Using `AdhereR` with various database technologies for processing very large datasets

In AdhereR: Adherence to Medications

knitr::opts_chunk$set(echo = TRUE)

Introduction

It is sometimes stated that R cannot be used with very large database as they don't fit in memory. Unfortunately, this has also been used as an argument against giving AhereR a try, especially in contexts where it would be most useful, namely for processing (very) large datasets.

Here we try to dispel this worry by presenting various methods for dealing with such cases, focusing on data stored in "classic" relational dabases (such as MySQL, MariaDB, SQLite, PostgreSQL, Microsoft SQL Server or Oracle Database) as well as on a widely-used paradigm for processing large datasets in a distributed manner, namely Hadoop's MapReduce paradigm.

We will present several techniques that can be used to access such data from AdhereR (without attempting to load the whole of it in memory!), to compute the adherence to medication and/or generate plots, and to optionally store the results back into the database.

Please note that while the code here was tested on an Ubuntu 18.04 "server" (an AMD Ryzen 7 2700X CPU with 8 physical cores and 16 logical ones, and 32 GB RAM) and a macOS High Sierra "client" (an early 2015 Macbook Air 11" 7,1 with an Intel Core i7-5650U CPU with 2 physical cores and 4 logical ones, and 8 GB RAM), actually running it (and, thus, compiling this very Rmarkdown script) requires a complex setup (detailed below). Therefore, we provide this vignette in compiled form as a PDF document[^1] together with detailed instructions on how to install the required components for those users that want to reproduce or extend it.

Prerequisites: load AdhereR and Rmarkdown setup bits

Before we start, load AdhereR and various options concerning the figures:

library(AdhereR);

# Various Rmarkdown output options:
# center figures and reduce their file size:
knitr::opts_chunk$set(fig.align = "center", dpi=150, dev="jpeg"); 

Relational databases

Relational databases have a venerable history and are very popular[^2] in a multitude of settings, including those of potential interest for the computation and visualization of patterns of adherence to treatment. In such databases, data is organized in one or more tables, each table comprising several columns (or variables) and rows (or entries) -- a representation familiar to most users of statistical software such as SPSS, SAS, Stata and R (in the latter, this is implemented by data.frame and friends). The querying of such databases is usually done through SQL (or Structured Query Language), which allows, among others, the selection of entries from such tables that meet certain requirements (for an introduction see, for example, @viescas_sql_2008 among many others).

Usually, in practice, there already exists such a relational database that contains the relevant patient info organized in one or more tables, hosted by one of the many commercial or free/open source solutions (or Relational Database Management Systems, RDBMS) available, and which we can access using SQL. For exemplification and reproductibility purposes, here we will also create these databases from the med.events dataset included in the AdhereR package (but these steps are obviously not part of the actual exploitation of the database).

We will focus here on MySQL, but these can be applied to other RDBMS with minimal changes. MySQL is free and widely used, being a good example from the wider class of such systems. MySQL (and its close relative, MariaDB) is a stand-alone server that can be accessed (locally or remotely) by various clients who use SQL to manipulate the stored data. Thus, MySQL stands for the usual scenario where the patient, prescription and event data are stored in a centralized RDBMS (possibly hosted on a dedicated hardware and software infrastructure) which can be queried (possibly over a network or even the Internet) by different specialized clients who perform particular tasks with (parts of) the data.

Currently, there are several ways of accessing a RDBMS from R, and we will focus here on two:

a) using SQL directly, and b) transparently generating SQL queries through the dplyr front-end.

At the time of this writing (November 2018), the MySQL Community Server is a free version of the MySQL RDBMS which can be installed on several Operating Systems including Microsoft Windows, Apple's macOS and various flavors of Linux and BSD (for details, please see https://dev.mysql.com/downloads/mysql/). MariaDB is an open source RDBMS that started as a fork of MySQL (and is still very similar to it) and can as well be installed on a multitude of Operating Systems (for details, please see https://mariadb.org/download/).

Installing MySQL

Generic installation instructions can be found on the MySQL Community Server's website, while step-by-step instructions geared towards R can be found, for example, in the Introduction to MySQL with R and Connecting R to MySQL/MariaDB, among others. How To Install MySQL on Ubuntu 18.04 is oriented specifically for Ubuntu 18.04 LTS (which we use on our test machine). In the following, we will assume that MySQL (Community Server) was successfully installed on the local machine.

In our case, we are on a machine running Ubuntu 18.04 LTS and we follow the procedure described in How To Install MySQL on Ubuntu 18.04, with the difference that we create a user named adherentuser with password AdhereR123! using the command

CREATE USER 'adherentuser'@'localhost' IDENTIFIED BY 'AdhereR123!';

instead of the generic

CREATE USER 'sammy'@'localhost' IDENTIFIED BY 'password';

(and please make sure you do not forget to grant the new user the needed privileges:)

GRANT ALL PRIVILEGES ON *.* TO 'adherentuser'@'localhost' WITH GRANT OPTION;

Creating a sample database

Normally, this step is superfluous as the data should already be present in the RDBMS. Nevertheless, for illustration purposes, we transfer here the med.events example dataset that comes included with AdhereR into several tables in a MySQL database.

