Set random seed for reproducibility
set.seed(1234)
library(tidyverse)
library(lubridate)
library(ggpubr)
library(lime) # ML local interpretation
library(caret) # ML model building
library(ranger)
library(vip)
library(pdp)
Read in data
all.df <- read.csv("./data/aml.all.df.csv")
Convert dates
all.df$dot <- ymd(all.df$dot)
all.df$dor <- ymd(all.df$dor)
all.df$bdate <- ymd(all.df$bdate)
all.df$pdate <- ymd(all.df$pdate)
Convert all character strings to factors
all.df <- all.df %>% mutate_if(is.character,as.factor)
Make outcome a binary variable (0/1 relapse)
all.df$rbin <- factor(all.df$rbin, levels = c("yes", "no"))
Filter out any tests that are post-relapse
all.df <- all.df[which(all.df$bdate < all.df$dor | is.na(all.df$dor)), ]
Filter out relapse >720 days
all.df <- all.df[which(all.df$rbin == "no" | all.df$rtime < 720),]
Filter out any missing tests
all.df <- all.df[!is.na(all.df$bmc_cdw) & !is.na(all.df$bmc_cd3) &
!is.na(all.df$bmc_cd15) & !is.na(all.df$bmc_cd34) &
!is.na(all.df$pbc_cdw) & !is.na(all.df$pbc_cd3) &
!is.na(all.df$pbc_cd15) & !is.na(all.df$pbc_cd34),]
all.df <<- all.df
# dat2 <- dat %>%
# select(rbin, txage, hla, tbi, abd, ci, mtx, mmf, agvhd, cgvhd,
# bmc_cdw, bmc_cd3, bmc_cd15, bmc_cd34,
# pbc_cdw, pbc_cd3, pbc_cd15, pbc_cd34, ID)
all.df <- all.df %>%
select(rbin, sex, txage,
rstatprtx, ghgp, tbi,
bmc_cdw, bmc_cd3, bmc_cd15, bmc_cd34,
pbc_cdw, pbc_cd3, pbc_cd15, pbc_cd34, ID)
all.df <- all.df %>%
mutate_if(is.character, as.factor) %>%
mutate_if(is.integer, as.numeric) %>%
# mutate(abd = tolower(abd)) %>%
drop_na() %>%
droplevels()
all.df2 <- all.df %>%
select(-ID)
fit.caret <- train(
rbin ~ .,
data = all.df2,
method = 'rf'
)
fit.caret
## Random Forest
##
## 102 samples
## 13 predictor
## 2 classes: 'yes', 'no'
##
## No pre-processing
## Resampling: Bootstrapped (25 reps)
## Summary of sample sizes: 102, 102, 102, 102, 102, 102, ...
## Resampling results across tuning parameters:
##
## mtry Accuracy Kappa
## 2 0.8059948 0.2008426
## 10 0.8158994 0.3713088
## 19 0.8147371 0.3971010
##
## Accuracy was used to select the optimal model using the largest value.
## The final value used for the model was mtry = 10.
Optional rf model –> probably not needed
fit.rf <- randomForest::randomForest(
rbin ~ .,
data = all.df2)
pfun <- function(object, newdata) {
# Need vector of predicted class probabilities when using log-loss metric
predict(object, newdata = newdata, type = "prob")[, "yes"]
}
vis <- vi(fit.rf, method = "permute", train = all.df2, target = "rbin",
metric = "roc_auc", pred_wrapper = pfun,
reference_class = "no", nsim = 100)
## Warning: Consider setting the `event_level` argument when using "roc_auc" as
## the metric; see `?vip::vi_permute` for details. Defaulting to `event_level =
## "first"`.
vip(vis, geom = "boxplot") # Figure 12
p <- ggplot(vis, aes(reorder(Variable, Importance), Importance)) +
geom_bar(stat="identity", color="black",
position=position_dodge()) +
geom_errorbar(aes(ymin = Importance-StDev,
ymax = Importance+StDev), width = 0.2) +
coord_flip() + theme_bw() + scale_x_discrete(name = "Variable")
print(p)
explainer_caret <- lime(all.df2, fit.caret, n_bins = 5)
summary(explainer_caret)
## Length Class Mode
## model 25 train list
## preprocess 1 -none- function
## bin_continuous 1 -none- logical
## n_bins 1 -none- numeric
## quantile_bins 1 -none- logical
## use_density 1 -none- logical
## feature_type 14 -none- character
## bin_cuts 14 -none- list
## feature_distribution 14 -none- list
all_patients = unique(all.df$ID)
for (i in 1:length(all_patients)) {
patientID <- which(all.df$ID == all_patients[i])
explanation_caret <- explain(
x = all.df2[patientID,],
explainer = explainer_caret,
n_permutations = 5000,
dist_fun = "gower",
kernel_width = .75,
n_features = 10,
feature_select = "highest_weights",
labels = "yes"
)
p1 <- plot_features(explanation_caret) +
ggtitle(paste("Patient", all_patients[i])) +
scale_fill_manual(values = c('firebrick', 'steelblue'), drop = FALSE)
print(p1)
#ggsave(paste0("./lime_plots/patient_",i,".pdf"), plot = p1)
}
Stanford Medicine, dcshyr@stanford.edu↩︎
University of Utah, simon.brewer@geog.utah.edu↩︎