A common way to fix a clinical model that underperforms for small patient groups, such as elderly Hispanic women in an ICU dataset, is to generate synthetic patient records for those groups and add them to training. Because generators can produce nonsense, these pipelines keep only the synthetic records that "look real." They decide this with statistical filters trained on the real data. Those filters learn what typical patients look like, and typical patients come mostly from the majority groups. So the filtering step may quietly throw away the very minority-group records the pipeline set out to create. It may also let through records that are statistically typical but medically impossible.
We propose replacing or supplementing density-based filtering with a plausibility check built from medical knowledge. The check uses a knowledge graph, a network of facts of the form (subject, relation, object) such as (drug X, contraindicated in, condition Y), extracted automatically from biomedical literature. A synthetic record is rejected when it breaks a stated medical constraint. It is not rejected merely for being rare. The central measurement is how much each stage of a debiasing pipeline adds to or removes from subgroup disparity: generation, filtering, training, and thresholding.
Core Claim
In augmentation pipelines aimed at intersectional bias (groups defined by combinations of attributes, such as age × race × gender), the quality filter is itself a source of bias. Discriminator- and density-based filters reject valid records from rare subgroups at higher rates than records from common subgroups. A knowledge-graph constraint check cuts this false rejection while catching medically impossible records that density filters accept. Together these should raise worst-subgroup predictive performance on real held-out patients more than current filtering does.
The closest work is MedEqualizer. It finds under-covered intersectional subgroups in MIMIC-III and generates targeted records with MedGAN, CTGAN and HealthGAN. It then filters those records with a One-Class SVM and a logistic-regression discriminator. Its authors explicitly flag the risk of hallucinating medical patterns for rare profiles, and they use no knowledge to check this. It also never tests whether its own filters are subgroup-neutral. We make both of those the object of study.
Research Program
The work has three parts.
Filter-induced bias. The first part measures filter-induced bias directly. We have a clean test for this: pass real held-out MIMIC records through each filter. Every real record is valid by definition, so any rejection of a real record is a false rejection. The subgroup-wise false-rejection rate therefore isolates the filter's bias from generator quality. We expect the rate to rise as subgroup frequency falls.
The knowledge-graph checker. The second part builds the checker. Following the generative triple extraction of DALK, an LLM reads PubMed abstracts retrieved for the diagnosis codes, drugs and lab concepts present in the MIMIC feature set. DALK built an Alzheimer's graph from 9,764 papers this way and found that a compact, high-precision graph beat noisy exhaustive extraction. We keep only relation types that can be turned into checkable constraints on a record:
- sex-specific conditions;
- age-restricted conditions;
- drug contraindications;
- drug-implies-indication links;
- lab-range consistency with a diagnosis.
Precision matters more than recall here, because a false constraint would wrongly reject valid minority records. Extraction hallucination and non-standard relations are known failure modes (Schema-adaptable Knowledge Graph Construction). We therefore audit a sample of extracted constraints against a curated biomedical graph and clinician review. We also measure false rejection of real records under the knowledge-graph check itself, which should be near zero.
Stage-wise attribution. The third part places the checker inside a full pipeline and tracks disparity through each stage. The stages are:
- targeted generation;
- filtering;
- training a mortality predictor, with or without a group-robust objective;
- optional per-group thresholding.
The taxonomy in A Comprehensive Taxonomy of Bias Mitigation Methods for Hate Speech Detection observes that late and intermediate stages are under-studied. Our stage-wise attribution addresses a filtering stage that existing taxonomies do not name as a bias source.
Two constraints shape the evaluation. First, MIMIC-IF shows that MIMIC-IV subgroups differ in baseline mortality, and that predictive AUC is inversely related to baseline mortality. For that reason we score worst-subgroup AUROC and calibration, not parity of predicted mortality rates, so that genuine physiological differences are not penalized. Second, the synthetic-data literature warns that LLM judges are lenient (Towards Understanding Bias in Synthetic Data for Evaluation reports a +0.28 Bland-Altman leniency bias for GPT-4 relevance labels). We therefore do not let an LLM grade plausibility in the final evaluation.
Baselines
No augmentation (naive reference). A mortality predictor trained on real MIMIC-III training data only. It uses the same architecture, and the same hyperparameter search budget, as every arm. It fixes the floor for worst-subgroup AUROC. As a sanity reference for the predictor class on the sister dataset, MIMIC-IF reports IMV-LSTM at AUROC 0.955 and AUPRC 0.769 for in-hospital mortality on MIMIC-IV.
MedEqualizer as published (strongest existing method). We reimplement MedEqualizer with its three generators, its uncovered-pattern threshold, and its One-Class SVM plus discriminator filters. Its reported results are subgroup coverage counts: equitably represented subgroups rose from 25 to 47 with HealthGAN, 31 to 44 with MedGAN, and 22 to 31 with CTGAN. No reported result was found on Lacuna for MedEqualizer's downstream per-subgroup mortality AUROC, so we will establish that number ourselves.