Use MySQL Workbench to connect to the database as user adherentuser using the password AdhereR123!. On our system, this means starting it either from the Desktop Environment's start menu or from the command prompt by typing:

mysql-workbench

after which, using the menu Database -> Connect to Database... we made a local connection using the desired user and credentials. Further, as described in Introduction to MySQL with R, create a database:

CREATE DATABASE med_events;

which should become visible in the left-hand panel under "schemas" (see Figure).

Then, either using MySQL Workbench or SQL commands, create four tables within the med_events database:

• event_patients with two columns:

  • id of type INT(11), primary key and non-null, and
  • patient_id of type INT(11) and non-null,

• event_date with two columns:

  • id of type INT(11), primary key and non-null, and
  • date of type DATE and non-null,

• event_info with four columns:

  • id of type INT(11), primary key and non-null,
  • perday of type INT(11) and non-null,
  • category of type VARCHAR(45) and non-null, and
  • duration of type INT(11) and non-null.

• patients with two columns:

  • id of type INT(11), primary key and non-null, and
  • sex of type CHAR(1) and non-null,

The SQL commands for these could be:

CREATE TABLE `med_events`.`event_patients` (
 `id` INT NOT NULL,
 `patient_id` INT NOT NULL,
 PRIMARY KEY (`id`));

CREATE TABLE `med_events`.`event_date` (
 `id` INT NOT NULL,
 `date` DATE NOT NULL,
 PRIMARY KEY (`id`));

CREATE TABLE `med_events`.`event_info` (
 `id` INT NOT NULL,
 `perday` INT NOT NULL,
 `category` VARCHAR(45) NOT NULL,
 `duration` INT NOT NULL,
 PRIMARY KEY (`id`));

CREATE TABLE `med_events`.`patients` (
 `id` INT NOT NULL,
 `sex` CHAR(1) NOT NULL,
 PRIMARY KEY (`id`));

Now, we will fill these tables using the data from AdhereR's med.events dataset from R itself. First, make sure the the package RMariaDB is installed on your system (otherwise simply install it as usual, whether from RStudio or with install.packages("RMariaDB")) and load the library:

library(RMariaDB);

Then, connect to the database:

med_events_db <- dbConnect(RMariaDB::MariaDB(),     # works also for MySQL
                           user="adherentuser",     # the username
                           password="AdhereR123!",  # and password (insecure but ok here)
                           dbname='med_events',     # which database
                           host='localhost'         # on which host (here, local machine)
                          );

and, just to check that things are OK, list the tables:

db_res_tables <- dbListTables(med_events_db); db_res_tables;

Clear the tables (in case there's some junk info in there):

# Truncate all the tables:
for( i in db_res_tables ) dbExecute(med_events_db, paste0("TRUNCATE ",i,";"));

Now, generate an extra column containing the event_id unique event id (really, just the row number) and make sure the format of the dates fits SQL's DATE type:

d <- med.events; # make working a copy
d$event_id <- 1:nrow(d); # simply the index of the event in med.events
d$DATE <- as.character(as.Date(d$DATE, format="%m/%d/%Y"), 
                       format="%Y-%m-%d"); # use the expected format for SQL's DATE

Fill in the patients table (and allocate some random sexes which we won't really use at all):

tmp <- unique(d[,c("PATIENT_ID"),drop=FALSE]); # select the relevant info
tmp$sex <- sample(c("F","M"), size=nrow(tmp), replace=TRUE);
names(tmp) <- c("id", "sex"); # match the column names
# For convenience, write the whole data.frame in one go:
dbWriteTable(med_events_db, name="patients", 
             value=tmp, row.names=FALSE, append=TRUE);

then spread the event info across the remaining tables:

# event_date:
tmp <- d[,c("event_id", "DATE")]; # select the relevant info
names(tmp) <- c("id", "date"); # match the column names
# For convenience, write the whole data.frame in one go:
dbWriteTable(med_events_db, name="event_date", 
             value=tmp, row.names=FALSE, append=TRUE);

# event_info:
tmp <- d[,c("event_id", "PERDAY", "CATEGORY", "DURATION")]; # select the relevant info
names(tmp) <- c("id", "perday", "category", "duration"); # match the column names
# For convenience, write the whole data.frame in one go:
dbWriteTable(med_events_db, name="event_info", 
             value=tmp, row.names=FALSE, append=TRUE);

