JMIR Publications

ABSTRACT

Background:

AI-driven clinical systems can improve diagnosis, prognosis, and resource allocation, but they may reproduce disparities encoded in historical healthcare data. Existing mitigation methods typically target a single source of bias, while clinical datasets often contain interacting representation, proxy, integrity, and temporal biases.

Objective:

This study develops and evaluates a unified multi-stage framework for detecting and mitigating multiple forms of bias in structured healthcare machine learning data.

Methods:

We designed a compositional pipeline, D_clean = T_temp(T_int(T_proxy(T_repr(D)))), in which each stage conditions on the corrected output of the previous stage. To address cross-dataset heterogeneity, all datasets were first mapped into a prespecified harmonized clinical-concept space with explicit missing-concept masks. The five anchor features used for alignment were race, sex, age_group, income_proxy, and n_prior_visits. The final harmonized representation contained 37 clinical concepts plus 37 corresponding binary mask indicators, yielding a 74-dimensional model input after categorical expansion and mask concatenation. The primary model was trained on the Diabetes 130-US Hospitals dataset. External validation used CMS SynPUF for readmission prediction and NHANES for stage-level fairness and distributional stress testing rather than unsupported direct outcome transfer. Integrity bias was assessed with distributional tests appropriate to each variable type; Benford-style leading-digit analysis was restricted to unbounded count or charge-like variables and was not applied to bounded physiological laboratory values such as HbA1c.

Results:

On the primary Diabetes 130-US Hospitals test split, the proposed pipeline improved AUC from 0.798 to 0.812 and reduced Demographic Parity Difference (DPD) from 0.134 to 0.052. The DPD reduction was statistically significant (bootstrap 95% CI -0.094 to -0.069; paired permutation P < .001). On CMS SynPUF after harmonized concept mapping, DPD decreased from 0.141 to 0.066. NHANES stage-level validation showed improved representation balance and proxy attenuation, while HbA1c integrity checks were evaluated with bounded-variable distributional baselines rather than Benford's Law. Mixture-of-experts processing isolated 7,938 of 101,766 records (7.8%) flagged for integrity concerns and improved fairness without discarding records

Conclusions:

A coordinated data-centric pipeline can improve both fairness and predictive performance when dataset heterogeneity, variable-specific integrity assumptions, and subgroup-specific processing are made explicit. The revised framework resolves the methodological risk of unsupported zero-shot transfer by introducing a harmonized concept layer and reporting outcome validation only where the prediction task and feature space are aligned