Practical 2: Electronic Medical Records
Finlay Maguire
2022-05-30
knitr::opts_chunk$set(echo = TRUE)
library(readr)
library(dplyr)##
## Attaching package: 'dplyr'## The following objects are masked from 'package:stats':
##
## filter, lag## The following objects are masked from 'package:base':
##
## intersect, setdiff, setequal, unionlibrary(tidyr)
library(ggplot2)
library(scales)##
## Attaching package: 'scales'## The following object is masked from 'package:readr':
##
## col_factorlibrary(tidytext)
library(textstem)## Loading required package: koRpus.lang.en## Loading required package: koRpus## Loading required package: sylly## For information on available language packages for 'koRpus', run
##
## available.koRpus.lang()
##
## and see ?install.koRpus.lang()##
## Attaching package: 'koRpus'## The following object is masked from 'package:readr':
##
## tokenizelibrary(clinspacy)## Welcome to clinspacy.## By default, this package will install and use miniconda and create a "clinspacy" conda environment.## If you want to override this behavior, use clinspacy_init(miniconda = FALSE) and specify an alternative environment using reticulate::use_python() or reticulate::use_conda().library(topicmodels)This practical is based on exploratory data analysis, named entity recognition, and topic modelling of unstructured medical note free-text data derived from electronic medical records (EMR). Real EMR data is very difficult to access without a specific need/request so this data set is derived from medical transcription data instead. I’ll also caveat that the options of natural language processing (NLP) in R are far inferior to those available in Python.
First, install the packages in the setup block
(install.packages(c("readr", "dplyr", "tidyr", "ggplot2", "tidtext", "textstem", "clinspacy", "topicmodels"))).
Note: To try and make it clearer which library certain functions are coming from clearer, I’ll try to do explicit imports throughout this notebook.
Data Parsing
After that we can grab the dataset directly from the
clinspacy library.
raw.data <- clinspacy::dataset_mtsamples()
dplyr::glimpse(raw.data)## Rows: 4,999
## Columns: 6
## $ note_id <int> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 1…
## $ description <chr> "A 23-year-old white female presents with complaint …
## $ medical_specialty <chr> "Allergy / Immunology", "Bariatrics", "Bariatrics", …
## $ sample_name <chr> "Allergic Rhinitis", "Laparoscopic Gastric Bypass Co…
## $ transcription <chr> "SUBJECTIVE:, This 23-year-old white female present…
## $ keywords <chr> "allergy / immunology, allergic rhinitis, allergies,…There is no explanation or data dictionary with this dataset, which is a surprisingly common and frustrating turn of events!
1 Using the output of dplyr’s glimpse
command (or rstudio’s data viewer by clicking on raw.data
in the Environment pane) provide a description of what you think each in
this dataset contains.
Let’s see how many different medical specialties are featured in these notes:
raw.data %>% dplyr::select(medical_specialty) %>% dplyr::n_distinct()## [1] 40So, how many transcripts are there from each specialty:
ggplot2::ggplot(raw.data, ggplot2::aes(y=medical_specialty)) + ggplot2::geom_bar()
Let’s make our life easier and filter down to 3 specialties: a diagonstic/lab, a medical, and a surgical specialty
analysis.data <- raw.data %>% dplyr::filter(medical_specialty %in% c("Neurology", "Radiology", "Neurosurgery")) Text Processing
Let’s now apply our standard pre-processing to the transcripts from
these specialties.
We are going to use the tidytext package to tokenise the
transcript free-text.
By default this tokenises to words but other options include characters,
n-grams, sentences, lines, paragraphs, or separation around a regular
expression.
tokenized.data <- analysis.data %>% tidytext::unnest_tokens(word, transcription, to_lower=TRUE)How many unique tokens are there in the transcripts from each specialty:
tokenized.data %>% dplyr::group_by(medical_specialty) %>% dplyr::distinct(word) %>% dplyr::summarise(n=dplyr::n())## # A tibble: 3 × 2
## medical_specialty n
## <chr> <int>
## 1 Neurology 9015
## 2 Neurosurgery 4263
## 3 Radiology 7313However, there are a lot of extremely common words e.g., “the”, “of”,
“to”, and so forth.