# event_patients:
tmp <- d[,c("event_id", "PATIENT_ID")]; # select the relevant info
names(tmp) <- c("id", "patient_id"); # match the column names
# For convenience, write the whole data.frame in one go:
dbWriteTable(med_events_db, name="event_patients", 
             value=tmp, row.names=FALSE, append=TRUE);

Check that the data was correctly written either using MySQL Workbench or from R:

# Fetch the whole results as they are small enough for display:
knitr::kable(dbGetQuery(med_events_db, "SELECT * FROM patients LIMIT 5;"), 
             align=c("l","l"), 
             caption="First 5 rows of the `patients` table (please note 
             that the `sex` may differ from your results and between runs).");
knitr::kable(dbGetQuery(med_events_db, "SELECT * FROM event_date LIMIT 5;"),
             align=c("l","l"), 
             caption="First 5 rows of the `event_date` table.");
knitr::kable(dbGetQuery(med_events_db, "SELECT * FROM event_info LIMIT 5;"),
             align=c("l","l"), 
             caption="First 5 rows of the `event_info` table.");
knitr::kable(dbGetQuery(med_events_db, "SELECT * FROM event_patients LIMIT 5;"),
             align=c("l","l"), 
             caption="First 5 rows of the `event_patients` table.");

Finally, don't forget to close the connection to the database:

dbDisconnect(med_events_db); 

Access the database and estimate adherence using explicit SQL

The first method uses explicit SQL statements directed at the MySQL server through the R library RMariaDB (despite its name, it also works for MySQL).

First, make sure it is installed on your system (for example by running the following or, if you're using RStudio, by checking the Packages panel):

require(RMariaDB); # there should be no warnings

If not, install it either using RStudio's Tools -> Install packages menu, or by running:

install.packages("RMariaDB", dep=TRUE); # also install the dependencies

(please make sure you check for any errors and follow any indication; for example, on Ubuntu 18.04 it is necessary to install prior to this the libmysqlclient-dev package using, for example, apt install libmysqlclient-dev).

Now, load it:

library(RMariaDB);

We will next illustrate a few scenarios concerning the computation and plotting of patterns of adherence accessing data from the database.

Connect to the database

med_events_db <- dbConnect(RMariaDB::MariaDB(),     # works also for MySQL
                           user="adherentuser",     # the username
                           password="AdhereR123!",  # and password (insecure but ok here)
                           dbname='med_events',     # which database
                           host='localhost'         # on which host (here, local machine)
                          );
db_res_tables <- dbListTables(med_events_db); db_res_tables; # check things are ok

How many patients are there?

no_patients <- dbGetQuery(med_events_db, 
                          # overkill, as `id` is the primary key:
                          "SELECT COUNT(DISTINCT id) FROM patients;"); 
no_patients <- as.numeric(no_patients[1,1]); # single value: convert it to numeric
no_patients;

How many events?

no_events <- dbGetQuery(med_events_db, 
                        "SELECT COUNT(DISTINCT id) FROM event_info;");
no_events <- as.numeric(no_events[1,1]); # single value: convert it to numeric
no_events;

And how many events per patient?

dbGetQuery(med_events_db, 
           "SELECT patient_id, COUNT(*) FROM event_patients GROUP BY patient_id;");

Get the list of patient_id's

patient_ids <- dbGetQuery(med_events_db, "SELECT id FROM patients;"); # a data.frame
patient_ids <- patient_ids$id; # convert it to a vector
patient_ids;

Retreive the data for a given (set of) patient(s)

We don't assume here that we can load a lot of patients in memory at a time (after all, the whole point of this exercise is to keep the data in the database as much as possible), so we wrote a simple (but clearly not the most efficient!) function that retrieves the info for a given (set of) patient(s):

# Given a (set of) patient(s) ID(s) and a database connection,
# retreive his/her/their info and 
# return it as a data.frame (or NULL)
retreive_patients_info <- function(patient_ids, # a single value or a vector
                                   db_connection) # the database connection
{
  s <- dbGetQuery(db_connection, paste0(" SELECT event_date.id,
                                                 event_date.date,
                                                 event_info.category,
                                                 event_info.duration,
                                                 event_info.perday,
                                                 event_patients.patient_id,
                                                 patients.sex
                                          FROM event_date
                                           JOIN event_info
                                            ON event_info.id = event_date.id
                                           JOIN event_patients
                                            ON event_patients.id = event_info.id
                                           JOIN patients
                                            ON patients.id = event_patients.patient_id
                                          WHERE patients.id IN (",
                                        paste0(patient_ids,collapse=","),
                                        ");")); 
  if( nrow(s) < 1 )
  {
    return (NULL);
  } else
  {
    return (s);
  }
}

