Fairness-Aware Interpretable Modelling (FAIM) for trustworthy machine learning in healthcare

FAIM Introduction

As artificial intelligence (AI) plays an increasingly vital role in high-stakes domains like healthcare, concerns about fairness and bias have grown. In healthcare, biased AI models can unfairly impact critical decisions, such as disease diagnosis and organ allocation, leading to health disparities. Achieving fairness and mitigating biases in AI models is challenging and critical in healthcare decision-making.

Existing bias mitigation methods operate at different stages of the modeling process, but many struggle with biases intersecting multiple protected attributes, and some methods compromise model effectiveness. Additionally, interpretability is a challenge, particularly in black-box AI models, making it difficult for clinicians to participate actively in the modeling process.

To address these challenges, we introduce the Fairness-Aware Interpretable Modeling (FAIM) framework. FAIM operates within a cloud of nearly-optimal models, ensuring fairness by leveraging varying degrees of model reliance on specific variables. This framework enhances interpretability and facilitates clinician involvement, achieving contextualized AI fairness in healthcare.

Installation

Execute the following command in Terminal/Command Prompt to install the Python library from GitHub:

• Linux/macOS:

pip install git+"https://github.com/nliulab/FAIM.git"

• Windows:

python.exe -m pip install git+"https://github.com/nliulab/FAIM.git"

Usage

FAIM consists of 3 general steps:

1. Nearly-optimal model generation
2. Fairness transmission
3. SHAP-based model explanation followed by the comparison with other bias-mitigation methods

Demo: FAIM for prediction hospital admission based on MIMIC-IV-ED data

The dataset can be extracted from MIMIC-IV-ED benchmark. With data splitted into dat_train, dat_test and dat_expl (for fairness tranmission), we can run the following code to generate the nearly-optimal model, transmit fairness, and explain the model.

Initial data processing (starting from MIMIC-IV-ED benchmark)

esi_dic = {1:"High risk", 2:"High risk", 3:"@Low risk", 4:"@Low risk", 5:"@Low risk"}
dat_train["triage_acuity"] = dat_train["triage_acuity"].map(esi_dic)
dat_expl["triage_acuity"] = dat_expl["triage_acuity"].map(esi_dic)
dat_test["triage_acuity"] = dat_test["triage_acuity"].map(esi_dic)

race_dic = {0:"Asian", 1:"Black", 2:"Hispanic", 3:"Others", 4:"@White"}
dat_train["race"] = dat_train["race"].map(race_dic)
dat_expl["race"] = dat_expl["race"].map(race_dic)
dat_test["race"] = dat_test["race"].map(race_dic)

gender_dic = {0:"Female", 1:"@Male"}
dat_train["gender"] = dat_train["gender"].map(gender_dic)
dat_expl["gender"] = dat_expl["gender"].map(gender_dic)
dat_test["gender"] = dat_test["gender"].map(gender_dic)

var_dict = {"Age": "Age", "gender": "Gender", "race":"Race", "triage_acuity": "ESI", "triage_o2sat": "SPO2", "triage_temperature":"Temperature", "n_hosp_365d":"Hospitalizations last year", "triage_pain": "Pain scale", "triage_heartrate":"Heartrate", "triage_resprate": "Respirate rate", "triage_dbp": "Diastolic blood pressure", "triage_sbp":"Systolic blood pressure"}

Initial setting up:

y_name = 'label'
colnames = ['Age', 'ESI', 'Systolic blood pressure', 'Heartrate', 'Diastolic blood pressure', 'Temperature', 'Pain scale', 'SPO2', 'Respirate rate', 'Hospitalizations last year', 'Gender', 'Race']
x_names_cat = ["ESI", "Gender", "Race"]
sen = ["Gender", "Race"]
sen_ref = {"Gender":"@Male", "Race":"@White"} 

output_dir = "output"
if not os.path.exists(output_dir):
    os.makedirs(output_dir)

Step 1: Nearly-optimal model generation

faim_obj = FAIMGenerator(
  dat_train, 
  selected_vars=colnames, 
  selected_vars_cat=x_names_cat, 
  y_name="label", 
  sen_name=sen, 
  sen_var_ref = sen_ref, 
  criterion="auc", m=800, n_final=200, output_dir=output_dir, without_sen="auto", pre=False)

Step 2: Fairness transmission

faim_obj.FAIM_model(dat_expl)

best_results, fair_idx_df = faim_obj.transmit(
  targeted_metrics = ["Equalized Odds", "Equal Opportunity", "BER Equality"]
  ) 
print(best_results['best_sen_exclusion'])
# print(best_results['best_coef'])

"Gender" and "Race" were excluded!

pred_test, fairmetrics_faim, fairsummary_faim = faim_obj.test(dat_test)

Step 3: Shap-based model explanation

shap_compare = faim_obj.compare_explain(overide=True)
shap_compare

Comparison with other bias-mitigation methods

fairbase = FairBase(
  dat_train, 
  selected_vars=colnames, 
  selected_vars_cat=x_names_cat, 
  y_name="label", 
  sen_name=sen, 
  sen_var_ref=sen_ref, 
  weighted=True, 
  weights={"tnr": 0.5, "tpr": 0.5}
  )

• Original LR

lr_results = fairbase.model(method="OriginalLR")
pred_ori, fairmetrics_ori, clametrics_ori = fairbase.test(dat_test, model=lr_results)
fairmetrics_ori

• Reweigh

_, rw_results, _ = fairbase.model(method_type="pre", method="Reweigh")
pred_rw, fairmetrics_rw, clametrics_rw = fairbase.test(dat_test, model=rw_results)
fairmetrics_rw

summary of results:

faircompare_df = pd.concat([fairmetrics_ori, fairmetrics_faim, fairmetrics_rw])

Metric	Equal Opportunity	Equalized Odds	Statistical Parity	Accuracy Equality	BER Equality
Original LR	0.316338	0.316338	0.39563	0.049166	0.301258
Reweigh	0.206211	0.206211	0.207493	0.047577	0.18257
FAIM	0.133691	0.146606	0.241103	0.04704	0.140149

clacompare_df = pd.concat([clametrics_ori, clametrics_faim, clametrics_rw])

	auc_low	auc	auc_high
Original	0.787	0.790	0.793
Reweigh	0.786	0.792	0.792
FAIM	0.783	0.786	0.789

Contact

• Mingxuan Liu (Email: [email protected])
• Nan Liu (Email: [email protected])