Writing custom extensions

cohortBuilder package is adapted to work with various data sources and custom backends. Currently there exists one official extension cohortBuilder.db package that allows you to use cohortBuilder with database connections.

The goal of this document is to explain how to create custom extensions to cohortBuilder.

In general to create the custom layer you need to:

Define set of S3 methods operating on your custom source.
Define selected filters (see vignette("custom-filters")).

It’s recommended to include all of the methods in your custom R package.

Before you start creating a new layer, you need to choose what data (connection) should your layer operate on.

For example, cohortBuilder uses tblist class object to operate on list of data frames , or db class for operating on database connections.

To start with create a function that will take required parameters to define data connection, such as tblist or dbtables in case of cohortBuilder.db. The function should return an object of selected class which is used to define required extension methods.

Below we describe all the required and optional methods you need to define within the created package.

set_source - method used for defining a new source

Required parameters:

dtconn

Details:

Define the method calling Source$new inside.
Declare necessary parameters used for defining data source (and pass them to Source$new). The arguments are then available at source$attributes object.
If valid, you may declare primary_keys and binding_keys parameters (see vignette("binding-keys")).
It’s also worth to declare source_code parameter that allows users to define code for creating source (visible in reproducible code) and description storing list of useful source objects descriptions.

Example:

cohortBuilder - tblist object (same for cohortBuilder.db - db object)

set_source.tblist <- function(dtconn, primary_keys = NULL, binding_keys = NULL, 
                           source_code = NULL, description = NULL, ...) {
  Source$new(
    dtconn, primary_keys = primary_keys,
    binding_keys = binding_keys, source_code = source_code,
    ...
  )
}

.init_source - structure data passed between filtering steps

Required parameters:

source - Source object

Details:

Within the function body define how data should be extracted from source, and structured.
A structure of the output is then used as an input for filter’s method (data_object argument).

Examples:

cohortBuilder - ‘tblist’ class. Operating on list of tables in each step.

.init_step.tblist <- function(source, ...) {
  source$dtconn
}

cohortBuilder.db - ‘db’ class.

cohortBuilder.db operates on db class object which is list of connection, tables and schema fields.

.init_step.db <- function(source) {
  purrr::map(
    stats::setNames(source$dtconn$tables, source$dtconn$tables),
    function(table) {
      tbl_conn <- dplyr::tbl(
        source$dtconn$connection,
        dbplyr::in_schema(source$dtconn$schema, table)
      )
      attr(tbl_conn, "tbl_name") <- table
      tbl_conn
    }
  )
}

.pre_filtering (optional) - modify data object before filtering

Required parameters:

source,
data_object - an object following the structure of .init_step,
step_id - id of the filtering step

Details:

An optional method used to prepare data for filtering (cleaning up attributes, creating middle step objects etc.).

Examples:

cohortBuilder - tblist class. Cleaning up filtered attribute for new step data.

.pre_filtering.tblist <- function(source, data_object, step_id) {
  for (dataset in names(data_object)) {
    attr(data_object[[dataset]], "filtered") <- FALSE
  }
  return(data_object)
}

cohortBuilder.db - creating temp tables for the current step in database and cleaning up filtered attributes.

.pre_filtering.db <- function(source, data_object, step_id) {
  purrr::map(
    stats::setNames(source$dtconn$tables, source$dtconn$tables),
    function(table) {
      table_name <- tmp_table_name(attr(data_object[[table]], "tbl_name"), step_id)
      DBI::dbRemoveTable(source$dtconn$connection, table_name, temporary = TRUE, fail_if_missing = FALSE)
      data_object[[table]] <- dplyr::compute(
        data_object[[table]],
        name = table_name
      )
      attr(data_object[[table]], "filtered") <- FALSE
      return(data_object[[table]])
    }
  )
}

.post_filtering (optional) - data object modification after filtering (before running binding).

Required parameters:

source,
data_object - an object following the structure of .init_step,
step_id - id of the filtering step

.collect_data - define how to collect data object into R.

Required parameters:

source,
data_object - an object following the structure of .init_step

Details:

cohortBuilder’s equivalent of collect method known for sourcing the object into R memory when working with remote environment (e.g. database).
When operating in R memory it’s enough to return data_object.

Examples:

cohortBuilder - operating in R memory, so return data_object.

.collect_data.tblist <- function(source, data_object) {
  data_object
}

cohortBuilder.db - collect tables from database and return as a named list.

.collect_data.db <- function(source, data_object) {
  purrr::map(
    stats::setNames(source$dtconn$tables, source$dtconn$tables),
    ~dplyr::collect(data_object[[.x]])
  )
}

.get_stats - collect data object stats

Required parameters:

source,
data_object

Details:

There are no obligatory statistics to be returned.
The calculated statistic is cached within Cohort object after each step filtering.
Returned stats plays mostly read-only role. The only situation the stats are used is within .get_attrition_count and shinyCohortBuilder integration.

Examples:

cohortBuilder - operating in R memory, so return data_object.

.get_stats.tblist <- function(source, data_object) {
  dataset_names <- names(source$dtconn)
  dataset_names %>%
    purrr::map(
      ~ list(n_rows = nrow(data_object[[.x]]))
    ) %>%
    stats::setNames(dataset_names)
}

cohortBuilder.db - collect tables from database and return as a named list.