These are known as stop words and we can remove them relative easily
using a list from tidytext::stop_words and
dplyr::anti_join()
2 How many stop words are there in
tidytext::stop_words?
no.stop.tokenized.data <- tokenized.data %>% dplyr::anti_join(tidytext::stop_words) ## Joining, by = "word"3 How many unique words are there in each category without stop words and numbers?
Sometimes we are interested in tokenising/segmenting things other
than words like whole sentences or paragraphs.
4 How many unique sentences are there in each category?
Hint: use ?tidytext::unnest_tokens to see the documentation
for this function.
Now that we’ve tokenized to words and removed stop words, we can find the most commonly word used within each category:
no.stop.tokenized.data %>%
dplyr::group_by(medical_specialty) %>%
dplyr::count(word, sort = TRUE) %>%
dplyr::top_n(5)## Selecting by n## # A tibble: 15 × 3
## # Groups: medical_specialty [3]
## medical_specialty word n
## <chr> <chr> <int>
## 1 Radiology left 701
## 2 Neurology left 672
## 3 Neurology patient 648
## 4 Radiology normal 644
## 5 Neurology 2 533
## 6 Neurology normal 485
## 7 Radiology 2 466
## 8 Neurology history 429
## 9 Radiology 1 409
## 10 Neurosurgery patient 374
## 11 Radiology 3 328
## 12 Neurosurgery c5 289
## 13 Neurosurgery c6 266
## 14 Neurosurgery procedure 247
## 15 Neurosurgery left 222We should lemmatize the tokenized words to prevent over counting of
similar words before further analyses.
Annoyingly, tidytext doesn’t have a built-in
lemmatizer.
5 Do you think a general purpose lemmatizer will work well for medical data? Why not?
Unfortunately, a specialised lemmatizer like in
clinspacy is going to be very painful to install so we will
just use a simple lemmatizer for now:
lemmatized.data <- no.stop.tokenized.data %>% dplyr::mutate(lemma=textstem::lemmatize_words(word))We can now calculate the frequency of lemmas within each specialty and note.
lemma.freq <- lemmatized.data %>%
dplyr::count(medical_specialty, lemma) %>%
dplyr::group_by(medical_specialty) %>%
dplyr::mutate(proportion = n / sum(n)) %>%
tidyr::pivot_wider(names_from = medical_specialty, values_from = proportion) %>%
tidyr::pivot_longer(`Neurosurgery`:`Radiology`,
names_to = "medical_specialty", values_to = "proportion")And plot the relative proportions
ggplot2::ggplot(lemma.freq, ggplot2::aes(x=proportion,
y=`Neurology`,
color=abs(`Neurology` - proportion))) +
ggplot2::geom_abline(color="gray40", lty=2) +
ggplot2::geom_jitter(alpha=0.1, size=2.5, width=0.3, height=0.3) +
ggplot2::geom_text(ggplot2::aes(label=lemma), check_overlap=TRUE, vjust=1.5) +
ggplot2::scale_x_log10(labels=scales::percent_format()) +
ggplot2::scale_y_log10(labels=scales::percent_format()) +
ggplot2::scale_color_gradient(limits=c(0, 0.001), low="darkslategray4", high="gray75") +
ggplot2::facet_wrap(~medical_specialty, ncol = 2) +
ggplot2::theme(legend.position="none") +
ggplot2:: labs(y="Neurology", x = NULL)## Warning: Removed 26214 rows containing missing values (geom_point).## Warning: Removed 26214 rows containing missing values (geom_text).
6 What does this plot tell you about the relative
similarity of lemma frequencies between neurosurgery and neurology and
between radiology and neurosurgery? Based on what these specialties
involve, is this what you would expect?