Compute CMA9 for all the patients and store it in the database

Also, for the sake of the exercise, we cannot assume that we can store the computed CMA values for all the patients in a vector in R, so, as soon as we estimate it for a single patient, we write it back to the database. For this, we will create a new table named cma9_estimates (with two columns: patient_id and cma) and we make sure it's empty:

dbExecute(med_events_db, "CREATE TABLE IF NOT EXISTS med_events.cma9_estimates (
                           patient_id INT NOT NULL,
                           cma REAL NULL,
                           PRIMARY KEY (patient_id));"); # create it if not already there
dbExecute(med_events_db, "TRUNCATE cma9_estimates;"); # make sure it's emty

For each patient in turn, extract his/her data from the database, estimate CMA9 and write back this estimate to the database (also, for the sake of illustration, we will assume that we can't even store the whole list of patient ID's, so we ask for each individual patient by selecting individual rows in the patients table):

# Time the whole thing:
start.time <- Sys.time();

# First (re)get the number of patients in the whole database:
no_patients <- dbGetQuery(med_events_db, "SELECT COUNT(DISTINCT id) FROM patients;");
no_patients <- as.numeric(no_patients[1,1]); # single value: convert it to numeric
no_patients;

# Select each patient in turn by its row number (remeber: rows count starts at 0 in SQL!):
for( i in 0:(no_patients-1) )
{
  # Get the patient ID in row i:
  patient_id <- dbGetQuery(med_events_db, 
                           paste0("SELECT id FROM patients LIMIT ",i,",1;"));
  if( is.null(patient_id) || nrow(patient_id) < 1 )
  { 
    warning(paste0("Cannot find patient number ",i+1,"!\n")); 
    next;
  }
  patient_id <- as.numeric(patient_id[1,1]); # single value: convert it to numeric

  # Extract the patient's data:
  s <- retreive_patients_info(patient_id, med_events_db); # retreive all the data
  if( is.null(s) )
  {
    warning(paste0("Cannot retrieve data for patient '",patient_id,"'!\n"));
    cma <- NA;
  } else
  {
    # Estimate the CMA:
    cma <- CMA9(data=s, # compute the CMA
                ID.colname="patient_id", 
                event.date.colname="date", 
                event.duration.colname="duration", 
                event.daily.dose.colname="perday", 
                medication.class.colname="category", 
                followup.window.start=0, 
                followup.window.start.unit="days", 
                followup.window.duration=2*365, 
                followup.window.duration.unit="days", 
                observation.window.start=30, 
                observation.window.start.unit="days", 
                observation.window.duration=1*365, 
                observation.window.duration.unit="days",
                date.format="%Y-%m-%d", 
                parallel.backend="none");
    if( is.null(cma) )
    {
      warning(paste0("Cannot estimate CMA for patient '",patient_id,"'!\n"));
      cma <- NA;
    } else
    {
      cma <- getCMA(cma)$CMA[1]; # retrieve the estimate
    }
  }

  # Write back this estimate (with replacement, 
  # if there's an already existing estimate for this patient):
  result <- dbExecute(med_events_db, 
                      paste0("REPLACE INTO cma9_estimates (patient_id, cma) VALUES (",
                             patient_id,
                             ", ",
                             ifelse(is.na(cma),"NULL",cma),");"));
}

end.time <- Sys.time();
cat(paste0("The SQL-mediated computation of CMA9 on the whole database took: ",
           round(difftime(end.time, start.time, units="secs"),2),
           " seconds\n"));

Also, do it within R so we can compare the results of the two procedures:

# Also, time it using the same procedure:
start.time <- Sys.time();
cma9r <- CMA9(data=med.events,
              ID.colname="PATIENT_ID", 
              event.date.colname="DATE", 
              event.duration.colname="DURATION", 
              event.daily.dose.colname="PERDAY", 
              medication.class.colname="CATEGORY", 
              followup.window.start=0, 
              followup.window.start.unit="days", 
              followup.window.duration=2*365, 
              followup.window.duration.unit="days", 
              observation.window.start=30, 
              observation.window.start.unit="days", 
              observation.window.duration=1*365, 
              observation.window.duration.unit="days",
              date.format="%m/%d/%Y", 
              parallel.backend="none");
end.time <- Sys.time();
cat(paste0("The computation of CMA9 in R on the whole database took: ",
           round(difftime(end.time, start.time, units="secs"),2),
           " seconds\n"));

# Retreive the results from the database:
cma9sql <- dbGetQuery(med_events_db, "SELECT * FROM cma9_estimates;");
knitr::kable(merge(getCMA(cma9r), 
                   cma9sql, 
                   by.x="PATIENT_ID", by.y="patient_id", all=TRUE),
             align=c("c","c","c"), 
             col.names=c("Patient ID", "CMA9 (R)", "CMA9 (MySQL)"),
             caption="The estimation of CMA9 on all the patients comparing 
             the `MySQL` and `R` estimates (they are, as expected, **identical**!).");

Plot a set of patients

Here we generate some plots of patients from the database.

