What v1.10.0 adds, annotated

The tab reads the loaded study’s subject-level and adverse events files, the same two the safety charts read, through the column mapping on the Data tab. There is no second study to load. On the pilot study those two files support three of gsm’s eight site metrics: adverse event rate, serious adverse event rate and study discontinuation rate.

Before the press. The tab says what it will run on, what starting R downloads and from where, and that the files stay in the browser. Nothing has been asked of R’s addresses yet.

Opened, the line lists what each file of the study gives gsm and which metrics that supports. It is worked out from the files’ columns before R starts, and R says the same when it runs.

While R starts, the tab says which step it is on and counts the seconds.

After the press: “R ran 3 of 8 metrics on the loaded study’s adsl.csv and adae.csv in 3.3 seconds, 28 seconds after Start R was pressed.” The overview has a row for each of the 17 sites, and site 705 leads it with a red flag for adverse events. Each tick and arrow is the flag R returned for that site and metric; the red and amber counts are gsm.viz’s count of them.

A metric that could not run says why

Five metrics need records the pilot study does not have: protocol deviations, lab toxicity grades, treatment completion and screening. Each stays in the row of metrics, marked “did not run”, and says which file it needs.

Study discontinuation rate, the third metric the pilot study supports.

“Screen Failure Rate needs Raw_ENROLL.csv, which is not loaded.” The sentence is R’s own.

It follows the mapping on the Data tab

R is handed only the columns the metrics need, under the app’s standard names, read through your mapping. So a study with its own column names works once its site, end-of-study day and end-of-study status are mapped. Clear a row of the mapping and R runs the study again at once, and names the column.

With the site’s row cleared: “Adverse Event Rate needs the column SITEID, which no column of adsl.csv is mapped to.” The file may hold the column under another name, so the tab says it is not mapped, not that it is missing.

On a phone

The overview at 390 pixels wide. Nothing scrolls sideways.

The metric buttons and a chart, stacked.

A study that has gsm’s raw files runs every metric. The demo app carries one as a fourth demo study, and the tab takes your own files too. A raw file that is loaded is used as it is; the study’s own files fill in only for a table that is not loaded.

The RBQM study: nine raw files in gsm’s format, for a synthetic study of 150 sites and 765 enrolled participants. “R ran 8 of 8 metrics on the 9 loaded files in 5 seconds, 30 seconds after Start R was pressed.” Here a site is labelled with its investigator’s name, because these files have one.

Three files chosen from disk: “3 files loaded, 2 placed in a gsm raw domain. They support 2 of 8 metrics.” A file the tab does not recognise is named and stays out of R.

R computes every rate, score and flag the tab shows. The app’s JavaScript decides when R is asked, hands R’s tables to gsm.viz and writes the sentences. It computes no rate, score or flag. gsm.viz, gsm’s own chart code, counts each site’s red and amber flags for the overview’s two count columns and orders the rows by them.

Whose code runs

• gsm.kri is not in the tab’s list of R’s versions because R cannot report a package it has not installed. The tab now names it in the sentence after that list.
• The steps gsm.kri’s workflows name (input, transform, analyze, flag, summarize) are gsm.core’s functions. Installing gsm.kri would also bring the packages it uses for tables and reports, which the tab does not need.
• Every copy is held to its release tag by a check that runs on every change. The four packages are built from their tags; a build cannot be compared with its source, so each is held to the checksum recorded when it was built.

How the loaded study becomes gsm’s tables

Five small workflows, in the same format as gsm’s own and run by the same engine in R, make gsm’s raw tables from the app’s standard columns. These are choices, and each can change:

• Days on study is the study’s end day.
• A participant completed the study when the end-of-study status is COMPLETED, and left it when it is DISCONTINUED.
• Everyone in the subject-level file counts as enrolled.
• An adverse events row with no term is a participant with no event, and is left out, as the app’s charts leave it out.
• The pilot study names no investigators, so a site is labelled with its number.
• The pilot study’s subject-level file gained one column for this, the site. Every other column of it is unchanged.

What it downloads, and what leaves

• On this capture, starting R downloaded 55.2 MB: 13.0 MB of R from webr.r-wasm.org, 40.0 MB of packages from repo.r-wasm.org, and 2.2 MB of gsm’s packages from the page’s own address. The tab says about 55 MB.
• The first result came 29 seconds after the press on this capture, over a home connection. On the test runner, which has a faster one, it takes 30 to 50 seconds. A later run takes a few seconds, because R is already up.
• Before the press, nothing is asked of R’s addresses.
• Your files are never sent. A test watches every request the tab makes on the pilot study and holds each to a read with no body, to three addresses: the page’s own, webr.r-wasm.org and repo.r-wasm.org. No address carries a participant’s ID.
• The single-file build of the app cannot start R, and its RBQM tab says so.

What R tells you when a study is not clean

• R’s warnings are listed under the sentence. A participant with no site is left out of every metric, and the tab says so in R’s words: “R warned: 26 cases of NA's in GroupID, cases are removed in output.”
• A file with none of its columns mapped does not stop the run: each metric names the columns it needs.
• When gsm’s reporting stops, as it does when no participant has a site, each metric that ran says so with R’s message.
• A cell that reads NA is a missing value, as R writes one.
• A participant listed twice is counted twice, and nothing says so yet (the task for it).

These came from the release review, which ran R on studies the tests had not tried. The first three were fixed before this page was put in front of the owner.

How the numbers are checked

• R in the browser is held to R on a desktop: all 1,186 site rows of the eight metrics on the RBQM study, and the 51 rows of three metrics on the pilot study, agree to eight decimal places.
• Every flag in the overview table is read back against the row R returned, and each count of flags against R’s flags counted.
• On the pilot study, R’s totals are counted another way from the files: 1,122 adverse events, 3 of them serious, and 144 of 254 participants who left the study early.
• The tab’s tests run in CI on every change, and their pictures are in the repository (the demo app’s evidence folder). The demo app has no evidence page on the site yet, as each chart has (the task for one).

Each chart links to its guide and its evidence

Under every chart in the demo app, a footnote links the chart’s clinical guide, where it has one, and its test evidence. The links open in a new tab, so a study you loaded stays loaded.

“Hepatic Safety Explorer: Clinical guide · Test evidence” Six charts have a clinical guide; every chart has test evidence.

The histogram’s two p-value settings are removed

The v1.9.1 release notes said v1.10.0 would remove test_normality and compare_distributions, and it does. Each put a p-value on the chart that safety.viz worked out in JavaScript with a shortcut. The histogram now draws no p-value; a test belongs in R. Three older charts still make a statistical call in JavaScript, which is not new in this release and waits for a ruling (the interval in the adverse events explorer, the threshold mark in the QT explorer and the band on the time-to-event chart).

The histogram given both removed settings. It draws as it would without them: no p-value on the main chart or on a panel.