How to extrapolate data from data

There are many occasions when a column of data needs to be created from an already existing column for ease of data manipulation. For example, perhaps you have a body of text as a pathology report and you want to extract all the reports where the diagnosis is ‘dysplasia’.

You could just subset the data using grepl so that you only get the reports that mention this word…but what if the data needs to be cleaned prior to subsetting like excluding reports where the diagnosis is normal but the phrase ‘No evidence of dysplasia’ is present. Or perhaps there are other manipulations needed prior to subsetting.

This is where data accordionisation is useful. This simply means the creation of data from (usually) a column into another column in the same dataframe.

The neatest way to do this is with the mutate function from the {dplyr} package which is devoted to data cleaning. There are also other ways which I will demonstrate at the end.

The input data here will be an endoscopy data set:

Age <- sample(1:100, 130, replace = TRUE)
Dx <- sample(c("NDBE", "LGD", "HGD", "IMC"), 130, replace = TRUE)
TimeOfEndoscopy <- sample(1:60, 130, replace = TRUE)

library(dplyr)

EMRdf <- data.frame(Age, Dx, TimeOfEndoscopy, stringsAsFactors = F)

Perhaps you need to calculate the number of hours spent doing each endoscopy rather than the number of minutes

EMRdftbb <- EMRdf %>% mutate(TimeOfEndoscopy / 60)

# install.packages("knitr")
library(knitr)
library(kableExtra)

# Just show the top 20 results

kable(head(EMRdftbb, 20))

Age	Dx	TimeOfEndoscopy	TimeOfEndoscopy/60
12	LGD	54	0.9000000
40	LGD	52	0.8666667
100	LGD	25	0.4166667
96	LGD	27	0.4500000
5	NDBE	39	0.6500000
54	LGD	38	0.6333333
65	HGD	41	0.6833333
11	IMC	53	0.8833333
2	HGD	40	0.6666667
59	HGD	21	0.3500000
19	NDBE	25	0.4166667
39	LGD	1	0.0166667
49	NDBE	51	0.8500000
21	LGD	22	0.3666667
23	LGD	1	0.0166667
60	NDBE	28	0.4666667
2	LGD	29	0.4833333
22	NDBE	46	0.7666667
44	NDBE	6	0.1000000
17	HGD	22	0.3666667

That is useful but what if you want to classify the amount of time spent doing each endoscopy as follows: <0.4 hours is too little time and >0.4 hours is too long.

Using ifelse() with mutate for conditional accordionisation.

For this we would use ifelse(). However this can be combined with mutate() so that the result gets put in another column as follows

EMRdf2 <- EMRdf %>%
  mutate(TimeInHours = TimeOfEndoscopy / 60) %>%
  mutate(TimeClassification = ifelse(TimeInHours > 0.4, "Too Long", "Too Short"))

# Just show the top 20 results

kable(head(EMRdf2, 20))

Age	Dx	TimeOfEndoscopy	TimeInHours	TimeClassification
12	LGD	54	0.9000000	Too Long
40	LGD	52	0.8666667	Too Long
100	LGD	25	0.4166667	Too Long
96	LGD	27	0.4500000	Too Long
5	NDBE	39	0.6500000	Too Long
54	LGD	38	0.6333333	Too Long
65	HGD	41	0.6833333	Too Long
11	IMC	53	0.8833333	Too Long
2	HGD	40	0.6666667	Too Long
59	HGD	21	0.3500000	Too Short
19	NDBE	25	0.4166667	Too Long
39	LGD	1	0.0166667	Too Short
49	NDBE	51	0.8500000	Too Long
21	LGD	22	0.3666667	Too Short
23	LGD	1	0.0166667	Too Short
60	NDBE	28	0.4666667	Too Long
2	LGD	29	0.4833333	Too Long
22	NDBE	46	0.7666667	Too Long
44	NDBE	6	0.1000000	Too Short
17	HGD	22	0.3666667	Too Short

Note how we can chain the mutate() function together.

Using multiple ifelse()

What if we want to get more complex and put several classifiers in? We just use more ifelse’s:

EMRdf2 <- EMRdf %>%
  mutate(TimeInHours = TimeOfEndoscopy / 60) %>%
  mutate(TimeClassification = ifelse(TimeInHours > 0.8, "Too Long", ifelse(TimeInHours < 0.5, "Too Short", ifelse(TimeInHours >= 0.5 & TimeInHours <= 0.8, "Just Right", "N"))))

# Just show the top 20 results

kable(head(EMRdf2, 20))

Age	Dx	TimeOfEndoscopy	TimeInHours	TimeClassification
12	LGD	54	0.9000000	Too Long
40	LGD	52	0.8666667	Too Long
100	LGD	25	0.4166667	Too Short
96	LGD	27	0.4500000	Too Short
5	NDBE	39	0.6500000	Just Right
54	LGD	38	0.6333333	Just Right
65	HGD	41	0.6833333	Just Right
11	IMC	53	0.8833333	Too Long
2	HGD	40	0.6666667	Just Right
59	HGD	21	0.3500000	Too Short
19	NDBE	25	0.4166667	Too Short
39	LGD	1	0.0166667	Too Short
49	NDBE	51	0.8500000	Too Long
21	LGD	22	0.3666667	Too Short
23	LGD	1	0.0166667	Too Short
60	NDBE	28	0.4666667	Too Short
2	LGD	29	0.4833333	Too Short
22	NDBE	46	0.7666667	Just Right
44	NDBE	6	0.1000000	Too Short
17	HGD	22	0.3666667	Too Short

Using multiple ifelse() with grepl() or string_extract

Of course we need to extract information from text as well as numeric data. We can do this using grepl() or string_extract() from the library(stringr).

Let’s say we want to extract all the samples that had IMC. We don’t want to subset the data, just extract IMC into a column that says IMC and the rest say ’Non-IMC’

Using the dataset above:

library(stringr)

EMRdf$MyIMC_Column <- str_extract(EMRdf$Dx, "IMC")

# to fill the NA's we would do:EMRdf$MyIMC_Column<-ifelse(grepl("IMC",EMRdf$Dx),"IMC","NoIMC")

# Another way to do this (really should be for more complex examples when you want to extract the entire contents of the cell that has the match)

EMRdf$MyIMC_Column <- ifelse(grepl("IMC", EMRdf$Dx), str_extract(EMRdf$Dx, "IMC"), "NoIMC")

So data can be usefully created from data for further analysis.

Hopefully this way of extrapolating data and especially using conditional expressions to categorise data according to some rules is a helpful way to get more out of your data.

Please follow @gastroDS on twitter

This article originally appeared on https://sebastiz.github.io/gastrodatascience/ and has been edited to render in Quarto and had NHS-R styles applied.