.get_stats.db <- function(source, data_object) {
  dataset_names <- source$dtconn$tables
  dataset_names %>%
    purrr::map(
      ~ list(
        n_rows = data_object[[.x]] %>%
          dplyr::summarise(n = n()) %>%
          dplyr::collect() %>%
          dplyr::pull(n) %>%
          as.integer()
      )
    ) %>%
    stats::setNames(dataset_names)
}

.run_binding - method defining how binding should be handled

Required parameters:

source,
binding_key - binding key definition,
data_object_pre - data object state before filtering in the current step,
data_object_post - data object state after filtering in the current step (including effect of previous bindings)

Details:

Binding is run after applying all the filters in the current filtering step (and applying .post_filtering if defined).
When filtering is finished, binding loop starts iterating over all the defined bindings. .run_binding takes care of handling a single iteration.
The returned bound data, should preserve the structure of .init_step method output.
You should preserve the assumed binding rules described in bindings-keys (i.e. handling post = TRUE/FALSE, activate = TRUE/FALSE and filtered attribute) but this is not obligatory.

Examples:

cohortBuilder

.run_binding.tblist <- function(source, binding_key, data_object_pre, data_object_post, ...) {
  binding_dataset <- binding_key$update$dataset
  dependent_datasets <- names(binding_key$data_keys)
  active_datasets <- data_object_post %>%
    purrr::keep(~ attr(., "filtered")) %>%
    names()

  if (!any(dependent_datasets %in% active_datasets)) {
    return(data_object_post)
  }

  key_values <- NULL
  common_key_names <- paste0("key_", seq_along(binding_key$data_keys[[1]]$key))
  for (dependent_dataset in dependent_datasets) {
    key_names <- binding_key$data_keys[[dependent_dataset]]$key
    tmp_key_values <- dplyr::distinct(data_object_post[[dependent_dataset]][, key_names, drop = FALSE]) %>%
      stats::setNames(common_key_names)
    if (is.null(key_values)) {
      key_values <- tmp_key_values
    } else {
      key_values <- dplyr::inner_join(key_values, tmp_key_values, by = common_key_names)
    }
  }

  data_object_post[[binding_dataset]] <- dplyr::inner_join(
    switch(
      as.character(binding_key$post),
      "FALSE" = data_object_pre[[binding_dataset]],
      "TRUE" = data_object_post[[binding_dataset]]
    ),
    key_values,
    by = stats::setNames(common_key_names, binding_key$update$key)
  )
  if (binding_key$activate) {
    attr(data_object_post[[binding_dataset]], "filtered") <- TRUE
  }

  return(data_object_post)
}

cohortBuilder.db - not implemented yet

.get_attrition_count - define how to get metric used for attrition data plot

Required parameters:

source,
data_stats - statistics related to each step data - list of .get_stats results for each step (and original data, assigned to step_id = 0),

Details:

The output should return vector of length n+1 where n is number of steps. The first element of the vector should describe statistic for the base, unfiltered data.
You may define additional parameters which can be passed to attrition method of Cohort object (e.g. dataset in the below example).

Examples:

cohortBuilder

.get_attrition_count.tblist <- function(source, data_stats, dataset, ...) {
  data_stats %>%
    purrr::map_int(~.[[dataset]][["n_rows"]])
}

cohortBuilder.db - not implemented yet

.get_attrition_label - define label displayed in attrition plot for the specified step

Required parameters:

source,
step_id - id of the step ("0" for original data case),
step_filters - list storing filters configuration for the selected step (NULL for original data case),

Details:

Please remember to define label for the initial data state (step_id = "0" case).
You may define additional parameters which can be passed to attrition method of Cohort object (e.g. dataset in the below example).

Examples:

cohortBuilder

get_attrition_label.tblist <- function(source, step_id, step_filters, dataset, ...) {
  pkey <- source$primary_keys
  binding_keys <- source$binding_keys
  if (step_id == "0") {
    if (is.null(pkey)) {
      return(dataset)
    } else {
      dataset_pkey <- .get_item(pkey, "dataset", dataset)[1][[1]]$key
      if (is.null(dataset_pkey)) return(dataset)
      return(glue::glue("{dataset}\n primary key: {paste(dataset_pkey, collapse = ', ')}"))
    }
  }
  filters_section <- step_filters %>%
    purrr::keep(~.$dataset == dataset) %>%
    purrr::map(~get_attrition_filter_label(.$name, .$value_name, .$value)) %>%
    paste(collapse = "\n")
  bind_keys_section <- ""
  if (!is.null(binding_keys)) {
    dependent_datasets <- .get_item(
      binding_keys, attribute = "update", value = dataset,
      operator = function(value, target) {
        value == target$dataset
      }
    ) %>%
      purrr::map(~names(.[["data_keys"]])) %>%
      unlist() %>%
      unique()
    if (length(dependent_datasets) > 0) {
      bind_keys_section <- glue::glue(
        "\nData linked with external datasets: {paste(dependent_datasets, collapse = ', ')}",
        .trim = FALSE
      )
    }
  }
  gsub(
    "\n$",
    "",
    glue::glue("Step: {step_id}\n{filters_section}{bind_keys_section}")
  )
}

cohortBuilder.db - not implemented yet