Using Constellation to Identify Sepsis

Mark Sendak

2017-11-30

Sepsis

Implementing a sepsis definition motivated development of the constellation package. There is ongoing debate in the medical literature about what constitutes sepsis, but our local team of clinical experts agreed to define sepsis as the occurrence of all 3 following criteria:

  1. 2 or more SIRS criteria
    • Temperature > 100.4 F or Temperature < 96.8 F (6 hours)
    • Heart Rate > 90 (6 hours)
    • Respiration Rate > 20 (6 hours)
    • WBC Count > 12, WBC Count < 4 (24 hours)
  2. Blood culture order (24 hours)
  3. End organ damage
    • Creatinine > 2.0 (24 hours)
    • INR > 1.5 (24 hours)
    • Total bilirubin > 2.0 (24 hours)
    • systolic BP < 90 or decrease in SBP by > 40 over 6 hours (6 hours)
    • Platelets < 100 (24 hours)
    • Lactate >= 2 (24 hours)

For each criteria, there is both a threshold to signify dysfunction as well as a relevant time window to search for events. A major challenge of working with medical data is that the sampling rate varies across measurements. For example, vital signs are sampled more frequently than blood labs, so their information is carried forward for a shorter period of time than lab results.

There are 3 steps to identify sepsis and each step corresponds to a constellation function:

  1. Identify instances of systolic blood pressure drops > 40 over 6 hours (value_change() function)
  2. Identify instances of SIRS >= 2 (constellate_criteria() function)
  3. Identify instances in which patients have SIRS >= 2, a blood culture order, and evidence of end organ damage (constellate() function)

View Vignette with Output

The .Rmd version of the vignette will not show code output. If you’d like to see code output, please do the following:

library(constellation)


vignette("identify_sepsis", package = "constellation")

Environment and Data Prep

First, load constellation and fasttime

library(constellation)


library(data.table)


library(fasttime)

Next, prep the timestamps

for (dt in list(vitals, labs, orders)) {


    date_col <- grep("TIME", names(dt), value = TRUE)


    set(dt, j = date_col, value = fastPOSIXct(dt[[date_col]]))


}

value_change() - identify systolic blood pressure drops

Use this function to identify all instances of a drop in systolic blood pressure of greater than 40 within 6 hours.

systolic_bp <- vitals[VARIABLE == "SYSTOLIC_BP"]




systolic_bp_drop <- value_change(systolic_bp, value = 40, direction = "down",


    window_hours = 6, join_key = "PAT_ID", time_var = "RECORDED_TIME",


    value_var = "VALUE", mult = "all")


head(systolic_bp_drop)
##    PAT_ID PRIOR_RECORDED_TIME PRIOR_VALUE CURRENT_RECORDED_TIME
## 1: 108546 2010-02-25 10:45:29    139.9967   2010-02-25 15:42:35
## 2: 108546 2010-03-01 10:57:24    136.8654   2010-03-01 11:07:00
## 3: 108546 2010-03-02 14:59:20    129.0167   2010-03-02 19:46:35
## 4: 108546 2010-03-02 15:49:00    110.1830   2010-03-02 19:46:35
## 5: 108546 2010-03-03 19:18:41    137.8095   2010-03-03 23:23:54
## 6: 108546 2010-03-03 21:13:39    130.3280   2010-03-03 23:23:54
##    CURRENT_VALUE
## 1:      80.07446
## 2:      88.88972
## 3:      69.94551
## 4:      69.94551
## 5:      82.16874
## 6:      82.16874

constellate_criteria() - identify SIRS >= 2

First, break out all the SIRS component tables applying dysfunction thresholds

temp <- vitals[VARIABLE == "TEMPERATURE" & (VALUE > 100.4 | VALUE < 96.8)]


pulse <- vitals[VARIABLE == "PULSE" & VALUE > 90]


resp <- vitals[VARIABLE == "RESPIRATORY_RATE" & VALUE > 20]


wbc <- labs[VARIABLE == "WBC" & (VALUE > 12 | VALUE < 4)]

Next, identify which criteria are met at every measurement time

sirs <- constellate_criteria(temp, pulse, resp, wbc,


        criteria_names = c("TEMPERATURE", "PULSE", "RESPIRATORY_RATE", "WBC"),


        window_hours = c(6, 6, 6, 24), join_key = "PAT_ID",


        time_var = "RECORDED_TIME")


head(sirs)
##    PAT_ID       RECORDED_TIME TEMPERATURE PULSE RESPIRATORY_RATE WBC
## 1: 108546 2010-02-25 00:46:27           0     1                0   0
## 2: 108546 2010-02-25 01:00:38           0     1                0   0
## 3: 108546 2010-02-25 01:37:13           0     1                1   0
## 4: 108546 2010-02-25 02:50:34           0     1                1   0
## 5: 108546 2010-02-25 03:19:12           0     1                1   0
## 6: 108546 2010-02-25 03:23:08           0     1                1   0