# Extract the data of the patient(s) to plot:
# (we relax the stringency and assume we can hold a the patient IDs in memory :)):
s <- retreive_patients_info(patient_ids[1:5], med_events_db);
cma0 <- CMA0(data=s, 
             ID.colname="patient_id",
             event.date.colname="date", 
             event.duration.colname="duration", 
             event.daily.dose.colname="perday", 
             medication.class.colname="category",
             followup.window.start.unit="days", 
             followup.window.duration=2*365, 
             followup.window.duration.unit="days", 
             observation.window.start=30, 
             observation.window.start.unit="days", 
             observation.window.duration=1*365, 
             observation.window.duration.unit="days",
             date.format="%Y-%m-%d", 
             parallel.backend="none");
plot(cma0);

# Extract the data of the patient(s) to plot:
# (we relax the stringency and assume we can hold a the patient IDs in memory :)):
s <- retreive_patients_info(patient_ids[1:5], med_events_db);
cma9 <- CMA9(data=s, 
             ID.colname="patient_id",
             event.date.colname="date", 
             event.duration.colname="duration", 
             event.daily.dose.colname="perday", 
             medication.class.colname="category",
             followup.window.start.unit="days", 
             followup.window.duration=2*365, 
             followup.window.duration.unit="days", 
             observation.window.start=30, 
             observation.window.start.unit="days", 
             observation.window.duration=1*365, 
             observation.window.duration.unit="days",
             date.format="%Y-%m-%d", 
             parallel.backend="none");
plot(cma9);

Interactive plotting

For the interactive plotting of patients directly from the database (i.e., without local caching of data in R), we make use of the getter functions allowed by plot_interactive_cma through its function arguments get.colnames.fnc, get.patients.fnc and get.data.for.patients.fnc. As detailed in the help page for plot_interactive_cma, these arguments allow a user to use other types of data storage than the default data.frame by overriding the default behaviors for listing the data column names, listing all the patient IDs and getting the actual data for a (set of) patient ID(s), respectively.

Here, we redefined these functions as follows:

# Retreive the column names of the data for the first patient:
get.colnames.fnc.MySQL <- function(d)
{
  return(names(retreive_patients_info(
    as.numeric(dbGetQuery(med_events_db, 
                          "SELECT id FROM patients LIMIT 0,1;")[1,1]), 
    d)
  ));
}
# List all the patient IDs:
get.patients.fnc.MySQL <- function(d, idcol)
{
  s <- dbGetQuery(d, "SELECT id FROM patients;"); 
  return(s$id);
}
# Retreive the data for given (set of) patient ID(s):
get.data.for.patients.fnc.MySQL <- function(patientid, d, idcol)
{
  retreive_patients_info(patientid, d);
}

resulting in the code (not run here as it needs an interactive R session with AdhereR and Shiny):

plot_interactive_cma(data=med_events_db, # use the MySQL connection as the data provider
                     ID.colname="patient_id",
                     event.date.colname="date",
                     event.duration.colname="duration",
                     event.daily.dose.colname="perday",
                     medication.class.colname="category",
                     date.format="%Y-%m-%d", # SQL's DATE specification
                     backend=c("shiny","rstudio")[1], # use Shiny
                     use.system.browser=TRUE, # use the system web browser
                     # Override the getter functions to use the MySQL database 
                     # (please note that, for consistency, we must keep the 
                     # function argments evens if we don't use all of them):
                     get.colnames.fnc=get.colnames.fnc.MySQL,
                     get.patients.fnc=get.patients.fnc.MySQL,
                     get.data.for.patients.fnc=get.data.for.patients.fnc.MySQL
);

Disconnect from the database

Finally, we need to disconnect from the database:

dbDisconnect(med_events_db); 

Using a non-local MySQL database

Usually, the MySQL database is located on a different machine on a network (possibly even somewhere else on the Internet). We will show here how to access the same MySQL database we used so far (let's called it here the "server") from a different machine (a Macbook Air 11" running macOS High Sierra with R 3.4.4; called the "client").

First, make sure that the RMariaDB package is installed on the "client".

Second, make sure that the user adherentuser is allowed to remotely access the MySQL database; for our setup, we did the following on the "server" (as suggested in https://stackoverflow.com/questions/8380797/enable-remote-mysql-connection-error-1045-28000-access-denied-for-user and https://stackoverflow.com/questions/8380797/enable-remote-mysql-connection-error-1045-28000-access-denied-for-user/21151255#21151255):

• as root in MySQL (obtained by running, for example, mysql -u root -p in a terminal), execute:

GRANT ALL PRIVILEGES ON *.* 
 TO 'adherentuser'@'%' 
 IDENTIFIED BY 'AdhereR123!' 
 WITH GRANT OPTION;
FLUSH PRIVILEGES;

• also as root, edit the /etc/mysql/my.cnf (or the /etc/mysql/mysql.conf.d/mysqld.cnf on newer versions of MySQL) and replace the line bind-address = 127.0.0.1 by bind-address = 0.0.0.0
• restart MySQL (for example, by sudo service mysql restart.