7 Modify the above plotting code to do a direct comparison of Neurosurgery and Radiology (i.e., have Neurosurgery or Radiology on the Y-axis and the other 2 specialties as the X facets)
TF-IDF Normalisation
Maybe looking at lemmas across all notes in a specialty is misleading, what if we look at lemma frequencies across a specialty.
lemma.counts <- lemmatized.data %>% dplyr::count(medical_specialty, lemma)
total.counts <- lemma.counts %>%
dplyr::group_by(medical_specialty) %>%
dplyr::summarise(total=sum(n))
all.counts <- dplyr::left_join(lemma.counts, total.counts)## Joining, by = "medical_specialty"Now we can calculate the term frequency / invariant document frequency (tf-idf):
all.counts.tfidf <- tidytext::bind_tf_idf(all.counts, lemma, medical_specialty, n) We can then look at the top 10 lemma by tf-idf within each specialty:
all.counts.tfidf %>% dplyr::group_by(medical_specialty) %>% dplyr::slice_max(order_by=tf_idf, n=10)## # A tibble: 31 × 7
## # Groups: medical_specialty [3]
## medical_specialty lemma n total tf idf tf_idf
## <chr> <chr> <int> <int> <dbl> <dbl> <dbl>
## 1 Neurology speech 89 62573 0.00142 0.405 0.000577
## 2 Neurology 93 87 62573 0.00139 0.405 0.000564
## 3 Neurology impression 86 62573 0.00137 0.405 0.000557
## 4 Neurology sleep 82 62573 0.00131 0.405 0.000531
## 5 Neurology b.i.d 30 62573 0.000479 1.10 0.000527
## 6 Neurology cn 78 62573 0.00125 0.405 0.000505
## 7 Neurology drug 77 62573 0.00123 0.405 0.000499
## 8 Neurology hx 70 62573 0.00112 0.405 0.000454
## 9 Neurology 96 69 62573 0.00110 0.405 0.000447
## 10 Neurology fhx 63 62573 0.00101 0.405 0.000408
## # … with 21 more rows8 Are there any lemmas that stand out in these lists? Why?
We can look at transcriptions using these unusual lemmas to check how
they are used with stringr::str_detect
analysis.data %>% dplyr::select(medical_specialty, transcription) %>% dplyr::filter(stringr::str_detect(transcription, 'b.i.d')) %>% dplyr::slice(1)## medical_specialty
## 1 Radiology
## transcription
## 1 CC:, Episodic monocular blindness, OS.,HX:, This 29 y/o RHF was in her usual healthy state until 2 months prior to her 3/11/96 presentation when she developed episodic arthralgias of her knees and ankles, bilaterally. On 3/3/96, she experienced sudden onset monocular blindness, OS, lasting 5-10 minutes in duration. Her vision "greyed out" from the periphery to center of her visual field, OS; and during some episodes progressed to complete blindness (not even light perception). This resolved within a few minutes. She had multiple episodes of vision loss, OS, every day until 3/7/96 when she was placed on heparin for suspected LICA dissection. She saw a local ophthalmologist on 3/4/96 and was told she had a normal funduscopic exam. She experienced 0-1 spell of blindness (OS) per day from 3/7/96 to 3/11/96. In addition, she complained of difficulty with memory since 3/7/96. She denied dysarthria, aphasia or confusion, but had occasional posterior neck and bioccipital-bitemporal headaches.,She had no history of deep venous or arterial thrombosis.,3/4/96, ESR=123. HCT with and without contrast on 3/7/96 and 3/11/96, and Carotid Duplex scan were "unremarkable." Rheumatoid factor=normal. 3-vessel cerebral angiogram (done locally) was reportedly "unremarkable.",She was thought to have temporal arteritis and underwent Temporal Artery biopsy (which was unremarkable), She received Prednisone 80 mg qd for 2 days prior to presentation.,On admission she complained of a left temporal headache at the biopsy site, but no loss of vision or weakness,She had been experiencing mild fevers and chills for several weeks prior to presentation. Furthermore, she had developed cyanosis of the distal #3 toes on feet, and numbness and rash on the lateral aspect of her left foot. She developed a malar rash on her face 1-2 weeks prior to presentation.,MEDS:, Depo-Provera, Prednisone 80mg qd, and Heparin IV.,PMH:, 1)Headaches for 3-4 years, 2)Heart murmur, 3) cryosurgery of cervix, 4)tonsillectomy and adenoidectomy, 5) elective abortion. She had no history of spontaneous miscarriage and had used oral birth control pill for 10 years prior to presentation.,FHX:, Migraine headaches on maternal side, including her mother. No family history of thrombosis.,SHX:, works as a metal grinder and was engaged to be married. She denied any tobacco or illicit drug use. She consumed 1 alcoholic drink per month.,EXAM: ,BP147/74, HR103, RR14, 37.5C.,MS: A&O to person, place and time. Speech was fluent without dysarthria. Repetition, naming and comprehension were intact. 