15
16
17
18
19
20
21
22
23
24
25
9
10
11
12
13
14
5
6
3
4
7
8
1
2
#Script to calculate and display a temporal phenome map using MIMIC2 data
#
#
#
#Author:
#
Jeremy L. Warner M.D., M.S.
Assistant Professor of Medicine and Biomedical Informatics jeremy.warner@vanderbilt.edu
#Version date: 07/12/2013
#
#
#
#Disclaimer: This script is open-source and you are free to distribute, modify, and
# use without limitations. The original author is not responsible for any errors associated with the script and offers no guarantees as to its functionality.
Attribution is appreciated.
#
#Note: This is a script file and is not functionalized. Individual sections
# are delineated by headers.
#RPostgreSQL library for SQL database interface with MIMIC2 library(RPostgreSQL) drv <- dbDriver("PostgreSQL")
#Password is in clear text; you must have MIMIC2 access. conn <- dbConnect(drv, dbname="MIMIC2", user="mimic2", password="XXXXXX")
#1 Run algorithm to determine length of hospital stay for all adult patients

46
47
48
49
50
51
40
41
42
43
44
45
34
35
36
37
38
39
29
30
31
32
33
26
27
28
#1.1 All adult admissions with admit and discharge dates all.hadm <- dbGetQuery(conn, "SELECT distinct admissions.subject_id as subject_id, hadm_id, admit_dt, disch_dt
FROM mimic2v26.admissions
WHERE admissions.subject_id in
(SELECT subject_id
FROM mimic2v26.icustay_detail
WHERE icustay_age_group = 'adult')") hadm <- paste(all.hadm$hadm_id, ',', sep='', collapse=' ') hadm <- substr(hadm, start = 1, stop = nchar(hadm) - 1)
#1.1.1 All adult admissions with admission and discharge dates and at least one POE order all.hadm.poe <- dbGetQuery(conn, "
SELECT hadm_id, min(enter_dt) as first_poe, max(stop_dt) as last_poe
FROM mimic2v26.admissions JOIN mimic2v26.poe_order using(hadm_id)
WHERE admissions.subject_id in
(SELECT subject_id
FROM mimic2v26.icustay_detail
WHERE icustay_age_group = 'adult')
GROUP BY admissions.subject_id, hadm_id, admit_dt, disch_dt")
#1.1.2 All adult admissions with at least one lab order within t -48 hours of admission all.hadm.lab <- dbGetQuery(conn, paste("
SELECT admissions.hadm_id as hadm_id, min(charttime) as first_lab
FROM mimic2v26.labevents JOIN mimic2v26.admissions using(subject_id)

66
67
68
69
70
71
60
61
62
63
64
65
72
73
74
75
76
77
52
53
54
55
56
57
58
59
WHERE admissions.hadm_id in (", hadm, ") AND charttime - admit_dt > interval '-48 hours'
GROUP BY admissions.hadm_id"))
#1.1.3 All adult discharges from the ICU all.hadm.icu <- dbGetQuery(conn, "SELECT subject_id, hadm_id, icustay_outtime as icu_out
FROM mimic2v26.icustay_detail
WHERE icustay_age_group = 'adult'")
#1.2 The following code determines the first event for each hadm_id all.hadm <- merge(all.hadm, all.hadm.lab, by='hadm_id') all.hadm <- merge(all.hadm, all.hadm.poe, by='hadm_id', all.x = TRUE) all.hadm$first.event <- as.POSIXct(NA) all.hadm$first.event[difftime(all.hadm$first_lab, all.hadm$admit_dt, units = 'h') <= 24] <- all.hadm$first_lab[difftime(all.hadm$first_lab, all.hadm$admit_dt, units = 'h') <= 24] foo <- which(is.na(all.hadm$first.event))
#1.2.1 Some hadm have enter dates more than 24 hours after admit_dt; fall back to admit_dt/disch_dt all.hadm$first.event[foo] <- all.hadm$first_poe[foo] all.hadm$first.event[foo][difftime(all.hadm$first.event[foo], all.hadm$admit_dt[foo], units = 'h') > 24] <-
NA all.hadm$first.event[is.na(all.hadm$first.event)] <- all.hadm$admit_dt[is.na(all.hadm$first.event)]

