When researchers pool thousands of clinical trials into one graph, they must decide which differently worded outcomes count as "the same outcome." Examples are "all-cause mortality at 30 days," "death within one year," and "overall survival." This proposal argues that the decision goes wrong in a predictable place. Trial graphs, like many real networks, are probably scale-free: a few nodes (placebo, "overall survival," "nausea") connect to thousands of trials, and most nodes connect to one or two. We hypothesize that these hub nodes are exactly where incompatible definitions pile up. Hubs also sit on most paths through the graph. So a small amount of hidden heterogeneity at a handful of hubs could distort a large share of downstream answers: which trials look similar, which drug–outcome links are predicted, which indirect treatment comparisons are made. This is the clinical version of a well-known property of scale-free networks. They tolerate random damage but are fragile when hubs are hit.
The concrete proposal has two parts. First, measure whether definition spread grows with node degree in a public clinical-trial knowledge graph, and whether corrupting hubs hurts downstream tasks far more than corrupting random nodes. Second, test a targeted fix. Instead of making all outcome categories finer, split only the hubs whose members disagree, and keep a parent link so the graph stays connected. The claim is new only if this targeted refinement beats uniform refinement at the same node budget. The experiments are designed to settle exactly that.
Background and the Gap
The substrate is CTKG. It builds a knowledge graph from 8,210 ClinicalTrials.gov studies in the AACT database (the public relational copy of that registry). It normalizes adverse events with MedDRA and drugs with MeSH, and groups free-text outcome titles into 200 clusters with CLUTO, linking 56.96% of outcome records. It then embeds the graph with TransE (a method that represents each relation as a translation between entity vectors) for link prediction and study retrieval. CTKG supplies the data and the baseline pipeline. It never examines the graph's degree distribution. It also never asks whether its 200 clusters merge clinically non-equivalent endpoints, a concern its own design invites.
Two lines of work motivate the hub hypothesis. The first is the ontology study How Matchable Are Four Thousand Ontologies on the Semantic Web. It found that alignment networks among 4,433 ontologies are scale-free, and that meaning drifts along mapping chains routed through hub ontologies. We borrow its diagnosis but move it from schema terms to trial endpoints, and we propose a correction it does not offer. The second is the protein-network result Lethality and centrality in protein networks. It showed that hubs are disproportionately consequential: about 62% of proteins with more than 15 links are essential, against about 21% for proteins with 5 or fewer, with r = 0.75. Similarly, Efficient Immunization Algorithm for Peer-to-Peer Networks showed that protecting about 3% of hubs halts epidemic spread. Both findings suggest that fixing a few hubs might recover most of the damage. We test whether that transfers to definitional noise.
Several papers constrain the design rather than motivate it. Scale-Free Graph-Language Models shows that cosine nearest-neighbor graphs over text naturally have heavy-tailed degree. That means any embedding-based clustering we use may manufacture hubs of its own, so degree must be measured on the registry-derived graph, not only on similarity graphs. The file also notes that incomplete sampling can produce apparent power laws. We therefore follow the maximum-likelihood fitting with Kolmogorov-Smirnov tests used in Distinguishing manipulated stocks via trading network analysis. We also compare against a lognormal fit, and do not claim "scale-free" unless it survives that comparison. Nothing downstream depends on the exponent being exactly a power law. The method only needs heavy-tailed degree.
PALM makes multivariate network meta-analysis tractable. Network meta-analysis combines direct and indirect trial comparisons across many treatments. PALM treats the treatment network purely as a design matrix and assumes trial-level definitions are commensurate. We use its setting as a secondary test of whether hub heterogeneity distorts indirect comparisons. We do not compete with its estimator.
Core Claim
In registry-derived trial graphs, the spread of definitions among records merged into a node rises with that node's degree. Hubs therefore carry disproportionate hidden heterogeneity. Refining only the high-spread hubs should recover more downstream accuracy per added node than refining every cluster uniformly.
We define a node's definition spread $D(v)$ as the mean pairwise distance between the outcome records merged into node $v$. The distance combines three parts: the embedding distance between outcome titles, a mismatch indicator for measurement unit, and a mismatch indicator for time frame. AACT stores the latter two as separate fields. Using a mean over pairs rather than a sum keeps $D$ from growing mechanically with cluster size. We test this further with a null model that reassigns records to clusters while preserving the cluster-size profile.
Research Program
The first stage is diagnostic. We rebuild the CTKG graph from AACT and fit degree distributions for outcome, adverse-event, and intervention nodes. We then estimate how $D(v)$ scales with degree $k$ against the size-preserving null. Next we run a perturbation study modeled on attack-tolerance experiments. We inject synthetic definitional noise, by reassigning a fixed number of outcome records to wrong but similar clusters, in two ways: concentrated on the top-degree nodes, or spread over random nodes, with the same total number of corrupted records. We then measure the drop in link prediction and retrieval quality. A clearly larger drop under hub-targeted corruption shows that hub heterogeneity matters downstream, not merely that it exists.
The second stage is the method, hub-adaptive refinement. For each node whose $D(v)$ exceeds a threshold chosen on validation data, we re-cluster its members into sub-nodes using the unit and time-frame fields and title embeddings. Each sub-node keeps a "subtype-of" edge to the original parent. This design choice addresses a real risk. Removing or splitting a hub in a scale-free graph can fragment it, leaving rare trials with no path to anything. The parent edge preserves the coarse connection while letting the embedding separate incompatible uses. The total number of new nodes is capped so that every comparison below uses the same node budget.