2/3 recall at 2 minutes.,CN: unremarkable.,Motor: unremarkable.,Coord: unremarkable.,Sensory: decreased LT, PP, TEMP, along the lateral aspect of the left foot.,Gait: narrow-based and able to TT, HW and TW without difficulty.,Station: unremarkable.,Reflexes: 2/2 throughout. Plantar responses were flexor, bilaterally.,Skin: Cyanosis of the distal #3 toes on both feet. There was a reticular rash about the lateral aspect of her left foot. There were splinter-type hemorrhages under the fingernails of both hands.,COURSE: , ESR=108 (elevated), Hgb 11.3, Hct 33%, WBC 10.0, Plt 148k, MCV 92 (low) Cr 1.3, BUN 26, CXR and EKG were unremarkable. PTT 42 (elevated). PT normal. The rest of the GS and CBC were normal. Dilute Russell Viper venom time was elevated at 27 and a 1:1 prothrombin time mix corrected to only 36.,She was admitted to the Neurology service. Blood cultures were drawn and were negative. Transthoracic and transesophageal echocardiography on 3/12/96 was unremarkable.,Her symptoms and elevated PTT suggested an ischemic syndrome involving anticardiolipin antibody and/or lupus anticoagulant. Her signs of rash and cyanosis suggested SLE. ANA was positive at 1:640 (speckled), RF (negative), dsDNA, 443 (elevated). Serum cryoglobulins were positive at 1% (fractionation data lost). Serum RPR was positive, but FTA-ABS was negative (thereby confirming a false-positive RPR). Anticardiolipin antibodies IgM and IgG were positive at 56.1 and 56.3 respectively. Myeloperoxidase antibody was negative, ANCA was negative and hepatitis screen unremarkable.,The Dermatology Service felt the patient's reticular foot rash was livedo reticularis. Rheumatology felt the patient met criteria for SLE. Hematology felt the patient met criteria for Anticardiolipin Antibody and/or Lupus anticoagulant Syndrome. Neurology felt the episodic blindness was secondary to thromboembolic events.,Serum Iron studies revealed: FeSat 6, Serum Fe 15, TIBC 237, Reticulocyte count 108.5. The patient was placed on FeSO4 225mg tid.,She was continued on heparin IV, but despite this she continued to have occasional episodes of left monocular blindness or "gray outs" up to 5 times per day. She was seen by the Neuro-ophthalmology Service. The did not think she had evidence of vasculitis in her eye. They recommended treatment with ASA 325mg bid. She was placed on this 3/15/96 and tapered off heparin. She continued to have 0-4 episodes of monocular blindness (OS) for 5-10 seconds per episodes. She was discharged home.,She returned 3/29/96 for episodic diplopia lasting 5-10 minutes per episode. The episodes began on 3/27/96. During the episodes her left eye deviated laterally while the right eye remained in primary gaze. She had no prior history of diplopia or strabismus. Hgb 10.1, Hct 30%, WBC 5.2, MCV 89 (low), Plt 234k. ESR 113mm/hr. PT 12, PTT 45 (high). HCT normal. MRI brain, 3/30/96, revealed a area of increased signal on T2 weighted images in the right frontal lobe white matter. This was felt to represent a thromboembolic event. She was place on heparin IV and treated with Solu-Medrol 125mg IV q12 hours. ASA was discontinued. Hematology, Rheumatology and Neurology agreed to place her on Warfarin. She was placed on Prednisone 60mg qd following the Solu-Medrol. She continued to have transient diplopia and mild vertigo despite INR's of 2.0-2.2. ASA 81mg qd was added to her regimen. In addition, Rheumatology recommended Plaquenil 200mg bid. The neurologic symptoms decreased gradually over the ensuing 3 days. Warfarin was increased to achieve INR 2.5-3.5.,She reported no residual symptoms or new neurologic events on her 5/3/96 Neurology Clinic follow-up visit. She continues to be event free on Warfarin according to her Hematology Clinic notes up to 12/96.9 Extract an example of one of the other unusual “top lemmas” by modifying the above code
Topic Modelling
In NLP, we often have collections of documents (in our case EMR transcriptions) that we’d like to divide into groups so that we can understand them separately. Topic modeling is a method for unsupervised classification of such documents, similar to clustering on numeric data.
