Clinical prediction models for rare ICU conditions have too few patients to train on. The usual remedy is to borrow patients from related, more common conditions and use adversarial domain adaptation to make the borrowed data look like the rare target. Adversarial domain adaptation trains a feature encoder to fool a second network, the discriminator, that tries to tell source patients from target patients. The open question is what "related" should mean. The strongest current system builds its map of which conditions resemble each other from hospital statistics: diagnosis co-occurrence, similar lab distributions, and shared drugs. It reports that this data-driven map beats the ICD-9 diagnosis hierarchy. Yet for the rarest conditions, those statistics come from a few dozen patients and are likely to be noisy. The biomedical Linked Open Data ecosystem offers a denser alternative that does not depend on sample size: diagnosis hierarchies, drug targets, and indications, all published as machine-readable knowledge graphs.
We propose to build the "domain graph", the graph whose nodes are conditions and whose edges say which conditions should share representations, from linked biomedical ontologies fused with empirical edges. This graph then tells the adversarial discriminator which conditions to pull together and which to keep apart. The central, falsifiable claim is that ontology edges and statistical edges are complements, not rivals, and that their value depends on prevalence. Statistical edges should win for moderately rare conditions, as the prior work found. Ontology edges should win below a prevalence threshold, where co-occurrence estimates collapse. A graph-structured discriminator that uses both should beat either alone across the whole prevalence range.
Where this sits relative to existing work
Two papers define the space. KnowRare builds a condition knowledge graph from MIMIC-III and eICU statistics, embeds it, picks the $k$ most similar source conditions by cosine similarity, and runs a standard binary adversarial discriminator between target and pooled sources. It reports gains of up to 17.0% AUPRC over the best domain-adaptation baseline. It motivates our work in two ways. It shows that graph-guided source choice matters; its "data-volume paradox" finds that pooling everything hurts. It also supplies the tension we test: its claim that empirical similarity beats ICD-9 hierarchies. It never uses the graph inside the adversarial game. Once sources are chosen, alignment is uniform.
Graph-Relational Domain Adaptation (GRDA) supplies the mechanism. It replaces the binary discriminator with one that tries to rebuild the domain graph from encoded samples. This forces the encoder to align neighboring domains strongly and distant ones weakly. On the TPT-48 climate task it shows that uniform DANN and ADDA do worse than no adaptation. GRDA assumes a single homogeneous, hand-given, noise-free graph and was never tested on semantic or clinical data.
Our proposal differs from KnowRare by putting the graph inside the discriminator and by drawing edges from linked ontologies. It differs from GRDA by building a multi-relational graph from Linked Open Data plus statistics, and by studying how edge reliability interacts with domain size.
The linked-data side supplies the plumbing. PDD Graph linked MIMIC-III records to the ICD-9 ontology and DrugBank, reaching 99.9% diagnosis linkage and 83.4% drug linkage. Bio2RDF-based drug indication work shows that SPARQL queries (queries over RDF knowledge graphs) can pull DrugBank, SIDER, and MedDRA relations into features. That same paper also constrains the design. It shows that evaluation on linked data leaks badly unless splits are disjoint by entity, so all our tuning uses condition-disjoint splits.
Hunter's review argues that ontologies supply structure that sparse statistics cannot. We treat that argument as a hypothesis to measure, not a premise.
Two neighboring lines are not our method:
- DAEA applies domain adaptation to entity alignment between knowledge graphs.
- AdaGCN adapts node classifiers across citation networks.
In both, the knowledge graph is the object being adapted. In ours, the knowledge graph describes the relations among domains.
Core Claim and Mechanism
Each ICU condition is a domain. We build relation-specific adjacency matrices $A_r$ over conditions from three linked sources:
- ICD-9 hierarchy: sibling and ancestor distance in the ICD-9 ontology.
- Shared drug targets: overlap in DrugBank targets of the drugs prescribed for each condition.
- Shared treatments: overlap in Bio2RDF indication links.
We add the empirical KnowRare-style matrix $A_{\text{emp}}$. The fused domain graph is
$$A = w_{\text{emp}} A_{\text{emp}} + \sum_r w_r A_r ,$$
where the nonnegative weights $w$ are chosen on a proxy task, described below. Following GRDA, the encoder $E$ and outcome predictor $F$ minimize prediction loss minus $\lambda$ times the loss of a discriminator $D$. $D$ receives the encodings of patients from two conditions $i, j$ and predicts $A_{ij}$. The encoder therefore removes condition identity in proportion to how strongly the graph links two conditions.
We also test a prevalence-aware variant, where each condition's empirical edges are down-weighted by a shrinkage factor that grows as its patient count falls. This is the formal version of our crossover hypothesis. The shrinkage form follows from the variance of a co-occurrence estimate, which scales as $1/n$ in the number of patients $n$.
The weights cannot be tuned on rare-target labels without leakage. We therefore use a leave-one-condition-out proxy. We treat common conditions as pseudo-rare by subsampling their patients to rare-condition sizes, and tune on those, keeping the true rare conditions untouched.
Research Program
The work proceeds in four steps.
Build the graph. We build the linked condition graph over the MIMIC-III and eICU diagnosis vocabularies, reusing PDD Graph-style linking. We report linkage coverage per relation, because unlinked conditions become isolated nodes and must be handled explicitly, via fallback to empirical edges only.
