crossgat-trial-eligibility-cv

CrossGAT Trial Eligibility — 5-fold CV checkpoints

A 3-class eligibility model that reranks clinical-trial retrieval candidates. It scores a patient description against a trial's inclusion/exclusion criteria as irrelevant / excluded / eligible, by fusing a graph encoder over parsed eligibility criteria with a text branch over whole-document embeddings.

These are the five cross-validation checkpoints — one per fold, so each is trained on 80% of the data and never saw its own fold's topics. Use them to reproduce the reported numbers, or ensemble them. For a single deployment model, see crossgat-trial-eligibility-full.

⚠️ This model is not self-contained

It does not take raw text. It takes parsed eligibility graphs — entity and relation structures extracted from trial criteria and patient descriptions by an LLM. Producing those needs the generator from the project repo plus an LLM endpoint (~38 GB VRAM).

If you just want to reproduce the results, the graphs for all 13,229 TREC CT trials are published as a dataset — no LLM needed: clinical-trials-eligibility-graphs-rerank.

What it is for, and what it is not for

It reranks. It does not retrieve. Trained on an eligibility target, it reorders within the candidate set a first-stage retriever already found — promoting eligible trials over merely topically-relevant ones.

It cannot judge disease relevance. Per-topic AUC for relevant-vs-rest is 0.533 — chance. Whether a trial is even on the right disease is entirely the first stage's job. This is why it must be fused with retrieval, never used to replace it.

Recommended configuration

score = (1 - w) * retrieval_norm + w * p_eligible_norm      with w = 0.5

Both terms min-max normalised within a topic. p_eligible = P(eligible) alone wins over the alternatives in all four runs, at every cutoff, in both metric families — even though expected_gain models the graded gain directly and is what training optimises.

Results

5-fold cross-validation, TREC Clinical Trials 2021–23, 162 topics, paired Wilcoxon signed-rank against each run's own first stage. Metric is graded NDCG@10 (gain 2^label - 1 = 0/1/3).

First stages: hybrid BM25+embeddinggemma-300m-medical (retrieval 0.5973) and dense-only MedEmbed-base-v0.1 (retrieval 0.5187).

*** p<0.001, ** p<0.01, * p<0.05.

Read both metric families. raw scores unjudged documents as 0 (punishing a system for surfacing trials TREC never pooled); condensed deletes them first (flattering a system that surfaces many). Neither is honest alone — the gap between them, read against judged@10, says how far apart they can be. Where the two agree the result is robust; at w=1.0 they diverge sharply and raw loses significance entirely.

Which checkpoint to load

Each fold ships two: ndcg.ckpt (selected on val_ndcg10) and f1.ckpt (selected on val_macro_f1).

Every number above uses ndcg. The f1 checkpoints classify better (macro F1 0.597 ± 0.020 vs 0.561 ± 0.035 on held-out folds) but have not been benchmarked for reranking — offered for completeness, not recommended blindly.

Note val_ndcg10 is flat across epochs 1–5, so the per-fold selected epochs (1/2/5/2/4) are largely noise around a plateau. Do not read them as convergence.

Usage

pip install "crossgat @ git+https://github.com/JDev2001/msc_v2"

from crossgat_hf import load, score, rerank_score

pipe = load("2001jdev/crossgat-trial-eligibility-cv", fold=0, metric="ndcg")
out  = score(pipe, patient_graph, inc_graph, exc_graph,
             patient_text="65yo male, type 2 diabetes, HbA1c 8.1% ...",
             trial_text="Metformin in adults with type 2 diabetes ...")

out["class_probs"]          # {'irrelevant': .., 'excluded': .., 'eligible': ..}
rerank_score(out)           # p_eligible — fuse this with your retrieval score

patient_text and trial_text are required: the text branch is part of the model, and omitting them raises rather than silently scoring an all-zero branch.

Training

• Data: TREC Clinical Trials 2021–23 qrels/topics; eligibility criteria parsed into graphs (CHIA-style entities/relations)
• Objective: cross-entropy + LambdaRank (k=10) on topic-grouped batches (4 topics/batch)
• Hard negatives: 40/topic drawn from the retrieval candidate distribution (+5,307 samples), mined from hybrid gemma-med run files
• Text branch: abhinand/MedEmbed-base-v0.1, whole-document embeddings of patient description and trial title+summary
• Splits: 5-fold CV over patients, seed fixed; each fold's test topics never seen in its training

Limitations

• The first stage is the ceiling. medembed with reranking still scores below gemma-med without it. A better retriever buys more than more reranker tuning.
• Do not use w=1.0. Pure model ranking loses significance on both retrievers (p=0.65 / 0.85). It ranks better inside the judged pool but drifts out of it — judged@10 falls from 0.92 to 0.80.
• The hard-negative gain is retriever-sensitive. +0.0175 (p=0.007) on gemma-med, +0.0148 (p=0.050) on medembed — the negatives were mined from gemma-med candidates and transfer only partly.
• Within-collection evidence only. 162 topics, one corpus, one set of assessors. Nothing about generalisation to another trial collection follows.
• Not a clinical device. Research artefact. It does not determine patient eligibility and must not be used to make enrolment decisions.

Citation & provenance

Trained on TREC Clinical Trials track data (2021–23). Users must comply with the TREC data usage agreements for the underlying collection and relevance judgments.

Checkpoint SHA-256 prefixes and the exact training configuration are in checkpoint_manifest.json. Every benchmark run in the source repository writes a manifest.json recording checkpoint hashes, git state and sanity checks.