Step-by-step Causal Analysis Of Ehrs To Ground Decision Making

Overview

This code was used to study the effect of a combination of Albumin+crystalloids compared to crystalloids only for sepsis patients on 28 day mortality in the Mimic-iv database.

The corresponding paper is:

Causal thinking for decision making on Electronic Health Records: why and how, Doutreligne M. and Struja T. and Abecassis J. and Morgand C. and Celi L. and Varoquaux G., arXiv preprint arXiv:2308.01605, 2023

Step 0: data acquisition

We used a combination of duckdb and postgresql to build mimic for a laptop into parquet files easy to process during the analysis.

Steps to reproduce data acquisition on a laptop

0. We first built the data using postgresql following [mimic-code instructions](https://github.com/MIT-LCP/mimic-code/tree/main/mimic-iv/buildmimic/postgres) *=30min-1h*.
1. We then used built the concepts from the [postgresql concepts scripts](https://github.com/MIT-LCP/mimic-code/tree/main/mimic-iv/concepts_postgres) *=1h-2h*.
2. We built the original database with duckdb *=16min* following [mimic-code instructions](https://github.com/MIT-LCP/mimic-code/tree/main/mimic-iv/buildmimic/duckdb). This is done to copy the large files to parquet without memory overflow.
3. We convert all tables from the original database to parquet files using duckdb: `cli.duckdb2parquet` *=5min*.
4. We convert all derived tables from postgresql to parquet using polars: `cli.mimiciv_derived2parquet` *=2min*.

NB: Step 2 and 3 could be done from polars or duckdb directly from the csv, I think. But I had some bugs with the `pl.sink_parquet` when trying to convert the csv.

**Computing setup:** The analysis was performed on a laptop running Ubuntu 22.04.2 LTS with the following hardware: CPU 12th Gen
Intel(R) Core(TM) i7-1270P with 16 threads and 15 GB of RAM.

Step 1: study design – Frame the question to avoid biases

• Code for main paper: caumim.framing.albumin_for_sepsis.py
• Step-by-step notebook: notebooks/_1_framing_albumin_for_sepsis.py

Step 2: identification – List necessary information to answer the causal question

We built the causal graph with a clinician, using the online graphical tool Daggity.

The practical implementation of the selection of these covariates in MIMIC-IV is done with the get_event_covariates_albumin_zhou function in caumim.variables.selection.py

Step 3: Estimation – Compute the causal effect of interest

• Code for main paper: caumim.experiments.sensitivity_albumin_for_sepsis.py
• Step-by-step notebook: notebooks/_3_estimation__albumin_for_sepsis.py

Step 4: Vibration analysis – Assess the robustness of the hypotheses

• Code for main paper:

  • Vibration analysis on immortal time bias: caumim.experiments.immortal_time_bias_albumin_for_sepsis.py
  • Vibration analysis on feature aggregations: caumim.experiments.sensitivity_feature_aggregation_albumin_for_sepsis.py
  • Vibration analysis on causal and statistical estimators: caumim.experiments.sensitivity_albumin_for_sepsis.py

• Step-by-step notebook: notebooks/_3_estimation__albumin_for_sepsis.py

Step 5: Treatment heterogeneity – Compute treatment effects on subpopulations

• Code for the main paper: caumim.experiments.cate_exploration_albumin_for_sepsis.py