frac{n_{complex}}{d^2(\hat{\theta})}$$

Clustering

• Grouping people by geographic regions
• SRS to choose a geographic region
  

Clustering

Stratification

Stratification

Weighting

• $N = 51$
  

Weighting

• $N_{men} = 30$
• $p_{men} = \frac{30}{51} = 0.588$
  

Weighting

• $N_{women} = 21$
• $p_{women} = \frac{21}{51} = 0.412$
  

Weighting

• $N_{women} = 21$
• Women Odds Ratio: $\frac{p_{women}}{p_{men}} = \frac{0.588}{0.412} = 1.427$
• Men Odds Ratio: $\frac{p_{men}}{p_{women}} = \frac{0.412}{0.588} = 0.701$

H-CUP Databases

• Healthcare Cost and Utilization Project
• Includes NIS and NRD
• Must be purchased
• Data usage agreement required for analysts

NIS Sampling Design

• Nationwide Inpatient Sample
  • 1988-2011: 100% sample of 20% of HCUP hospitals
• National Inpatient Sample
  • 2012-present: 20% sample of 100% of HCUP hospitals

HCUP Complex Survey Design

• Clustered on hospital ID
• Weights included in discwt field for national estimates
• 1988-2011: Stratified by census region and bed size
• 2012-present: Stratified by census division and bedside
  • Region 1 (Northeast)
    • Division 1 (New England) -Division 2 (Mid Atlantic)
  • Region 2 (Midwest)
    • Division 3 (East North Central)
    • Division 4 (West North Central) (incl. Nebraska)
  • Region 3 (South)
    • Division 5 (South Atlantic)
    • Division 6 (East South Central)
    • Division 7 (West South Central)
  • Region 4 (West)
    • Division 8 (Mountain)
    • Division 9 (Pacific)

Importance of survey design

• Treating as SRS

summary(lm(los~age, data=cdiff))

## 
## Call:
## lm(formula = los ~ age, data = cdiff)
## 
## Residuals:
##    Min     1Q Median     3Q    Max 
## -14.19  -7.12  -3.98   2.27 349.01 
## 
## Coefficients:
##              Estimate Std. Error t value            Pr(>|t|)    
## (Intercept) 14.190744   0.187955   75.50 <0.0000000000000002 ***
## age         -0.043275   0.002687  -16.11 <0.0000000000000002 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Residual standard error: 13.94 on 73264 degrees of freedom
## Multiple R-squared:  0.003528,   Adjusted R-squared:  0.003514 
## F-statistic: 259.4 on 1 and 73264 DF,  p-value: < 0.00000000000000022

Importance of survey design

• Accounting for survey design with R survey package

library('survey')

cdiff.design <- svydesign(ids = ~hospid, data = cdiff, weights = ~discwt,  strata = ~nis_stratum, nest=TRUE)
summary(svyglm(los~age, design=cdiff.design))

## 
## Call:
## svyglm(formula = los ~ age, design = cdiff.design)
## 
## Survey design:
## svydesign(ids = ~hospid, data = cdiff, weights = ~discwt, strata = ~nis_stratum, 
##     nest = TRUE)
## 
## Coefficients:
##             Estimate Std. Error t value             Pr(>|t|)    
## (Intercept) 13.95231    0.55033  25.353 < 0.0000000000000002 ***
## age         -0.04657    0.00637  -7.311    0.000000000000627 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## (Dispersion parameter for gaussian family taken to be 180.579)
## 
## Number of Fisher Scoring iterations: 2

SRS vs. complex design

SRS line in red, complex design in blue

Research design checklist

NIS Features

• 2009 - Added DX and DXCCS codes 16-25
• 2012 - NIS redesign (removed hospital-specific features)
• 2014 - Added DX and DXCCS codes 26-30

NIS dimensions

• Big data?
• Definitely large data
• ~3 GB per year (raw CSV)
• 108,683,763 total rows
• 10868376.3 per year

