Analyse the data

The starting point: mortality_dataset.csv, 111 stays and 21 variables, with each column's statistics on the right. Click to enlarge.

Summary

The Lab is a project's analysis space. It starts from a dataset in wide format — one row per stay, one column per variable — onto which you attach analyses configured without writing code (descriptive table, regression), plus an IDE for R and Python for anything that needs to be bespoke.

This page continues the journey from the data warehouse: the warehouse is for exploring and selecting, the Lab for analysing. The data is still the MIMIC-IV Demo.

Datasets

A dataset is a table in wide format: one row per stay, one column per variable. This is the shape nearly every statistical tool expects — and the opposite of the warehouse's long format, where each measurement takes up its own row. The article Organising your data covers both shapes and how to move between them.

The demo dataset holds 111 stays and 21 variables: age, sex, admission unit, length of stay, then the lab extrema of the first 24 hours (lactate, creatinine, bicarbonate, Glasgow coma score…) and the variable to predict, in_hospital_death.

Selecting a column shows its completeness, statistics and distribution in the right-hand panel — mean, median, quartiles, box plot and histogram. This is the first thing to look at before any analysis: a variable with 39% missing values cannot be treated like a complete one.

The age column selected: 100% complete, a median of 63, and the distribution of all 111 stays as a histogram and box plot.

This dataset was produced by the project's R script, which queries the MIMIC-IV database and aggregates the measurements of the first 24 hours. A dataset can also be imported straight from a CSV, Excel or Parquet file.

Analyses

Below the file list, the Analyses panel gathers the analyses attached to the dataset. The demo has two: a Tableau descriptif (descriptive table) and a Regression. Each one is a form on the left, a result on the right.

The descriptive table grouped by in_hospital_death: 96 survivors against 15 deaths, each variable summarised as median [IQR] or count (%).

The descriptive table — the Table 1 of every clinical paper — is configured in a few fields: the variables to include, the Group by variable, and the numeric summary (here median [IQR]). Numeric variables are summarised by their median and interquartile range, categorical ones by count and percentage, and missing values get their own row when Missing row is ticked.

The contrast between the two groups can be read directly: patients who died are older (66 versus 61), stay longer (4.9 days versus 2.3) and have higher lactate and creatinine. The Export button produces the table ready for publication.

Analyses are recomputed each time they are opened, from the current dataset: they follow the data rather than freezing a copy of it.

IDE

Not everything fits in a form. The built-in IDE opens the project's scripts — R Markdown, Jupyter notebooks, SQL, Python, R — and runs them in the browser, with nothing to install.

The 01_eda_mortality.Rmd script mid-run: a SQL cell queries MIMIC-IV, its result appears below, and the document outline is listed on the right.

The project ships five scripts, including the exploratory analysis in R Markdown above and a modelling notebook in Python. Cells run one at a time (Run cell and advance), alternating prose, code and results, and the outline on the right makes a long document navigable.

The key point: the dropdown at the top names the database being queried — here MIMIC-IV Demo. So a SQL cell hits the warehouse directly, and its result comes back as a table the following R code can reuse. That chain is what produces the dataset at the top of this page.

In client-only mode, R and Python run in the browser (WebR and Pyodide), which covers most analysis work. For full environments with your own packages, you need a deployment with a backend.

Going further

Explore a health data warehouse — where the data analysed here comes from.
Long format and wide format — why the warehouse and the Lab do not use the same table shape.
Deployment modes — what runs in the browser and what needs a server.