Baselines
The naive reference point is CTKG exactly as published, with 200 CLUTO outcome clusters and TransE trained through DGL-KE. We rerun it on the same temporal split with the same embedding dimension (200) and training budget. The only downstream figure the file reports for it is that 50% of the top-10 predicted drug-repurposing pairs had supporting evidence (CTKG). No reported result was found on Lacuna for CTKG link-prediction mean reciprocal rank or Hits@10, so we will establish those numbers ourselves.
The strongest competitor is uniform refinement: re-running CLUTO with more clusters so the total node count equals the hub-adaptive graph's. This is the obvious fix a reviewer would propose. If heterogeneity is spread evenly rather than concentrated at hubs, it should do just as well. It receives the same embedding model, hyperparameter search budget, and split.
The ablation that isolates the key ingredient is degree-only splitting. It splits the top-degree nodes to the same budget but ignores $D(v)$. This separates two possibilities: that hubs simply need more capacity because they are large, or that the gain comes from locating where definitions actually disagree. A second ablation removes the parent edges, to measure whether fragmentation costs are real.
We also include a no-clustering text baseline, in which each outcome record keeps its own node and is attached to neighbors by title-embedding similarity. We use this rather than a harmonization classifier like the one in Automated Data Harmonization in Clinical Research, because that classifier's reported top-5 accuracy of 98.95% was obtained within the cohorts it trained on. Its cross-cohort area under the curve fell to 0.78–0.85, and that drop is the relevant warning for new registries.
Beating the baselines means the following. Hub-adaptive refinement must exceed uniform refinement and degree-only splitting on held-out link prediction and on clinical-equivalence precision of retrieved trials. The margins must lie outside bootstrap confidence intervals over three or more seeds. The gain must also be concentrated on queries whose paths cross refined hubs.
Why not just…?
Why not just use more clusters everywhere? This is the main threat, and the proposal is built so that it can lose to it. If $D(v)$ does not rise with degree, or if uniform refinement at a matched node budget performs as well, the hub-concentration claim is false. Our contribution would then shrink to a measurement that CTKG's granularity is too coarse. The perturbation study and the matched-budget comparison decide this directly.
Why not just skip clustering and match outcome text with a language model? Text similarity alone collapses exactly the distinctions at issue. "Mortality at 30 days" and "mortality at 12 months" are nearly identical as strings. The harmonization evidence also shows that text matchers degrade on unseen cohorts: area under the curve of 0.78–0.85 across separated cohorts in Automated Data Harmonization in Clinical Research. We run the no-clustering baseline so this argument rests on measured results rather than on assertion.
Why not just use standard meta-analytic heterogeneity statistics and leave-one-out sensitivity checks? Between-study variance estimates such as I² work within one pooled comparison, after someone has already decided which outcomes are pooled. They cannot detect that a graph node silently merged incompatible definitions across thousands of trials. Our secondary experiment shows this in practice, by checking whether indirect comparisons routed through high-spread hubs disagree more with direct evidence than routes through low-spread hubs do.
Why not just downweight high-degree nodes during embedding? Degree normalization treats every hub as suspect, including clean ones such as well-defined adverse-event terms. The degree-only ablation captures this family. If it matches the spread-driven method, heterogeneity measurement adds nothing, and we will report that.
Proposed Experiments
All models train on trials registered before a fixed cutoff year and are evaluated on later trials. This reflects the real use case and prevents leakage from near-duplicate registrations such as the paired NCT records CTKG highlights.
The first task is link prediction for study-to-outcome and drug-to-adverse-event edges, scored by mean reciprocal rank and Hits@10. Results are stratified by the degree and spread of the target node.
The second task is retrieval of similar trials. Precision is judged by clinical equivalence of the retrieved trial's primary outcome: an automated check requires matching unit and time-frame fields, and clinicians annotate a sample of several hundred pairs.
The third is the indirect-comparison check. From AACT's reported group-level results, we identify drug pairs that have direct head-to-head trials and are also connected indirectly through a shared comparator hub. We measure the disagreement between direct and indirect estimates as a function of the hub's $D(v)$, before and after refinement. Following the univariate-versus-multivariate concern raised by PALM, we restrict this to single-outcome comparisons, so estimator choice cannot confound the result.
The decisive checks come in order:
- Does spread rise with degree beyond the size-preserving null?
- Does hub-targeted corruption hurt more than random corruption at an equal number of corrupted records?
- Does spread-driven refinement beat uniform refinement and degree-only splitting at a matched node budget, with gains located on hub-crossing queries?
- Does refinement reduce disagreement between indirect and direct estimates?
A negative answer to the first check ends the hub story. A negative answer to the third means the right lesson is simply finer clusters everywhere. Positive answers to all four would show that trial-graph builders should spend curation effort on a small, identifiable set of hubs.
References
- A knowledge graph of clinical trials ($$\mathop {\mathtt {CTKG}}\limits$$)
- How Matchable Are Four Thousand Ontologies on the Semantic Web
- Lethality and centrality in protein networks
- Efficient Immunization Algorithm for Peer-to-Peer Networks
- Scale-Free Graph-Language Models
- Distinguishing manipulated stocks via trading network analysis
- PALM: Patient-centered treatment ranking via large-scale multivariate network meta-analysis
- Automated Data Harmonization in Clinical Research: Natural Language Processing Approach