98
99
100
101
102
103
92
93
94
95
96
97
86
87
88
89
90
91
81
82
83
84
85
78
79
80
#1.2.2 For patients who went from hospital to ICU and then discharged in the same day, use ICU discharge as out
#and first POE order as in. all.temp <- merge(all.hadm.icu, all.hadm, by='subject_id', all.x = TRUE) all.temp$diff <- difftime(all.temp$icu_out, all.temp$disch_dt, units = 'days') all.temp <- all.temp[which(all.temp$diff >= 0),] all.temp <- all.temp[which(all.temp$diff <= 1),] all.temp$diff <- difftime(all.temp$icu_out, all.temp$first.event, units = 'hours') all.temp <- all.temp[order(all.temp$hadm_id.y),]
#1.3 The following code determines the last event for each hadm_id all.hadm$last.event <- as.POSIXct(NA)
#1.3.1 Adds the times to the all.hadm table all.hadm$last.event[which(all.hadm$hadm_id %in% all.temp$hadm_id.y)] = all.temp$icu_out
#1.3.2 Some hadm have stop dates less than 24 hours after disch_dt foo <- which(is.na(all.hadm$last.event))
#1.3.3 Some hadm have enter dates more than 24 hours after admit_dt; fall back to admit_dt/disch_dt all.hadm$last.event[foo] <- all.hadm$last_poe[foo] all.hadm$last.event[foo][difftime(all.hadm$last.event[foo], all.hadm$disch_dt[foo], units = 'h') > 24] <-
NA
#1.3.4 Some have labs less than 24 hours after disch_dt

118
119
120
121
122
123
124
125
126
127
128
112
113
114
115
116
117
104
105
106
107
108
109
110
111 hadm <- all.hadm$hadm_id[is.na(all.hadm$last.event)] hadm <- paste(hadm, ',', sep='', collapse=' ') hadm <- substr(hadm, start = 1, stop = nchar(hadm) - 1) last.lab <- dbGetQuery(conn, paste("
SELECT admissions.hadm_id as hadm_id, max(charttime) as last_lab
FROM mimic2v26.labevents JOIN mimic2v26.admissions using(subject_id)
WHERE admissions.hadm_id in (", hadm, ") AND charttime - disch_dt <= interval '24 hours' AND charttime - disch_dt > interval '0 hours'
GROUP BY admissions.hadm_id")) last.lab <- last.lab[order(last.lab$hadm_id),] all.hadm$last.event[which(all.hadm$hadm_id %in% last.lab$hadm_id)] = last.lab$last_lab
#1.3.5 Some will just default to the disch_dt all.hadm$last.event[is.na(all.hadm$last.event)] <- all.hadm$disch_dt[is.na(all.hadm$last.event)]
#1.4 Calculate the difference, in hours, between discharge and admit times all.hadm$diff <- difftime(all.hadm$last.event,all.hadm$first.event, units = 'hours') all.hadm$diff <- as.numeric(all.hadm$diff)
#1.4.1 If the difference is a negative number, truncate to 0 hours all.hadm$diff[all.hadm$diff < 0] <- 0

143
144
145
146
147
148
149
150
151
152
153
137
138
139
140
141
142
129
130
131
132
133
134
135
136
#2 Specify bin size and determine interval cutoffs and subpopulations
#2.1 Determine the optimal number of bins with goal ~n patients per bin
#Example: n = 100 n <- 100 bin.n <- ceiling(dim(all.hadm)[1]/n)
#2.2 Prepare to create cutoffs t.cuts <- quantile(as.numeric(all.hadm$diff), probs=seq(from=0, to=1, length.out=bin.n))
#2.3 Calculate subpopulations based on intervals sub.hadm <- as.list(rep(0,(length(t.cuts)-1))) for (a in 1:(length(t.cuts)-1)) { hadm <- all.hadm$hadm_id[which(all.hadm$diff > t.cuts[a] & all.hadm$diff <= t.cuts[a+1])] hadm.n <- length(hadm) hadm <- paste(hadm, ',', sep='', collapse=' ') hadm <- substr(hadm, start = 1, stop = nchar(hadm) - 1) sub.hadm[[a]][1] <- hadm.n sub.hadm[[a]][2] <- hadm

168
169
170
171
172
173
174
175
176
177
178
162
163
164
165
166
167
154
155
156
157
158
159
160
161
}
#3 Obtain ICD-9 counts for the subpopulations
#3.1 Initialize table with all possible ICD-9 in MIMIC icd9.table <- dbGetQuery(conn, "SELECT distinct code FROM mimic2v26.icd9") for (a in 1:(length(t.cuts)-1))
{
temp.colname <- paste("t",a,sep="_")
temp.icd9 <- dbGetQuery(conn, paste(
"SELECT code, count(*) AS", temp.colname,
"FROM mimic2v26.icd9",
"WHERE hadm_id in (",
sub.hadm[[a]][2],
") GROUP BY code"))
icd9.table <- merge(icd9.table,temp.icd9,all.x=T)
}
#3.2 Replace all NA created in this process with 0's icd9.table[is.na(icd9.table)] <- 0

