Overview

Search…
Overview
Background
CEDARS (Clinical Event Detection and Recording System) is a computational paradigm for collection and aggregation of time-to-event data in retrospective clinical studies. Developed out of a practical need for a more efficient way to conduct medical research, it aims to systematize and accelerate the review of electronic health record (EHR) corpora. It accomplishes those goals by deploying natural language processing (NLP) as a tool to assist detection and characterization of clinical events by human abstractors. In its current iteration, CEDARS is available as an open-source R package under GPL-3 license. The latest stable package can be downloaded from CRAN and the most recent development version can be cloned from GitHub. Full documentation is available here.
Requirements
R 3.5.0 or above and package dependencies
RStudio
MongoDB
Unified Medical Language System (UMLS) MRCONSO.RRF file (desirable but not required)
CEDARS can be installed locally or on a server. In the latter case, Shiny Server (open source or commercial version) will be required. A business-grade server installation of MongoDB is vastly preferred, even if CEDARS is run locally. Because by definition CEDARS handles protected health information (PHI), special consideration should be given to ensure HIPAA (Health Insurance Portability and Accountability Act) compliance, including but not limited to using HTTPS, encryption at rest, minimum password requirements and limiting operation to within institutional firewalls where indicated. CEDARS is provided as-is with no guarantee whatsoever and users agree to be held responsible for compliance with their local government/institutional regulations. All CEDARS installations should be reviewed with institutional information security authorities.
The UMLS is a rich compendium of biomedical lexicons. It is maintained by the National Institutes of Health (NIH) and requires establishing an account in order to access the associated files. Those files are not included with the CEDARS R package, but CEDARS is designed to use them natively so individual users can easily include them in their annotation pipeline. NegEx (Chapman et al, Stud Health Technol Inform. 2013; 192: 677–681.) is included with CEDARS.
Basic Concepts
CEDARS Workflow
Sentences with keywords or concepts of interest are presented to the end user one at a time and in chronological order. The user assesses each sentence, determining whether or not a clinical event is being reported. The whole note or report drawn from the EHR is available for review in the GUI. If no event is declared in the sentence, CEDARS presents the next sentence for the same patient (#1). If an event date is entered, CEDARS moves to the next unreviewed sentence before the event date. If there are no sentences left to review before the event, the GUI moves to the next patient (#2) and the process is repeated with the following record (#3 and #4), until all selected sentences have been reviewed.
In order for CEDARS to be sufficiently sensitive and not miss and unacceptable number of clinical events, the keyword/concept search query must be well thought and exhaustive. The performance of CEDARS will vary by medical area, since the extent of medical lexicon will vary substantially between event types.
Operational Schema
CEDARS Operational Schema
CEDARS is modular and all information for any given annotation project is stored in one MongoDB database. User credentials, original clinical notes, NLP annotations and patient-specific information are stored in dedicated collections. Once clinical notes have been uploaded, they are passed through the NLP pipeline. Currently only UDPipe is supported and integrated with CEDARS. If desired, the annotation pipeline can include negation and medical concept tagging by NegEx and UMLS respectively.
Multiple users can load the web GUI and annotate records at the same time. Once accessed, a given patient record is locked for the user.
Sample Code
The R CEDARS package includes a small simulated clinical notes corpus. This corpus is fictitious and does not contain information from real patients. Once access to MongoDB has been achieved, you can install and test drive CEDARS with the following code:
1
devtools::install_github("simon-hans/CEDARS", upgrade="never")
2
library(CEDARS)
3
​
4
# The code below creates an instance of CEDARS project on a public test MongoDB cluster, populated
5
# with fictitious EHR corpora.
6
​
7
# MongoDB credentials
8
db_user_name <- "testUser"
9
db_user_pw <- "testPW"
10
db_host <- "cedars.yvjp6.mongodb.net"
11
db_replica_set <- NA
12
db_port <- NA
13
​
14
# Using standard MongoDB URL format
15
uri_fun <- mongo_uri_standard
16
​
17
# Name for MongoDB database which will contain the CEDARS project
18
# In this case we generate a random name
19
mongo_database <- find_project_name()
20
​
21
# We create the database and all required collections on a test cluster
22
create_project(uri_fun, db_user_name, db_user_pw, db_host, db_replica_set, db_port, mongo_database,
23
  "CEDARS Example Project", "Dr Smith")
24
​
25
# Adding one CEDARS end user
26
add_end_user(uri_fun, db_user_name, db_user_pw, db_host, db_replica_set, db_port, mongo_database, "John",
27
  "strongpassword")
28
​
29
# Negex is included with CEDARS and required for assessment of negation
30
negex_upload(uri_fun, db_user_name, db_user_pw, db_host, db_replica_set, db_port, mongo_database)
31
​
32
# Uploading the small simulated collection of EHR corpora
33
upload_notes(uri_fun, db_user_name, db_user_pw, db_host, db_replica_set, db_port, mongo_database,
34
  simulated_patients)
35
​
36
# This is a simple query which will report all sentences with a word starting in
37
# "bleed" or "hem", or an exact match for "bled"
38
search_query <- "bleed* OR hem* OR bled"
39
use_negation <- TRUE
40
hide_duplicates <- TRUE
41
skip_after_event <- TRUE
42
save_query(uri_fun, db_user_name, db_user_pw, db_host, db_replica_set, db_port, mongo_database, search_query,
43
  use_negation, hide_duplicates, skip_after_event)
44
​
45
# Running the NLP annotations on EHR corpora
46
# We are only using one core, for large datasets parallel processing is faster
47
automatic_NLP_processor(NA, "latin1", "udpipe", uri_fun, db_user_name, db_user_pw,
48
  db_host, db_replica_set, db_port, mongo_database, max_n_grams_length = 0, negex_depth = 6, select_cores = 1)
49
​
50
# Pre-searching based on query
51
# This is optional but will speed-up the interface
52
pre_search(patient_vect = NA, uri_fun, db_user_name, db_user_pw, db_host, db_replica_set, db_port, mongo_database)
53
​
54
# Start the CEDARS GUI locally
55
# Your user name is "John", password is "strongpassword"
56
# Once you have entered those credentials, click on button "ENTER NEW DATE" and CEDARS will seek the first record to annotate
57
# Try out the interface, adjudicating sentences, entering event dates, comments, moving between sentences and searching for records
58
# Once you have entered some data, close the GUI
59
start_local(db_user_name, db_user_pw, db_host, db_replica_set, db_port, mongo_database)
60
​
61
# Obtaining events and info associated with data entry
62
# The annotations entered in the GUI are now available in this dataframe
63
event_output <- download_events(uri_fun, db_user_name, db_user_pw, db_host, db_replica_set, db_port, mongo_database)
64
​
65
# Remove project from MongoDB
66
terminate_project(uri_fun, db_user_name, db_user_pw, db_host, db_replica_set, db_port, mongo_database, fast=TRUE)
Copied!
If your systems use a different MongoDB URI string standard, you will have to substitute your string-generating function.
Future Development
We are currently documenting the performance of CEDARS with a focus on oncology clinical research. At the present time, we wish to solidify the CEDARS user interface and ensure a smooth experience in multi-user settings. In the longer term, plug-in modules featuring enhanced query generation and adaptive learning will be integrated into the R workflow. Support for other NLP engines and extensive parallel processing are also desirable.
Please communicate with package author Simon Mantha, MD, MPH ([email protected]) if you want to discuss new features or using this software for your clinical research application.
Administrator Manual
Last modified 2d ago
Export as PDF
Copy link
Contents