Reproduce the reference points. We reproduce KnowRare's pipeline and protocol on the same five tasks: 90-day mortality, 30-day readmission, ICU mortality, length of stay, and phenotyping. We also reproduce GRDA on TPT-48 as a check that our discriminator implementation behaves as published.
Swap the domain graph. We vary only the domain graph (empirical only, ICD-9 only, full linked, fused, fused with shrinkage). The encoder, the time-series pretraining, and the discriminator architecture stay fixed.
Analyze by prevalence. We plot performance gain against target-condition prevalence and test for the predicted crossover.
Baselines
The naive reference point is a source-only model: KnowRare's pretrained encoder with a predictor fine-tuned on the rare condition's few labeled patients, with no adversarial term. A second naive point is pooled uniform DANN over all conditions. We expect it to underperform source-only on some tasks, given the TPT-48 negative-transfer result in GRDA and KnowRare's data-volume paradox. If it does not, that weakens the premise that alignment structure matters.
The strongest existing method is KnowRare itself. The file gives only a relative figure: "up to 17.0% AUPRC improvement over the best domain adaptation method" on MIMIC-III/eICU. We found no reported absolute AUPRC on Lacuna for KnowRare. We will therefore rerun it ourselves with the same splits, encoder, tuning budget (equal hyperparameter search trials), and compute, and report absolute numbers.
We run two direct rivals from GRDA, using its graph discriminator with (a) the plain ICD-9 tree and (b) KnowRare's empirical graph. This separates "structured discriminator" from "linked-data graph".
Two ablations isolate our key ingredients:
- KnowRare plus ontology selection: KnowRare with its cosine source selection replaced or combined with ontology distance, still with a binary discriminator. This isolates whether the graph must live inside the discriminator.
- Fused graph without shrinkage: this isolates the prevalence-aware weighting.
What counts as beating KnowRare. We require a statistically reliable AUPRC gain, using paired bootstrap across rare conditions, on at least three of five tasks. The gain must be concentrated in the lowest-prevalence tertile, as the claim predicts. A uniform small gain across all tertiles would support the fused graph but not our mechanism.
Why not just…?
Why not just use the ontology for source selection, as KnowRare does with statistics, and keep the binary discriminator? This is the most serious threat, because it is much simpler. KnowRare's own finding against ICD-9 suggests that ontology-based selection alone may not help. It is still possible that combining the two selection signals captures most of the gain. The ablation above tests this directly. If it matches the graph-discriminator versions within noise, we will report that the value of Linked Open Data lies in selection, not in structured alignment, and drop the discriminator claim.
Why not just hand GRDA the ICD-9 tree? GRDA assumes a clean, given graph. The ICD-9 tree encodes billing taxonomy, not physiological similarity. KnowRare's result suggests it is a poor sole anchor for common conditions. We run exactly this baseline. Our hypothesis predicts it wins only at the lowest prevalence and loses elsewhere, which is the pattern the fused graph is meant to fix.
Why not just align with MMD instead of an adversarial game? MMD (maximum mean discrepancy) is a closed-form distance between two feature distributions. Evidence is mixed. AdaGCN reports relative gains of 4.41% Micro-F1 and 5.81% Macro-F1 over the MMD-based CDNE on citation graphs. GraphAlign gets a 2.16% average gain with MMD-based data distillation. Neither result was on clinical time series. We therefore include a graph-weighted MMD variant, which penalizes MMD between condition pairs in proportion to $A_{ij}$, with the same graph. This also answers the instability objection. If it matches the adversarial version, we prefer it and say so.
Why not just pool more data or pretrain larger? KnowRare already pretrains a condition-agnostic encoder and reports that indiscriminate pooling hurts rare targets. We include pooled training as a baseline, but do not expect it to be competitive.
Decisive Checks
Three results decide the claim:
- Prevalence crossover. Relative to empirical-only GRDA, ontology-only GRDA gains should increase monotonically as prevalence falls, and the sign should flip. Absence of this crossover falsifies the mechanism, whatever the average numbers show.
- Discriminator placement. The fused graph-discriminator model must beat KnowRare with combined ontology selection. Otherwise the structured discriminator is unnecessary.
- Cross-institution transfer. Training graphs and weights on MIMIC-III and evaluating rare conditions in eICU tests whether ontology edges, which do not depend on the hospital, transfer better than hospital-specific co-occurrence edges. Ontology edges should degrade less.
All splits are disjoint by condition, following Çelebi et al.
References
- Bridging data gaps of rare conditions in ICU: a multi-disease adaptation approach for clinical prediction
- Graph-Relational Domain Adaptation
- PDD Graph: Bridging Electronic Medical Records and Biomedical Knowledge Graphs via Entity Linking
- Machine Learning Based Drug Indication Prediction Using Linked Open Data
- Knowledge-based biomedical Data Science
- DAEA: Enhancing Entity Alignment in Real-World Knowledge Graphs Through Multi-Source Domain Adaptation
- Graph Transfer Learning via Adversarial Domain Adaptation With Graph Convolution
- Can Modifying Data Address Graph Domain Adaptation?