For fairness, all arms get the same synthetic record budget before filtering and the same generator checkpoints. The downstream predictor and tuning budget are identical across arms, and all arms share the same real test split with intersectional subgroups held out by patient. The proposal beats MedEqualizer if it meets three conditions. It must raise worst-subgroup AUROC with a confidence interval (from patient-level bootstrap) excluding zero. It must not lower overall AUROC by more than a pre-registered small margin. And it must reduce clinician-audited implausibility among accepted records.
Subgroup-conditional density filter. We fit the One-Class SVM and discriminator separately per subgroup, or use per-subgroup acceptance thresholds. This is the cheapest fix for filter-induced bias.
Curated-graph checker. The same constraint check built from a curated biomedical graph, iBKH as used by CLINGEN, instead of automatically extracted triples.
Knowledge-only versus knowledge-plus-density ablation. This isolates whether the knowledge graph should replace the statistical filter or run alongside it.
Reweighting without augmentation. An in-processing group-robust objective on real data alone. It tests whether synthesis is needed at all.
Why not just…?
Why not just make the density filter subgroup-conditional? This is the most serious threat, and it may fully remove false rejection of rare-group records. Its weakness is statistical. The subgroups we care about are, by construction, those below a frequency threshold. A density model fit on a few dozen patients is poorly estimated, so it will be either too permissive or too strict. A permissive model also has no way to reject a record that is rare and impossible, such as an obstetric code on a male record. The false-rejection test settles the first question, and the controlled impossible-record injection below settles the second. If the conditional filter matches the knowledge checker on both, we will report that the knowledge graph adds nothing beyond plausibility auditing.
Why not just use a curated knowledge graph or hand-written clinical rules? This is also a real threat. If iBKH-derived constraints perform as well, then automated construction is unnecessary for MIMIC. Automated extraction earns its place only if it covers constraints for the long tail of codes and drugs that curated graphs miss. We measure constraint coverage over MIMIC's code vocabulary for both graphs and compare downstream results head to head. A tie is a legitimate and useful finding.
Why not just skip synthesis and reweight real patients? Reweighting cannot add information about subgroups with almost no records. Synthetic pre-training helps most when labels are scarce: UCTR-ST reports 12.1 to 56.9 F1 with 50 labels. Even so, the reweighting arm is run directly. If it matches augmentation on worst-subgroup AUROC, the whole augmentation line is unnecessary for this task.
Why not ask an LLM whether each record is plausible? LLM graders are lenient and biased toward their own outputs (Towards Understanding Bias in Synthetic Data for Evaluation). An LLM grader may also share the same majority-group priors as the density filter. We include it as an extra filter arm and run the same real-record false-rejection test on it.
Proposed Experiments
Filter bias. On MIMIC-III, and replicated on MIMIC-IV, we compute false-rejection rates of real held-out records for each filter, by intersectional subgroup, and regress these rates on subgroup frequency. The hypothesis fails if the density filters show no frequency dependence.
Controlled impossibility injection. We corrupt real records with known-impossible combinations: sex-specific codes on the wrong sex, age-restricted diagnoses, and contraindicated drug-condition pairs. Each filter's recall on these records is measured separately for common and rare subgroups.
Downstream evaluation. Each full pipeline arm trains the mortality predictor. We report overall AUROC and AUPRC, worst-subgroup AUROC, per-subgroup calibration error, and the coverage counts MedEqualizer uses, with disparity attributed stage by stage.
Clinician audit. A blinded clinician audit of accepted synthetic records from each arm gives the implausibility rate.
Decisive outcome. The claim stands if two things hold together. The knowledge checker must show lower rare-subgroup false rejection than the published filters and higher impossible-record recall than the subgroup-conditional filter. And that advantage must carry through to worst-subgroup AUROC on real patients. If only the first holds, the contribution narrows to a diagnosis of filter bias with a plausibility tool. If neither holds, the density filters were not the bottleneck.
References
- MedEqualizer: A Framework Investigating Bias in Synthetic Medical Data and Mitigation via Augmentation
- DALK: Dynamic Co-Augmentation of LLMs and KG to answer Alzheimer's Disease Questions with Scientific Literature
- Schema-adaptable Knowledge Graph Construction
- A Comprehensive Taxonomy of Bias Mitigation Methods for Hate Speech Detection
- MIMIC-IF: Interpretability and Fairness Evaluation of Deep Learning Models on MIMIC-IV Dataset
- Towards Understanding Bias in Synthetic Data for Evaluation
- Knowledge-Infused Prompting: Assessing and Advancing Clinical Text Data Generation with Large Language Models
- Toward a Unified Framework for Unsupervised Complex Tabular Reasoning