NRD Features

• Years 2010-2013 had 116 features
• 2014 had 126 features
  • Added DX and DXCCS 26-30
• ~10 GB per year (raw CSV)
• 87,699,909
• 17539981.8 per year

Features of interest

• Age
• Female
• Length of Stay
• Died
• Hospital categorization
• DXn codes (ICD-9-CM codes)
  • C. diff - 00845
  • AKI - 584, 584.5-584.9
  • CKD - 585, 585.1-585.6, 585.9
  • Renal Failure, unspecified - 586

Dealing with big data

• Column-store relational database
  • Data warehouse

Getting data into the database

• Lots of Excel work

  

• NIS CSV: ~3GB
• read_csv(…NIS…): ~64 GB in RAM
• NRD CSV: ~6GB
• read_csv(…NRD…): CRASHED
  • Had to split data

Getting data back out

SELECT *
  FROM nis
 WHERE nis.dx1  = '00845' 
    OR nis.dx2  = '00845' 
    OR nis.dx3  = '00845' 
    OR nis.dx4  = '00845' 
    OR nis.dx5  = '00845' 
    OR nis.dx6  = '00845' 
    OR nis.dx7  = '00845' 
    OR nis.dx8  = '00845' 
    OR nis.dx9  = '00845' 
    OR nis.dx10 = '00845' 
    OR nis.dx11 = '00845' 
    OR nis.dx12 = '00845' 
    OR nis.dx13 = '00845' 
    OR nis.dx14 = '00845' 
    OR nis.dx15 = '00845'
    OR nis.dx16 = '00845' 
    OR nis.dx17 = '00845' 
    OR nis.dx18 = '00845' 
    OR nis.dx19 = '00845' 
    OR nis.dx20 = '00845' 
    OR nis.dx21 = '00845' 
    OR nis.dx22 = '00845' 
    OR nis.dx23 = '00845' 
    OR nis.dx23 = '00845' 
    OR nis.dx25 = '00845' 
    OR nis.dx26 = '00845' 
    OR nis.dx27 = '00845' 
    OR nis.dx28 = '00845' 
    OR nis.dx29 = '00845'
    OR nis.dx30 = '00845'

Readmissions

• nrd_visitlink allows analyst to link visits
• los and nrd_daystoevent used to determine sequence and length of stay
• Up to analyst to determine “index” admission and readmissions

Other considerations

• 2010-2014: (1 ≤ DMONTH ≤ 12 − ceil(d/30))
  • Cut off index events with enough time to track readmissions
• DIED $\ne$ 0
  • A death on index does not allow for readmission
• Length of stay > 0
  • LOS == zero represents transfers and same-day stays (more complex)
• AGE > 0
  • Infants are often asymptomatic carriers of C. diff

Models

• Linear logistic regression
• Start with all C. diff patients
• Fit each year separately
  • Accounts for changes in years
  • Allowed fitting to complete under hardware limitations

readmitted ~ hosp_hcontrl_govt +
             hosp_hcontrl_priv_np +
             hosp_urcat4 +
             hosp_ur_teach_metro +
             hosp_ur_teach_metro_teaching +
             hosp_bedsize +
             female +
             acute_kidney_failure +
             chronic_kidney_disease2 +
             chronic_kidney_disease3 +
             chronic_kidney_disease4 +
             chronic_kidney_disease5 +
             chronic_kidney_disease6 +
             chronic_kidney_disease_unk +
             renal_failure_unspecified

Trends

Trends

Trends

Trends

Trends

Trends

Trends

Trends

Model fit

Conclusions

• CDI spreading into younger demographics
• Age - correlation or causation?
  • Other comorbidities associated with age
• ESRD, AKI, and some CKD stages strong predictors for CDI readmission

Future work

• Age as a risk factor?
• What is causing spread into younger groups?
• Why are females more likely to get CDI?
• Mortality?
• Treatment studies?

Very special thanks

• Renuga Vivekanandan, M.D.
• Ryan Walters, Ph.D.
• Dora Matache, Ph.D.

Thank you!