Now, on the "client" we should be able to connect to the remote "server":

med_events_db <- dbConnect(RMariaDB::MariaDB(),     # works also for MySQL
                           user="adherentuser",     # the username
                           password="AdhereR123!",  # and password (insecure but ok here)
                           dbname='med_events',     # which database
                           host='remote.server.sql' # the host's name or IP address
                          );

and check that everything is OK:

db_res_tables <- dbListTables(med_events_db); db_res_tables;

followed by all the other things you would do with a local database, including closing it at the end (as shown above). Of course, the access and data transfer times will be worse than for a local database, but this can be addressed by judiciously selecting the patients one needs to process and by processing them in chunks of more than one patient.

Optimisations and security

Especially for remote databases (but also for local ones), the access and data transfer are potentially very slow; moreover, AdhereR is heavily optimized for processing (possibly in parallel) multiple patients (using data.table and other techniques). Therefore, probably the following ideas may help speed up the processing:

• as far as possible, information extraction and preparation should be done server-side (i.e., using SQL) instead of client-side (i.e., transferring raw data and processing it in R);
• transfer locally only the data that is strictly needed for computing the adherence and/or plotting (e.g., if computing CMA1, it doesn't make any sense to also transfer the column containing the daily dose) and only for the patients for which this needs to be done (i.e., perform an SQL pre-selection of the patients and events);
• to minimize server querying and transfer, and to maximally use the optimization built in AdhereR, split the whole set of patients to process into chunks that are relatively large (but not too large so as not to fit comfortably in the client's RAM even when processed in parallel);
• collect the computation results in R as much as possible and only write them back to the database in large chunks.

In these examples we stored the server name, username and password in clear in the R code: while this OK for this demo, it is a very bad idea in a production environment! There are various techniques for mitigating this, ranging from storing these info in a separate file that is read at run-time, to asking the user to interactively give the username and password, to using the keyring package (https://github.com/r-lib/keyring) -- see also https://db.rstudio.com/best-practices/managing-credentials/ for a discussion. Please chose the most appropriate one for your use-case, but whatever you do, don't store your credentials in the R script!!!

Use dplyr and DBI to transparently access the database

This is based on the more extended discussion in https://db.rstudio.com/dplyr/ . While using SQL for interacting with the database is very flexible and allows for fine-grained optimization, it requires a certain level of expertise and can be argued to create "chimeras" of R and SQL that might be more difficult to understand, debug and maintain.

The R packages dbplyr and DBI offer an alternative way, where the SQL is mostly "hidden" behind a familiar R code. Moreover, this allows to use the same code to access a variety of databases, such as MySQL, PostgreSQL and even Google’s BigQuery.

For example, we can connect and list the patients with (of course, you need to install first the needed packages, such as dbplyr, dplyr and RMySQL):

library(dplyr);
db_dplyer <- DBI::dbConnect(RMySQL::MySQL(),
                            user="adherentuser",     # the username
                            password="AdhereR123!",  # and password (insecure but ok here)
                            dbname='med_events'      # which database
); 

DBI::dbListTables(db_dplyer); # list the tables in the database

db_dplyer_patients <- tbl(db_dplyer, "patients"); # connect to the "patients" table
# db_dplyer_patients; # print it if you want to

Let's get the data from patient with id 1, compute CMA9 and plot it:

# connect to the various event info tables as well:
db_dplyer_event_date <- tbl(db_dplyer, "event_date");
db_dplyer_event_info <- tbl(db_dplyer, "event_info");
db_dplyer_event_patients <- tbl(db_dplyer, "event_patients");

# Select and join the various pieces of info for patient with patient_id == 1
# (Please note that the data data is actually transfered from MySQL to R only 
# at the end, when invoking collect()!!!)
s <- inner_join(inner_join(db_dplyer_event_info, 
                           # filter the events for patient 1:
                           db_dplyer_event_patients %>% filter(patient_id == 1), 
                           by="id"), # join with event_info
                db_dplyer_event_date, by="id") %>% collect(); 
cma9 <- CMA9(data=s, # compute CMA9 as usual
             ID.colname="patient_id",
             event.date.colname="date", 
             event.duration.colname="duration", 
             event.daily.dose.colname="perday", 
             medication.class.colname="category",
             followup.window.start.unit="days", 
             followup.window.duration=2*365, 
             followup.window.duration.unit="days", 
             observation.window.start=30, 
             observation.window.start.unit="days", 
             observation.window.duration=1*365, 
             observation.window.duration.unit="days",
             date.format="%Y-%m-%d", 
             parallel.backend="none");
plot(cma9); # plot it

Finally, close the connection to the database:

DBI::dbDisconnect(db_dplyer);

How about SAS and Stata

If the dataset is relatively small, probably the easiest is to export it into a format that R can read (such as CSV); if this is not possible, then maybe one of the methods for importing their native formats in R may work (see, for example, https://www.statmethods.net/input/importingdata.html or https://cran.r-project.org/doc/manuals/r-devel/R-data.html#EpiInfo-Minitab-SAS-S_002dPLUS-SPSS-Stata-Systat ).