193
194
195
196
197
198
199
200
201
202
203
187
188
189
190
191
192
179
180
181
182
183
184
185
186
#4 Determine p-values for the developed table with Fisher's Exact Test p.table <- data.frame(matrix(nrow=dim(icd9.table)[1],ncol=dim(icd9.table)[2]),stringsAsFactors=F) colnames(p.table) <- colnames(icd9.table) p.table[,1] <- icd9.table[,1] icd9.table.r <- rowSums(icd9.table[,2:dim(icd9.table)[2]]) for (j in 2:dim(p.table)[2])
{
# Number in subgroup, rounded up n3 <- as.integer(sub.hadm[[j-1]][[1]]) if(n3 > 0) { temp <- rep(0,dim(p.table)[1]) for (i in 1:dim(p.table)[1]) temp[i] <- fisher.test(matrix(c(icd9.table[i,j], icd9.table.r[i]-icd9.table[i,j],
n3,
dim(all.hadm)[1]-n3),
nrow = 2, ncol = 2))[['p.value']]

218
219
220
221
222
223
224
225
226
227
228
212
213
214
215
216
217
204
205
206
207
208
209
210
211 p.table[,j] <- temp } else p.table[,j] <- 1
}
#4.1 Replace p=0 with p=1.5e-322 (smallest size reported in R) p.table[p.table==0] <- 1.5e-322
#4.2 Adjust P-value for FDR on each bin individually p.table.adj <- p.table for (i in 2:dim(p.table)[2]) p.table.adj[,i] <- p.adjust(p.table[,i], method='BH')
#5 Data preparation for visualization
#5.1 Create numeric correlates to V-codes and E-codes code <- data.frame(codec=as.character(icd9.table$code), coden=as.character(icd9.table$code),stringsAsFactors=F) temp.index <- grep("V", code$codec) v.rep <- code$codec[temp.index] code$coden[temp.index] <- as.numeric(sub('V','',v.rep))+1050 rm(v.rep) temp.index <- grep("E", code$codec)

243
244
245
246
247
248
249
250
251
252
253
237
238
239
240
241
242
229
230
231
232
233
234
235
236 e.rep <- code$codec[temp.index] code$coden[temp.index] <- as.numeric(sub('E','',e.rep))+400 rm(e.rep) code$coden <- as.numeric(code$coden)
#5.2 Create an ICD-9 chapter color mapping table icd9.col <- data.frame(cbind( lower=c(0,140,240,280,290,320,360,390,460,520,580,630,680,710,740,760,780,
800,1025,1175), color=c('magenta','cyan','blue','green3','black','orange',
'magenta','cyan','blue','green3','black','orange',
'magenta','cyan','blue','green3','black','orange',
'purple','gray')
), stringsAsFactors = FALSE) icd9.col$lower <- as.numeric(icd9.col$lower)
#5.3 Create data frames containing localization, point sizing, and color information
#5.3.1 First data frame, all adjusted p not equal to 1 temp.df1 <- data.frame(codec=NA,coden=NA,y1=NA,y2=NA,p=NA) for (i in 2:dim(p.table.adj)[2]) { temp <- which(p.table.adj[,i] != 1)

268
269
270
271
272
273
274
275
276
277
278
262
263
264
265
266
267
254
255
256
257
258
259
260
261 if (length(temp) > 0) for (j in 1:length(temp))
} temp.df1 <- rbind(temp.df1, cbind(code[temp[j],],y1=i-1,y2=i,p=p.table.adj[temp[j],i])) temp.df1 <- na.omit(temp.df1) temp.df1 <- temp.df1[temp.df1$p < 0.99,] temp.df1 <- temp.df1[order(temp.df1$coden),] temp.df1$t1 <- round(t.cuts[temp.df1$y1]/24,digits=1) temp.df1$t2 <- round(t.cuts[temp.df1$y2]/24,digits=1) temp.df1$color <- '' for (j in 1:dim(temp.df1)[1]) temp.df1$color[j] <- icd9.col$color[max(which(icd9.col$lower <= temp.df1$coden[j]))]
#5.3.2 Second data frame, only adjusted p <= 0.05 temp.df2 <- temp.df1[temp.df1$p <= 0.05,]
#
#
#6 Temporal phenome maps based on temp.df1 and temp.df2
#
If you wish to generate a TIFF, run the entire section.
If you wish to view the graphic within the R environment, exclude the tiff() and dev.off()
If you wish to "zoom" into parts of the map, change the x.min, x.max, y.min, y.max parameters

293
294
295
296
297
298
299
300
301
302
303
287
288
289
290
291
292
279
280
281
282
283
284
285
286
#6.1 Set the borders of the map
#6.1.1 Default x-axis values range from 0 to 1400 x.min <- 0 x.max <- 1400
#6.1.2 Default y-axis values range from 1st to last time interval y.min <- 0 y.max <- dim(p.table)[2]
#6.2 Open pipe to TIFF tiff(filename = paste('temporal.phenome.',n,'.',Sys.Date(),'.tif',sep=''), width=10, height=8, res=300, units='in', compression='lzw')
#6.3 Set the environment to show two graphs in one TIFF/window par(mfrow=c(2,1))
#6.4 Create the first temporal phenome map
#6.4.1 Create temporal phenome map borders and labels, first plot(x=-100, xlim=c(x.min, x.max), xaxt='n', ylim=c(y.min, y.max),

