Machine-learning research

eICU-KG

First-author clinical knowledge-graph research on leakage-audited, long-tail medication-concept ranking.

My roleLed benchmark construction, leakage audits, and model evaluation with Dr. Lihui Liu at Wayne State.

Accepted for poster presentation, AIDaR Workshop at NeurIPS 2026

Interactive walkthrough

Inspect what the model can see.

A ranking result is only useful when the held-out answer stays out of its input. Compare a clean split with an answer-link leak.

Inspect the evaluation boundary

Select a stage to inspect its role in the workflow.

Step 4: Evaluate. Use the answer for evaluation only.

Graph context + candidate spaceAnswer held out

Step 04 / Inspect

Use the answer for evaluation only.

The known held-out label is used to assess the ranking, not supplied as an answer link in model input. CSR-HARD additionally uses known HARD membership and remains an evaluation-only case study.

evaluation_label: Concept B (synthetic)
valid_input_boundary: true
new_results_computed: false

Audit passes in this illustration: the answer edge is excluded, while background graph context remains.

Synthetic 147-concept graph. Held-out answers are evaluation-only; no model scores are computed.

About the source dataset

Results

147

Candidate concepts

1,886

HARD test queries

+27.5%

Relative HARD Hits@10 gain

Reported manuscript results. HARD gain: evaluation-only case study with known HARD membership.

Diagnosis-frequency prior: 61.7% Hits@10 overall and 2.4% on HARD, with shared-rank tie handling.
CSR-HARD: mean top-10 hits increased from 46 to 58.7 of 1,886 HARD queries across three gate initialisations.

Engineering details

Frequent medication concepts can dominate aggregate ranking scores. eICU-KG measures performance on the long tail with audited answer-link exclusion.

Define the ranking task

Transductive recovery of held-out DRUG_OF_CONCEPT edges over 147 concepts, with patient-disjoint target splits.

Audit the benchmark

Deterministic preprocessing and held-out answer-link checks support evaluation on the full test set and the long-tail HARD slice.

Evaluate graph reranking

CSR-HARD selects 7% of HARD queries for relational graph reranking in a label-dependent, evaluation-only case study.

More work