However, if the dataset is (very) large, then probably one should either:

• export it into a RDBMS such as MySQL where R can access it (as described above), or
• experimentally, use ODBC to access it directly from SAS (which might or might not work).

Finally, let's look at Hadoop and MapReduce!

Apache Hadoop is a general framework for the distributed processing of large datasets using a simple programming paradigm (MapReduce; see, for example, https://wiki.apache.org/hadoop/ProjectDescription ). Given that AdhereR can compute adherence estimates for sets of at least one patient at a time, it is very easy to embed within the MapReduce framework using a variety of approaches (see, among others, the outlines here or here).

Just as an example, we will install here Hadoop 3.0.3 on the Ubuntu 18.04 machine and use it to compute CMA9 on all the patients.

Installing Hadoop 3.0.3 on Ubuntu 18.04

We are following the tutorial here. All of these happen in a shell terminal.

Install Java

sudo apt update
sudo apt install default-jdk
java -version

Install Hadoop 3.0.3

Download the 3.0.3 binary version:

cd ~/Downloads
wget http://mirrors.ircam.fr/pub/apache/hadoop/common/hadoop-3.0.3/hadoop-3.0.3.tar.gz

then unpack and move it to /usr/local/:

tar -xzvf hadoop-3.0.3.tar.gz
sudo mv hadoop-3.0.3 /usr/local/hadoop

Configure Hadoop

Add to /usr/local/hadoop/etc/hadoop/hadoop-env.sh the line export JAVA_HOME=$(readlink -f /usr/bin/java | sed "s:bin/java::").

Test Hadoop

Start it first:

/usr/local/hadoop/bin/hadoop

and test it:

/usr/local/hadoop/bin/hadoop jar /usr/local/hadoop/share/hadoop/mapreduce/hadoop-
mapreduce-examples-3.0.3.jar grep ~/input ~/grep_example 'allowed[.]*'
cat ~/grep_example/*

which should produce:

Output
19  allowed.
1   allowed

Installing RHadoop on Ubuntu 18.04

We are adapting here various guidelines for RedHat and macOS. All these happen in R (unless specified).

Set the needed environment variables

Sys.setenv("HADOOP_HOME"="/usr/local/hadoop"); # Hadoop HOME
Sys.setenv("HADOOP_CMD"="/usr/local/hadoop/bin/hadoop"); # Hadoop binary
Sys.setenv("HADOOP_STREAMING"=   # Streaming
             "/usr/local/hadoop/share/hadoop/tools/lib/hadoop-streaming-3.0.3.jar");

Install the needed R packages

(You probably already have some installed).

install.packages(c("rJava", 
                   "Rcpp", 
                   "RJSONIO", 
                   "bitops", 
                   "digest", 
                   "functional", 
                   "stringr", 
                   "plyr", 
                   "reshape2"), 
                 dep=TRUE);

Now, manually download the rhdfs, rhbase and rmr2 packages from Revolution Analytics' GitHub repository. In a shell terminal:

cd ~/Downloads
wget https://github.com/RevolutionAnalytics/rmr2/releases/download/3.3.1/rmr2_3.3.1.tar.gz
wget https://github.com/RevolutionAnalytics/rhdfs/blob/master/build/rhdfs_1.0.8.tar.gz?raw=true
wget https://github.com/RevolutionAnalytics/rhbase/blob/master/build/rhbase_1.2.1.tar.gz?raw=true

and then, in R:

install.packages(c("~/Downloads/HADOOP/rhdfs_1.0.8.tar.gz?raw=true", 
                   "~/Downloads/HADOOP/rmr2_3.3.1.tar.gz", 
                   "~/Downloads/HADOOP/rhbase_1.2.1.tar.gz"), 
                 repos=NULL, dep=TRUE);

Using Hadoop through RHadoop to compute CMA9

With all these in place, we can compute CMA9 for each participant, as follows (of course, this is not optimized in any sense). Everything here takes place in R.

