This repository contains code for our MLHC 2024 paper FairEHR-CLP: Towards Fairness-Aware Clinical Predictions with Contrastive Learning in Multimodal Electronic Health Records

This project proposes a general framework for Fairness-aware Clinical Predictions with Contrastive Learning in EHRs. FairEHR-CLP operates through a two-stage process:

• Data Generation: Synthetic counterparts are created for each patient to introduce diverse demographic identities while maintaining essential health information.
• Fairness-Aware Predictions: Contrastive learning is employed to align patient representations across sensitive attributes, which are jointly optimized with an MLP classifier using a softmax layer for clinical classification tasks.

FairEHR-CLP: Our proposed method is implemented in patient_MIMICIII/MIMICIV/Stanford_exp.py. Note that synthetic notes are generated using Llama2_for_notes.py.

• Baselines:

  • Demographic-free Classification (DfC): This approach assumes that excluding demographic features, which are often central to socially sensitive biases, should lead to minimal differences in model performance. (See patient_MIMICIII/MIMICIV/Stanford_no_demo_exp.py)

  • Adversarial Debiasing (AdvDebias): A debiasing strategy tailored for EHR that simultaneously trains a classifier and an adversary model to neutralize bias. (Model link: https://github.com/yangjenny/adversarial_learning_bias_mitigation)

  • Fair Patient Model (FPM): Employs a Stacked Denoising Autoencoder and a weighted reconstruction loss to ensure equitable patient representations. (https://www.sciencedirect.com/science/article/pii/S1532046423002654)

  • RoBERTa-large: A widely-used embedding method for general applications. (Model link: https://huggingface.co/FacebookAI/roberta-large)

  • ClinicalBERT: A healthcare-specific embedding method designed for medical applications. (Model link: https://huggingface.co/medicalai/ClinicalBERT)

We evaluate our proposed framework using three EHR datasets: STAnford medicine Research data Repository (STARR) from Stanford Medicine, MIMIC-III, and MIMIC-IV. The focus is on surgical patients aged 50 or older, a cohort prone to age-related biases like impaired cognition. To avoid patient data overlap, we use the MIMIC-III Clinical Database CareVue subset. The study targets three tasks: classifying delirium, OUD, and 30-day readmission, chosen for their impact on postoperative care and patient safety. Demographic indicators are excluded from clinical notes to concentrate on health conditions. We extract patient data from a 24-hour postoperative period and use MICE imputation to address missing values. Each task is approached as a binary classification problem. The table below represents the class distribution for each task.

Please get in touch with the corresponding author for sample data inputs.

We use STARR dataset as an demonstration example (patient_Stanford_exp.py).

Step 1: Load necessary packages

import pandas as pd
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import random
import pdb
import nlpaug.augmenter.word as naw
import torch.optim as optim
from torch.utils.data import DataLoader, TensorDataset, Dataset, Subset
from transformers import BertTokenizer, BertModel, RobertaTokenizer, RobertaModel
from scipy.stats import wasserstein_distance

Step 2: Load datasets and generate synthetic counterparts

df = pd.read_csv("structured_Stanford.csv")
df_notes = pd.read_csv("unstructured_Stanford.csv")

df_demographics = df[['pat_deid', 'sex', 'ethnic_group', 'race', 'age', 'surg_family', 'product_type', 'tobacco_user', 'readmission_30_days_label']]
df_demographics = df_demographics.drop_duplicates()
df_longitudinal = df[['pat_deid', 'Heart Rate', 'Pulse', 'Resp', 'SpO2','Temp', 'Systolic_BP', 'Diastolic_BP', 'ALT (SGPT), Ser/Plas', 'Albumin, Ser/Plas', 'Anion Gap', 'BUN, Ser/Plas', 'CO2, Ser/Plas', 'Calcium, Ser/Plas', 'Chloride, Ser/Plas', 'Creatinine, Ser/Plas', 'Glucose, Ser/Plas', 'Hematocrit', 'Hemoglobin', 'MCH', 'MCHC', 'MCV', 'Platelet count', 'Potassium, Ser/Plas', 'RBC', 'RDW', 'Sodium, Ser/Plas', 'WBC']]
...
train_dataset = Subset(PatientPairDataset(real_data, synthetic_data, labels, use_synthetic=True), train_indices)
test_dataset = Subset(PatientPairDataset(real_data, synthetic_data, labels, use_synthetic=False), test_indices)
)

Step 3: Model training and evaluation

model = FairnessAwareModel().to(device)
for epoch in range(num_epochs):
    # Training Phase
    model.train()
    ...
    model.eval()
    ...
    np.save(f'{epoch_str}_task_test_ground_truth.npy', np.array(test_ground_truth))
    np.save(f'{epoch_str}_task_test_predictions.npy', np.array(test_predictions))

Step 4: Obtain the output .npy for ground truth and model predictions for further performance and fairness evaluation.

Step 5: Feature Extraction and Fairness Analysis

Please see feature_extraction_and_fairness_analysis.ipynb for how features from the STARR dataset are extracted and for fairness evaluation using EO and EDDI scores.

If you find this work helpful, please consider citing as follows:

@article{wang2024fairehr,
  title={FairEHR-CLP: Towards Fairness-Aware Clinical Predictions with Contrastive Learning in Multimodal Electronic Health Records},
  author={Wang, Yuqing and Pillai, Malvika and Zhao, Yun and Curtin, Catherine and Hernandez-Boussard, Tina},
  journal={arXiv preprint arXiv:2402.00955},
  year={2024}
}