Session 5: Live-demo script

Download the PDF · Back to Session 5

Foundations of Clinical Research · Week 5, Saturday 19:30 to 22:30 (online) · Neudata · instructor script

Part A: cleaning Part B: Table 1 and figures
Slot 5.2 Meet the course dataset (about 20:25) 5.4 Describe the course dataset (about 21:31)
Duration 8 minutes 6 minutes
Script Course/Data/R/01_clean.R Course/Data/R/02_describe.R
Participants watch your shared screen and predict in the chat (they do not run R) same

FLUID-ICU is SIMULATED. Say so every time you open the data: no real patient, hospital or country. All numbers below are from Course/Data/key_results.md and Course/Data/cleaning_log.md.

Before the session (night before + 10 minutes before)

Rules of the demo: run one block at a time (Ctrl+Enter). Before each block, ask the class “what do you expect?” and read two or three chat answers After it, say what was fixed and that it goes in the log. Never type new code live, because the scripts are complete.

Part A: From raw to clean (01_clean.R, 8 minutes)

Step (script block) What you show Expected output Say
0. Read raw as text dim(raw), glimpse(raw) 623 rows, 29 columns, everything <chr> “The study enrolled 620 patients. Why 623 rows?”
1. Column names names(dat) short lower-case names (map_mmhg, vent_free_days_28) “Names without spaces: same in every script.”
2. Duplicates get_dupes(study_id) 3 study IDs appear twice, identical rows; after distinct() 620 “Each patient counted once.”
3. Sites count(site) before/after variants such as icu a, ICU-A: become 4 sites “Otherwise R sees many different ICUs.”
4. Dates the three formats 2023-03-14, 14/03/2023, 14-Mar-2023: become one date; 0 failures “Is 03/04 March or April? Never guess.”
5–6. Sex and yes/no tabyl() before/after many spellings (M, m, male, F, f, ” Female”; Y, 1, yes) become Male/Female, Yes/No “1 and Yes must mean the same thing.”
7–8. Numbers and codes summary() of lactate, MAP, APACHE II maxima of 999, minima of -99: set to NA “A lactate of 999 would wreck the mean.”
9. Units lactate median by ICU ICU C about 9 times higher (mg/dL): divided by 9.01 “Found only because we summarised by ICU. Ask the site lab.”
10. Impossible values range checks age 210, weight 7 kg, a negative fluid output: set to NA + query; patient kept “We cannot guess 21 or 71. We never delete the patient.”
12. Derived variables the fluid-balance formula cum_fluid_balance_d3_ml_kg, positive_balance_d3, aki_any computed “Derived values are calculated, never typed.”
13. What is left colSums(is.na(dat)) lactate 93 missing, a handful elsewhere “These are REAL gaps. We handle them in Session 6.”
14. Reproducibility final cat() Cleaned data identical to the reference clean file: TRUE / byte-for-byte identical: TRUE “Anyone can re-run this and get exactly the same data.”

If short of time: run Steps 0, 2, 9, 10 and 14 only (4 minutes) and show the cleaning log for the rest.

Close Part A by opening cleaning_log.md: “Every block you saw is one row here: problem, how many, what we did, why.” Then show the slide From 623 rows to 527 complete cases (620 enrolled, then 618 with a day-3 balance, then 527 complete cases).

Part B: Table 1 and figures (02_describe.R, 6 minutes)

Block What you show Expected output (key_results.md) Say
1. Load dim(icu), count(positive_balance_d3) 620 patients, positive balance 373/618 (60.4%), 2 without a balance “618, not 620: two patients have no computable balance.”
2. Shape first mean/SD of age, median/IQR of lactate; quick histograms age mean 62.6 (SD 15.2), lactate skewed “Symmetric: mean (SD). Skewed: median (IQR).”
3. Table 1 the gtsummary table by group negative n = 245, positive n = 373; SOFA 6 (4 to 8) vs 8 (6 to 10); vasopressor 24.9% vs 55.0%; ventilation 39.6% vs 57.9%; lactate 2.2 vs 2.9 mmol/L; heart failure 22.0% vs 11.5%; missing counts shown; no p-values “Were the groups alike at the start? No: the positive group was sicker.”
4. Histogram fluid balance by group median 8.8 ml/kg (IQR -10.3 to 26.2) “Continuous exposure, cut at 0.”
5. Boxplot SOFA by group, with dots medians 6 vs 8 “Show every patient.”
6. Bar chart infection source, sorted respiratory most common “Categories: sorted bars, not a 3-D pie.”
7. Outcomes described, not compared 28-day mortality 179/620 (28.9%), ICU stay mean 7.3 vs median 6.0 days “Described today, compared next Saturday.”

Close Part B: “This Table 1 is in every results pack, Q1 to Q5. Now we read it together as a reviewer would.” Then move to the whole-class reading of Table 1 (Exercise 6).

Fallback plan

Problem Do this
R or a package fails Show raw_vs_clean.png and talk through the problems (Part A); show table1_snapshot.png, fluid_balance_hist.png, sofa_by_group.png (Part B). Teaching points are identical.
Screen sharing does not show RStudio legibly Use the PNG figures and cleaning_log.md in a browser at 150% zoom.
Running late Part A: Steps 0, 2, 9, 10, 14. Part B: blocks 3 and 5 only.
A participant asks to run it themselves Point to Course/Data/README.md and the free Clinical Data Analysis in R course, and offer to help after the session.

Anticipated questions

“Is this real data?” No. It is simulated, designed to behave like ICU data and to show the classic problems. Any manuscript must say so.

“Why not fix the errors in Excel?” Hand edits leave no trace and cannot be repeated, and one wrong paste is invisible. A script plus a log is checkable by a reviewer and repeatable by anyone.

“Why set age 210 to missing instead of guessing 21?” It could be 21, 71 or 10, and guessing invents data. Set to missing, query the site, correct only with the source document.

“Why no p-values in Table 1?” Table 1 describes the groups. The SAP says so explicitly. In an RCT any difference is chance. In a cohort it is confounding to be adjusted for, not tested.