Load the libraries and initialize things

library(rhdfs); # access to HDFS
hdfs.init(); # initialize HDFS

library(rmr2); # MapReduce in R

Store the med.events data in HDFS

First, we will store the med.events example AdhereR dataset in HDFS (the Hadoop Distributed File System) as a set of key-value pairs:

# Transfer med.events to HDFS as a set of key-value pairs:
med_events_hdfs <- to.dfs(keyval(med.events$PATIENT_ID, med.events));

This is not a very efficient and scalable way of doing it, but it works for this example (normally, the data would be transferred to HDFS using specialized tools such as Sqoop). What this does is create a handler (med_events_hdfs) to a temporary HDFS storage where the rows of the med.events data.frame (the values) are stored paired with the corresponding PATIENT_IDs (the keys). Importantly the med_events_hdfs handler ensures that the actual data is never fully loaded in memory, which is very important for large databases.

Use MapReduce to compute CMA9

s <- mapreduce(input=med_events_hdfs, # read the data from med_events_hdfs
               # the mapping function here doesn't do much; 
               # k (= the key) = PATIENT_ID; v (= value) = med.events rows:
               map=function(k,v) keyval(k,v), 
               # the reduce function computes CMA9 on the data (the values v) 
               # associated with a given PATIENT_ID (the key k):
               reduce=function(k,v)  
                 {
                  cma <- CMA9(v,
                             ID.colname="PATIENT_ID", 
                             event.date.colname="DATE", 
                             event.duration.colname="DURATION", 
                             event.daily.dose.colname="PERDAY", 
                             medication.class.colname="CATEGORY", 
                             followup.window.start=0, 
                             followup.window.start.unit="days", 
                             followup.window.duration=2*365, 
                             followup.window.duration.unit="days", 
                             observation.window.start=30, 
                             observation.window.start.unit="days", 
                             observation.window.duration=1*365, 
                             observation.window.duration.unit="days",
                             date.format="%m/%d/%Y", 
                             parallel.backend="none");
                  # Return the computed CMA as the value paired with the PATIENT_ID key:
                  keyval(k, getCMA(cma)[1,"CMA"]); 
                 });

The basic idea is to first map the (key, value) pairs in the input (here, the PATEINT_IDs with their associated event info) to (potentially) different (key, value) pairs (here, we don't do basically anything at this stage), followed by reducing all the data associated with a key to a summary value (here, CMA9), returning a new summary (key, value) pair. The result s is also a pointer to the actual results stored in HDFS (temporarily), so that again we don't load anything large in memory.

Load and use the results

Just for the sake of illustration, only now do we load the full results in memory for display as a data.frame (to be compared with the other ways of computing CMA9):

# Only *now* force the loaing of the full results in memory!
s <- from.dfs(s); 

# Convert the (key, value) pairs to a nice data.frame format:
cma9hadoop <- data.frame("PATIENT_ID"=s$key, "CMA9"=s$val);

# And plot them for comparison with the other methods discussed here:
knitr::kable(merge(merge(getCMA(cma9r), 
                         cma9sql, 
                         by.x="PATIENT_ID", by.y="patient_id", all=TRUE),
                   cma9hadoop, 
                   by="PATIENT_ID", all=TRUE),
             align=c("c","c","c","c"), 
             col.names=c("Patient ID", "CMA9 (R)", "CMA9 (MySQL)", "CMA9 (Haddop)"),
             caption="The estimation of CMA9 on all the patients comparing the `R`, 
             `MySQL` and `Hadoop` estimates (they are, as expected, **identical**!).");

Conclusions

We hope to have shown here that, far from being constrained by the "'R' can't process large datasets that don't fit in memory" myth, AdhereR can, in fact, nicely and easily interface with both "classical" relational databases (such as MySQL)) and with newer technologies (such as Apache Hadoop), allowing it to access potentially huge databases (even across the Internet) and process them efficiently (even across massively parallel heterogeneous computational infrastructures).

Thus, AdhereR can seamlessly scale up from being used for the real-time visualization and computation of adherence of a small-to-medium local database (stored either in a "flat" CSV file or in an SQLite database) on a consumer-grade laptop, up to running on large heterogenenous computer clusters where it can process huge amounts of data stored across several RDBMS tables (managed by, for example, MySQL) or as (key, value) pairs in a Hadoop HDFS file system. The examples (with actual code and step-by-step instructions) presented here should provide a starting point for real-world implementations optimized for the problems at hand.

There is a vast literature concerning both "classic" SQL and newer NoSQL databases, and there is an already mature ecosystem for accessing them from R. However, even for very specific cases where this is not feasible, we have provided a flexible framework for seamlessly using AdhereR from other programming languages (such as Python) for the computation and plotting of adherence.

[^1]: We provide only the PDF within the package itself as the HTML form is rather big due to the embedded images and generates a NOTE when building the package. Moreover, to void an unsuccessful attempt at compiling the vignette during the package build, the actual .Rmd source and related images are in the specialVignettes subfolder of the package.

[^2]: Despite the development of alternative architectures (generically known as "NoSQL"; see @harrison_next_2015), RDBMSs are still extremely popular and will very probably continue to be so for the foreseeable future.

References

AdhereR documentation built on July 5, 2022, 5:08 p.m.