Latent Dirichlet allocation (LDA) is a particularly popular method for fitting a topic model. It treats each document as a mixture of topics, and each topic as a mixture of words. This allows documents to “overlap” each other in terms of content, rather than being separated into discrete groups, in a way that mirrors typical use of natural language.
Every document is a mixture of topics. We imagine that each document may contain words from several topics in particular proportions. For example, in a two-topic model we could say “Document 1 is 90% topic A and 10% topic B, while Document 2 is 30% topic A and 70% topic B.”
Every topic is a mixture of words. For example, we could imagine a two-topic model of American news, with one topic for “politics” and one for “entertainment.” The most common words in the politics topic might be “President”, “Congress”, and “government”, while the entertainment topic may be made up of words such as “movies”, “television”, and “actor”. Importantly, words can be shared between topics; a word like “budget” might appear in both equally.
LDA is a mathematical method for estimating both of these at the same time: finding the mixture of words that is associated with each topic, while also determining the mixture of topics that describes each document. There are a number of existing implementations of this algorithm, and we’ll explore one of them in depth.
First lets calculate a term frequency matrix for each transcription:
lemma.counts <- lemmatized.data %>% dplyr::count(note_id, lemma)
total.counts <- lemma.counts %>%
dplyr::group_by(note_id) %>%
dplyr::summarise(total=sum(n))
all.counts <- dplyr::left_join(lemma.counts, total.counts)## Joining, by = "note_id"emr.dcm <- all.counts %>% tidytext::cast_dtm(note_id, lemma, n)Then we can use LDA function to fit a 3 topic (k=3)
LDA-model
emr.lda <- topicmodels::LDA(emr.dcm, k=3, control=list(seed=42))
emr.topics <- tidytext::tidy(emr.lda, matrix='beta')Then we can extract the top terms per assigned topic:
top.terms <- emr.topics %>% dplyr::group_by(topic) %>%
dplyr::slice_max(beta, n=10) %>%
dplyr::ungroup() %>%
dplyr::arrange(topic, -beta)
top.terms %>%
dplyr::mutate(term=tidytext::reorder_within(term, beta, topic)) %>%
ggplot2::ggplot(ggplot2::aes(beta, term, fill=factor(topic))) +
ggplot2::geom_col(show.legend=FALSE) +
ggplot2::facet_wrap(~ topic, scales='free') +
tidytext::scale_y_reordered()
Now we can ask how well do these assigned topics match up to the medical specialties from which each of these transcripts was derived.
specialty_gamma <- tidytext::tidy(emr.lda, matrix='gamma')
# we need to join in the specialty from the note_id
note_id_specialty_mapping <- lemmatized.data %>%
dplyr::mutate(document=as.character(note_id)) %>%
dplyr::select(document, medical_specialty) %>%
dplyr::distinct()
specialty_gamma <- dplyr::left_join(specialty_gamma, note_id_specialty_mapping)## Joining, by = "document"specialty_gamma %>%
dplyr::mutate(medical_specialty = reorder(medical_specialty, gamma * topic)) %>%
ggplot2::ggplot(ggplot2::aes(factor(topic), gamma)) +
ggplot2::geom_boxplot() +
ggplot2::facet_wrap(~ medical_specialty) +
ggplot2::labs(x = "topic", y = expression(gamma))
Interestingly, neurosurgery assigns mostly to a single topic but radiology and neurology are both more diverse in transcriptions. We’d possibly expect this from radiology due to referring to imaging for many different diagnoses/reasons. However, this may all just reflect we are using too few topics in our LDA to capture the range of possible assignments.
10 Repeat this with a 6 topic LDA, do the top terms from the 3 topic LDA still turn up? How do the specialties get split into sub-topics?
Credits
Examples draw heavily on material (and directly quotes/copies text)
from Julia Slige’s tidytext textbook.