mmid

A Python package for Multi-Modal Integration and Downstream analyses for healthcare analytics

Official implementation of the paper "mmid: Multi-Modal Integration and Downstream analyses for healthcare analytics in Python" (arXiv 2026)

• mmid
  • Getting started
    • Requirements
    • Installation
    • Run integration and downstream analysis
  • Data format
    • Configuration files
    • Input datasets
    • Optional inputs
  • Examples
  • Citation
  • License

Getting started

Requirements

Python 3.7

Installation

# Create a conda environment (recommended)
conda create -n mmid python=3.7
conda activate mmid

# Install mmid
pip install git+https://github.com/ht-diva/mmid

Run integration and downstream analysis

mmid runs in Python with a single function call. The datasets to analyze and the selected integration and downstream algorithms are passed to the function through configuration files.

Here is an example:

# Import mmid
from mmid.mmid import mmid

# Run analysis
mmid(config_data="/path/to/data_config/data_config.yml", 
    config_model="/path/to/model_config/model_config.yaml", 
    cohort_path="/path/to/cohort/", 
    cohort_file="cohort.csv", 
    end_study_date="2025-12-31", 
    out_path="/path/to/output/")

The function has the following parameters:

Parameter name	Format	Default	Description	Example
config_data	Path		Fully-qualified path of the yaml data configuration file	"/path/to/data_config/data_config.yml"
config_model	Path		Fully-qualified path of the yaml model configuration file	"/path/to/model_config/model_config.yaml"
cohort_path	Path		Path of the cohort_file's directory	"/path/to/cohort/"
cohort_file	Filename		Cohort file name (csv); if inside a directory, the directory name will be interpreted as the analyzed disease	"disease/cohort.csv"
end_study_date	yyyy-mm-dd		Last observation date (i.e., end of the study)	"2025-12-31"
out_path	Path		Path where to store the results	"/path/to/output/"
cohort_cov	String	""	Comma-separated covariates of cohort_file considered for the downstream analysis: blank for not considering covariates from cohort_file for the downstream analysis; "Age_at_baseline" for considering age	"CAT_Sex,Age_at_baseline"
baseline_field	String	"Assessment_date"	Field for initial observation date of the study: single (as default) or multiple (comma-separated); in the latter case, optimal_baseline must be set to True to keep the more recent baseline	"Assessment_date,Second_assessment_date"
optimal_baseline	Boolean	False	Whether to use the optimal (i.e., more recent) subject-dependent initial observation date; it must be set to True when multiple baselines are specified in baseline_field	True
latent_impute	Boolean	False	True to solve missingness in the latent space, if possible	False
years_risk_classification	Integer	5	N for N-years risk classification (ignored if the downstream task is not classification)	5
feature_selection_classification	Boolean	False	True to use Sequential Forward Feature Selection based on logistic regression to select features for classification (ignored if the dowsntream task is not classification)	False
withdrawals	Path	None	Fully-qualified path of the withdrawals file (csv)	"/path/to/withdrawals/withdrawals.csv"
genetic_kinship	Path	None	Fully-qualified path of the genetic kinship file (csv); None for ignoring genetic kinship in the analyses	"/path/to/genetic_kinship/genetic_kinship.csv"
genetic_kinship_exclude	Boolean	False	True to exclude subjects with some genetic kinship from cross-validation and test, according to the genetic_kinship file; if genetic_kinship is None, then genetic_kinship_exclude must be set to False	True
scaling	Boolean	True	Whether to scale features or not	True
n_folds	Integer	5	Number of folds for k-fold cross-validation downstream model assessment	5
test_size	Float	0.2	Proportion of the dataset to include in the held-out test split for downstream analyses	0.2
analysis_tag	String	"test"	Tag to identify the run/analysis	"test"
log_path	Path	None	Fully-qualified path of the log file to save in a file in out_path ("" or None to ignore this operation)	""
dpi	Integer	300	Figure quality (dpi)	300
seed	Integer	42	Random state (seed)	42

Data format

Configuration files

mmid relies on two YAML configuration files:

• A data configuration file, where the user inserts details on the modality datasets (features to consider, number of factors in the integration, ...)
  • Data configuration example
• A model configuration file, where the user selects the integration and downstream algorithms (and their hyperparameters) for the analysis
  • Model configuration example

Input datasets

mmid expects the following main input datasets:

• One or more modality datasets (csv), each one containing features from a specific data source
  • One row per subject
  • One column per feature (feature name must not contain the substring _target)
    • Categorical feature name must start with the prefix CAT_ and be in a number format (e.g., 0/1 allowed, A/B/C not allowed)
  • uid feature (mandatory) for unique integer subject identifiers
• A cohort dataset (csv) describing the cohort for downstream analyses, with the following structure (feature name must not contain the substring _target and must contain the prefix CAT_ when categorical):

Feature	Format	Description	Example
uid	Integer	Unique subject identifier	123456
DOB	yyyy-mm	Date of birth	1998-06
Assessment_date (feature name can be changed)	yyyy-mm-dd	Initial observation date	2026-03-20
CAT_baselineEvent	Binary	1 if the subject experienced a baseline event (e.g., a cardiovascular disease) that implies exclusion from downstream analyses if happened before baseline, 0 otherwise	1
baselineEvent_date	yyyy-mm-dd	Date of the baseline event (blank when CAT_baselineEvent is 0)	2026-01-01
CAT_endpointEvent	Binary	1 if the subject experienced the target event (e.g., coronary artery disease)	1
endpointEvent_date	yyyy-mm-dd	Date of the target event (blank when CAT_endpointEvent is 0)	2026-03-27
CAT_Exit	Binary	1 if the subject early exited the observation study, 0 if observed until the end of the study	0
Exit_date	yyyy-mm-dd	Date when the subject exited the observation study (blank when CAT_Exit is 0)
CAT_Death	Binary	1 if the subject died, 0 if alive	0
Death_date	yyyy-mm-dd	Date of subject's death (blank when CAT_Death is 0)
CAT_Sex (optional)	Binary	1 for males, 0 for females	0

Optional inputs

Users can optionally specify:

• Subjects who withdrew from the study and should thus be excluded from analyses (text file with one unique integer subject identifier per line, and no header)
• Genetic kinships between study participants (csv), with the following structure:

Feature	Format	Description	Example
uid	Integer	Unique subject identifier	123456
Genetic_kinship_to_other_participants	Binary	1 if subject has genetic kinship with at least another study participant, 0 otherwise	1

Examples

example.py and example.ipynb show how to use mmid. To run these example files, mmid has to be installed. Additionally, notebook execution requires ipykernel.

example.py can be executed with:

# Activate existing conda environment with mmid already installed (recommended)
conda activate mmid

# Navigate to the examples directory
cd examples/

# Run the example script
python example.py

The example scripts rely on some (simulated mockup) input datasets and configuration files to run the analyses, and save results in the output directory.

Citation

If you use mmid in your research, please cite:

@misc{vergani2026mmidmultimodalintegrationdownstream,
      title={mmid: Multi-Modal Integration and Downstream analyses for healthcare analytics in Python}, 
      author={Andrea Mario Vergani and Valeria Iapaolo and Emanuele Di Angelantonio and Marco Masseroli and Francesca Ieva},
      year={2026},
      eprint={2604.08334},
      archivePrefix={arXiv},
      primaryClass={stat.CO},
      url={https://arxiv.org/abs/2604.08334}, 
}

License

This project is licensed under the MIT License. See LICENSE for details.