Predicting Surgical Site Infections (SSI) using Electronic Medical Records

Rebecca Barter, Prabhu Shankar, Parul Dayal,

Karl Kumbier, Hien Nguyen, Bin Yu

A collaborative journey

Dr Prabhu Shankar

Dr Parul Dayal

Dr Hien Nguyen

Prof Bin Yu

(almost Dr) Rebecca Barter

Dr Karl Kumbier

Zimlichman et al. 2013, Merkow et al. 2015, Klevens et al. 2002, Magill et al. 2014

Surgical Site Infections

~160,000

Cases per year

>8000

Deaths associated with SSI per year

11%

ICU deaths are associated with SSI

$3.2 billion

Attributable cost per year to hospitals

11 days

Additional hospitalization for the average SSI patient

Surgical Site Infections

Closer monitoring

Preventative treatment

Take quicker action

If we can predict which patients are at high risk of SSI...

Better patient outcomes

Lower costs for hospitals

To generate a prediction... we need data!

SSI Surveillance by the CDC

30 days

Infection*?

All hospitals have compulsory SSI reporting to the CDC

National Healthcare Safety Network (NSHN)

*How do you define an infection?

+

NHSN surveillance data

EHR data

Current approaches

Surgery

Patient

SSI

Vitals

Labs

Medications

... Our idea

Obtaining NHSN and EHR data

The data is large and messy

Data was split across

one file per year (2014-2017)
multiple sheets within each excel file

for multiple types of data

Labs
Medications
Previous diagnoses
Problem list
Vitals
NHSN file (patient + surgery info)

Total: 26 excel files, each with multiple sheets

Combining the data was tricky...

The different sheets had slightly different column names, so initially everything beyond the first sheet was missing

Filenames differed slightly by year:

2014: Prob_List.xlsx
2015: Prob_list.xlsx
2016: Problem_list.xlsx
2017: Problem_List.xlsx

In 2017 there were two vitals datasets: *Vitals.xlsx and *Vitals_2

Defining complete datasets

Combined within data-types (across sheets and years)

Goal: combine into a single covariate matrix

NHSN denominator data

39,174 rows and 44 variables

Lab EHR data

12,927,273 rows across 30 lab variables

Vitals EHR data

8,666,375 rows across 9 vitals measurements (including pulse and temperature)

Medication EHR data

7,637,621 rows across 50 medication classes

Converting from long to wide form

Lab data

(long-form)

Lab data

(wide-form)

Want one row per surgery...

Summarising irregularly measured variables

Take average value in the 30 days pre- and post-surgery

Remove labs that have missing values for more than 30% of patients

Combining by patient ID and date

Each row is a single surgery

Handling missing data

Impute

missing values

Remove variables with fewer than 70% non-missing values

Data imputation

Imputation via Random Forest missRanger

Stekhoven, D. J. & Buhlmann, P. (2011) MissForest- non-parametric missing value imputation for mixed-type data.

Fit RF model

Plug into RF model

2.3

2.0

3.2

1.6

Data imputation

Imputation via Random Forest missRanger

Stekhoven, D. J. & Buhlmann, P. (2011) MissForest- non-parametric missing value imputation for mixed-type data.

Fit RF model

Plug into RF model

2.3

2.0

1.6

3.2

1.6

2.3

2.0

3.2

Run through all variables

Iterate until matrix change between iterations is small

Data oddities

Four patients reported with SSI whose surgeries did not appear in the main NHSN dataset
- Decided to remove them
Missing surgery times in 13% of patients
- discussion and exploration revealed that UCD switched to an EPIC EHR system in July 2014
- Decided to remove all patients before July 2014
Diagnosis codes were a mixture of ICD9 and ICD10
- Converted all diagnosis codes to ICD9
Mistakes in definitions in the codebook
- Asked collaborators for clarifications
Lab values that are >10000 but all others are in [0, 20]
- Did some research to discover that these values are possible...