In awconway/eicu_nitro: What the Package Does (One Line, Title Case)

library(tidyverse)
library(tidymodels)

distributions of features

 edaData <- recipe(sbp_post ~ sbp_pre +
    nitro_pre +
    nitro_post +
    nitro_diff_mcg +
    total_nitro +
    n_nitro +
    nitro_time +
    sbp_pre_mean_60 +
    sbp_pre_mean_120 +
    sbp_pre_median_60 +
    sbp_pre_sd_60+
    sbp_pre_sd_120,
    data = targets::tar_read(training)
  ) |>
    #these predictors are right skewed (many zeros)
    step_YeoJohnson(
       nitro_time, total_nitro, n_nitro, sbp_pre_sd_60
    ) |>
    step_normalize(
     all_numeric_predictors()
    )|>
    prep()|>
    juice()

edaData|>
ggplot(aes(x = n_nitro))+
geom_histogram(bins=200)

edaData|>
ggplot(aes(x = nitro_diff_mcg))+
geom_histogram(bins=200)

sbp diff and nitro diff

targets::tar_read(training)|>
ggplot(aes(x=nitro_diff_mcg, y=sbp_diff))+
geom_point()

sbp post and sbp_pre_mean_60

edaData|>
ggplot(aes(x = sbp_pre_mean_60))+
geom_histogram(bins = 200)

edaData|>
ggplot(aes(x=sbp_pre_mean_60, y=sbp_post))+
geom_point()+
geom_smooth(method="lm")

edaData|>
ggplot(aes(x=sbp_pre, y=sbp_post))+
geom_point() +
geom_smooth(method="lm")

sbp_median_60 is pretty much just sbp_pre - sbp_mean_60 is a little more sensitive to other values over the previous hour. maybe shoud look into a measure of how varied the bp was in the previous time period as a feature?

targets::tar_read(training)|>
ggplot(aes(x=sbp_pre, y=sbp_pre_mean_60))+
geom_point() +
geom_smooth(method="lm")

edaData|>
ggplot(aes(x=sbp_pre, y=sbp_pre_median_60))+
geom_point() +
geom_smooth(method="lm")

Doesn't seem to be much to gain from lengthening the rolling average window to 2 hours

edaData|>
ggplot(aes(x=sbp_post, y=sbp_pre_sd_60))+
geom_point() +
geom_smooth(method="lm")

rolling volatility of previous blood pressure measurements (standard deviation)

edaData|>
ggplot(aes(x=sbp_pre_mean_60, y=sbp_pre_sd_60))+
geom_point() +
geom_smooth(method="lm")

targets::tar_read(training)|>
ggplot(aes(x=sbp_diff, y=sbp_pre_sd_120))+
geom_point() +
geom_smooth(method="lm")

edaData|>
ggplot(aes(x=sbp_pre_sd_60, y=sbp_pre_sd_120))+
geom_point() +
geom_smooth(method="lm")

cor.test(edaData$sbp_pre_sd_60, edaData$sbp_pre_sd_120)

edaData|>
ggplot(aes(x = sbp_pre_sd_60))+
geom_histogram(bins = 200)

Time on nitro

• As time on nitro progresses, the difference in sbp from pre to post nitro titration seems to be more centered around zero.
• May be an interaction with dosage??
• Plots look very similar between nitro_time and time except

targets::tar_read(training)|>
ggplot(aes(x=sbp_diff, y=nitro_time, colour = nitro_diff_mcg))+
geom_point() 

targets::tar_read(training)|>
ggplot(aes(x=sbp_diff, y=time))+
geom_point()

Time and time on nitro is perfectly correlated for some and others nitro started from a non-zero time point. I think just using nitro_time is best considering bioplausibity of effect (i.e. less sensitive the more time on the infusion)

targets::tar_read(training)|>
ggplot(aes(x=nitro_time, y = time))+
geom_point()

Total nitro dose

Would make sense that time on nitro and nitro_dose are not perfectly correlated. Some patients on low dose for long period and others on high dose for a short period.

edaData|>
ggplot(aes(x=total_nitro, y = nitro_time))+
geom_point()+
geom_smooth(method = "lm")

cor.test(edaData$total_nitro, edaData$nitro_time)

awconway/eicu_nitro documentation built on Feb. 21, 2022, 4:35 p.m.