MixEHR-Nest

MixEHR-Nest: Identifying Subphenotypes within Electronic Health Records through Hierarchical Guided-Topic Modeling

Overview

(a) Initialization

For each patient d, MixEHR-Nest initializes the Phenotype topic prior for subtopic m within Phenotype topic k (α_{d,k_m}) by aligning the patient's ICD9-Code with Phenotype topics using an ICD-Phecode mapping.

• Seed Topics: The topic prior for subtopics associated with seed topics is set to 0.9.
• Regular Subtopics: The topic prior for regular subtopics is stochastically sampled from a range of 0.001 to 0.01.

(b) Integration and Feature Generation

MixEHR-Nest integrates information from multiple modalities within EHR data to generate multi-view EHR features. These features are then decomposed into Phenotype topic distributions for each modality t, as well as patient-topic mixture membership.

(c) Probabilistic Graphical Model

In the probabilistic graphical model (PGM) representing the data generative process for patient d and EHR feature j:

• The patient-topic mixture θ_d is drawn from a K×M-dimensional Dirichlet distribution with hyperparameter α_d.
• The Phenotype topic z_{d,j} is sampled from a Categorical distribution with probabilities determined by θ_d.
• Given the topic assignment z_{d,j} = k_m, the EHR feature X_{d,j} is then sampled from another Categorical distribution with probabilities specified by φ_{.,z{d, j}}, where φ_{.,z{d, j}} is a V-dimensional Dirichlet variable with a flat hyperparameter β.

Model Running Steps

This repository contains three steps to run the model using the provided Python scripts.

Steps

Step 0: Initialization

This step initializes the parameters for the model.

Usage

python step0_initialization.py --subtopic_num SUBTOPIC_NUMBER --output_dir STORE_DIR --metadata_folder METADATA_FOLDER

• --subtopic_num: Number of subtopics per phenotype (default: 3)
• --output_dir: Directory to store parameters (default: 'MixEHR_param')
• --metadata_folder: Metadata folder name (required)

Step 1: Train Model

This step trains the model using the initialized parameters.

Usage

python step1_trainModel.py --subtopic_num SUBTOPIC_NUMBER --param_dir PARAM_DIR --out_dir OUT_DIR --eta ETA --iteration ITERATION

• --subtopic_num: Number of subtopics per phenotype (default: 3)
• --param_dir: Path to directory storing initialized parameters (default: 'MixEHR_param')
• --out_dir: Path to output directory (default: 'MixEHR_out')
• --eta: The value to downscale the contribution of non-ICD modality in training (default: 0.6)
• --iteration: Training iterations (default: 20)

Step 2: Infer New Patient

This step uses the trained model to infer new patient data.

Usage

python step2_inferNewPat.py --subtopic_num SUBTOPIC_NUMBER --param_dir PARAM_DIR --train_out_dir TRAIN_OUT_DIR --out_dir OUT_DIR --iteration ITERATION

• --subtopic_num: Number of subtopics per phenotype (default: 3)
• --param_dir: Path to directory storing initialized parameters (default: 'MixEHR_param')
• --train_out_dir: Path to directory storing training results (default: 'MixEHR_out')
• --out_dir: Path to directory storing new patients infer results (default: 'MixEHR_infer')
• --iteration: Training iterations (default: 20)

Example

Here is an example of how to run the steps sequentially:

# Step 0: Initialization
python step0_initialization.py --subtopic_num 3 --output_dir MixEHR_param --metadata_folder Admit

# Step 1: Train Model
python step1_trainModel.py --subtopic_num 3 --param_dir MixEHR_param --out_dir MixEHR_out --eta 0.6 --iteration 20

# Step 2: Infer New Patient
python step2_inferNewPat.py --subtopic_num 3 --param_dir MixEHR_param --train_out_dir MixEHR_out --out_dir MixEHR_infer --iteration 20