318
319
320
321
322
323
324
325
326
327
328
312
313
314
315
316
317
304
305
306
307
308
309
310
311 yaxt='n', xlab="ICD-9-CM Code", ylab="Time, days")
#6.4.2 Add line segments for (i in 1:dim(temp.df2)[1]) segments(x0=temp.df2$coden[i], y0=temp.df2$y1[i], y1=temp.df2$y2[i]+1, lwd=3*log(-log10(temp.df2$p[i])), col=temp.df2$color[i])
#6.4.3 Place tick marks on the x-axis axis(side=1, labels=c('V Codes','E Codes'), at=c(1100,1300)
)
#6.4.4 Place tick marks on the y-axis axis(side=2, labels=round(quantile(x=t.cuts, probs = seq(0, 1, 0.1))/24),
at=quantile(x=1:length(t.cuts), probs = seq(0, 1, 0.1)))
#6.4.5 Next, add title (default title is "A") mtext(text = "A", font=2, side = 3, adj = 0, padj = -1, cex=1.5)

343
344
345
346
347
348
349
350
351
352
353
337
338
339
340
341
342
329
330
331
332
333
334
335
336
#6.4.6 Add light gray dashed vertical lines to divide by ICD-9 chapter for (i in c(140,240,280,290,320,360,390,460,520,580,630,680,710,740,760,780,
800,1025,1175)) abline(v = i, lty = 2, col = "light gray")
#6.4.7 Adds horizontal lines at 25th, median, 75th percentile
# Default text placement will have to be changed if borders are changed foo <- quantile(1:length(t.cuts), probs=seq(from=0.25, to=.75, length.out=3)) abline(h = foo[1], lty = 2, col = "black") text(1200, foo[1], "25th percentile", pos = 3) abline(h = foo[2], lty = 2, col = "black") text(1200, foo[2], "50th percentile", pos = 3) abline(h = foo[3], lty = 2, col = "black") text(1200, foo[3], "75th percentile", pos = 3)
#6.5 Create the second temporal phenome map
#6.5.1 Create temporal phenome map borders and labels, first plot(x=-100, xlim=c(x.min, x.max), xaxt='n',

368
369
370
371
372
373
374
375
376
377
378
362
363
364
365
366
367
354
355
356
357
358
359
360
361 ylim=c(y.min, y.max), yaxt='n', xlab="ICD-9-CM Code", ylab="Time, days")
#6.5.2 Add line segments for (i in 1:dim(temp.df1)[1]) segments(x0=temp.df1$coden[i], y0=temp.df1$y1[i], y1=temp.df1$y2[i]+1, lwd=3*log(-log10(temp.df1$p[i])), col=temp.df1$color[i])
#6.5.3 Place tick marks on the x-axis axis(side=1, labels=c('V Codes','E Codes'), at=c(1100,1300)
)
#6.5.4 Place tick marks on the y-axis axis(side=2, labels=round(quantile(x=t.cuts, probs = seq(0, 1, 0.1))/24),
at=quantile(x=1:length(t.cuts), probs = seq(0, 1, 0.1)))
#6.5.5 Next, add title (default title is "B")

393
394
395
396
397
398
399
400
401
402
403
387
388
389
390
391
392
379
380
381
382
383
384
385
386 mtext(text = "B", font=2, side = 3, adj = 0, padj = -1, cex=1.5)
#6.5.6 Add light gray dashed vertical lines to divide by ICD-9 chapter for (i in c(140,240,280,290,320,360,390,460,520,580,630,680,710,740,760,780,
800,1025,1175)) abline(v = i, lty = 2, col = "light gray")
#6.5.7 Adds horizontal lines at 25th, median, 75th percentile
# Default text placement will have to be changed if borders are changed foo <- quantile(1:length(t.cuts), probs=seq(from=0.25, to=.75, length.out=3)) abline(h = foo[1], lty = 2, col = "black") text(1200, foo[1], "25th percentile", pos = 3) abline(h = foo[2], lty = 2, col = "black") text(1200, foo[2], "50th percentile", pos = 3) abline(h = foo[3], lty = 2, col = "black") text(1200, foo[3], "75th percentile", pos = 3)
#7 Output pertinent data into a RData file with a specified naming convention:
# temporal.phenome.[bin size].[date of creation].RData save(sub.hadm,all.hadm,icd9.table,p.table,
file=paste('temporal.phenome.',n,'.',Sys.Date(),'.RData',sep=''))