Sum values in the 4 columns to calculate SIRS score at every measurement time and subset to instances where SIRS >= 2

sirs[, SIRS_SCORE := TEMPERATURE + PULSE + RESPIRATORY_RATE + WBC]


sirs <- sirs[SIRS_SCORE >= 2]


head(sirs)
##    PAT_ID       RECORDED_TIME TEMPERATURE PULSE RESPIRATORY_RATE WBC
## 1: 108546 2010-02-25 01:37:13           0     1                1   0
## 2: 108546 2010-02-25 02:50:34           0     1                1   0
## 3: 108546 2010-02-25 03:19:12           0     1                1   0
## 4: 108546 2010-02-25 03:23:08           0     1                1   0
## 5: 108546 2010-02-25 03:30:46           0     1                1   0
## 6: 108546 2010-02-25 04:47:14           0     1                1   0
##    SIRS_SCORE
## 1:          2
## 2:          2
## 3:          2
## 4:          2
## 5:          2
## 6:          2

constellate() - identify instances where 3 sepsis criteria are met

Compile end organ damage

First, build a table that combines all measurements for end organ damage:

## Subset values


end_organ <- labs[


    (VARIABLE == "CREATININE" & VALUE > 2.0) |


    (VARIABLE == "INR" & VALUE > 1.5) |


    (VARIABLE == "BILIRUBIN" & VALUE > 2.0) |


    (VARIABLE == "PLATELETS" & VALUE < 100) |


    (VARIABLE == "LACTATE" & VALUE >= 2.0)


    ]




## normalize systolic_bp_drop


systolic_bp_drop <- systolic_bp_drop[,.(PAT_ID, CURRENT_RECORDED_TIME, CURRENT_VALUE)]


systolic_bp_drop[, VARIABLE := "SYSTOLIC_BP_DROP"]


setnames(systolic_bp_drop, c("CURRENT_RECORDED_TIME", "CURRENT_VALUE"), c("RECORDED_TIME", "VALUE"))




## subset low systolic_bp


systolic_bp <- systolic_bp[VALUE < 90]




## Combine SBP drop and < 90


end_organ <- rbind(end_organ, systolic_bp, systolic_bp_drop)

Normalize blood culture orders

Normalize blood culture order data to feed into constellate() function. The timestamp variable in all time series data frames passed to constellate() must be identical.

setnames(orders, "ORDER_TIME", "RECORDED_TIME")

Run constellate()

Calculate every instance in which all 3 criteria are met for every patient.

## Find all sepsis events


sepsis <- constellate(sirs, orders, end_organ, window_hours = c(24, 24, 24),


    join_key = "PAT_ID", time_var = "RECORDED_TIME", event_name = "SEPSIS",


    mult = "all")


head(sepsis)
##    PAT_ID         SEPSIS_TIME
## 1: 108546 2010-02-25 17:49:06
## 2: 108546 2010-02-25 18:11:45
## 3: 108546 2010-02-25 18:21:19
## 4: 108546 2010-02-25 18:46:15
## 5: 108546 2010-02-25 18:55:56
## 6: 108546 2010-02-25 19:13:26

Calculate the first instant in which all 3 criteria are met for every patient.

## Find first sepsis events


sepsis <- constellate(sirs, orders, end_organ, window_hours = c(24, 24, 24),


    join_key = "PAT_ID", time_var = "RECORDED_TIME", event_name = "SEPSIS",


    mult = "first")


head(sepsis)
##    PAT_ID         SEPSIS_TIME
## 1: 108546 2010-02-25 17:49:06
## 2: 112374 2010-03-09 10:48:09
## 3: 244011 2010-12-05 09:05:25
## 4: 262980 2010-10-31 05:01:11
## 5: 296914 2010-06-16 03:28:30
## 6: 315107 2010-11-09 17:46:47

Calculate the last instant in which all 3 criteria are met for every patient.

## Find last sepsis events


sepsis <- constellate(sirs, orders, end_organ, window_hours = c(24, 24, 24),


    join_key = "PAT_ID", time_var = "RECORDED_TIME", event_name = "SEPSIS",


    mult = "last")


head(sepsis)
##    PAT_ID         SEPSIS_TIME
## 1: 108546 2010-06-03 12:26:05
## 2: 112374 2010-03-10 08:43:00
## 3: 244011 2010-12-08 23:47:59
## 4: 262980 2010-10-31 05:01:11
## 5: 296914 2010-06-16 09:57:53
## 6: 315107 2010-11-09 20:20:17

Technical References

If you’d like to learn more about our sepsis model, please see our publications from the International Conerence on Machine Learning and Machine Learning in Health Care.