How Knowledge Graphs Improve Multimodal Clinical Risk Prediction

Why clinical risk prediction needs more than one modality

A patient's condition is distributed across time and across systems. A waveform captures rapid physiological change. A clinical note records symptoms, observations, and context. Structured records identify diagnoses, medications, procedures, and laboratory findings. Each source is useful, but each is also incomplete.

A signal-only model may detect tachycardia without knowing whether the patient has a related diagnosis. A text-only model may recognise that diagnosis while missing a deteriorating vital trend. Simple early fusion—concatenating both feature sets—can improve prediction, but it does not automatically explain how the pieces are clinically related.

This is the gap a knowledge graph can address. Instead of treating inputs as an undifferentiated feature vector, the graph represents encounters, diagnoses, symptoms, medications, laboratory markers, signal phenotypes, and evidence snippets as connected entities. Typed relations make the connections explicit and inspectable.

The MedKG-Signal architecture in five layers

1. Encode physiological and clinical context separately

The signal encoder and text encoder first produce independent representations. Keeping these pathways separate at the start allows each encoder to preserve the structure of its own data before fusion.

2. Build a heterogeneous medical knowledge graph

The graph connects patient encounters to clinical concepts through typed relations such as encounter-has-diagnosis, medication-treats-diagnosis, lab-supports-diagnosis, and signal-phenotype-indicates-condition. Edges also carry provenance, confidence, and timestamp metadata so a traversal can become an audit trail rather than a hidden similarity match.

3. Treat signal phenotypes as first-class evidence

Five clinically grounded phenotype scores are derived for each cohort using distances from relevant thresholds. Examples include tachycardia, hypotension, low Glasgow Coma Scale, hypoxia, and tachypnea in the PhysioNet cohort. These phenotypes become graph nodes rather than staying isolated numeric features.

4. Learn relation-aware graph context

A relation-aware graph encoder passes messages across typed edges. Cross-modal attention then links phenotype activation with graph structure, including neighbourhood centrality and hub-code intensity. The final patient representation combines the fused signal-and-text embedding with graph and phenotype context.

5. Constrain explanations to valid paths

For each prediction, the framework retrieves the strongest supporting graph neighbourhoods and evidence snippets. A language model may turn that evidence into a concise rationale, but a claim is retained only if it maps to a valid graph path. Unsupported statements are removed or marked as uncertain.

That last constraint changes the role of generation. The language model is not free to invent a plausible clinical story; it is limited to communicating evidence the system can trace.

A consistent five-model benchmark

The evaluation applies the same five-model contract to four complementary cohorts:

• Signal-only uses physiological or observation-derived features.
• Text-only uses clinical text or coded-record features.
• Early Fusion concatenates signal and text features.
• KG without signal tests graph context without physiological evidence.
• MedKG-Signal combines text, graph, signal, phenotypes, and cross-modal attention.

Each feature block feeds a calibrated logistic regression under a stratified 65/15/20 split with a fixed random seed. This keeps the final classifier deliberately simple, making the contribution of the representation easier to inspect.

What the results do—and do not—show

Multimodal structure helped across the real ICU cohorts

MedKG-Signal achieved the highest reported AUROC on the primary GNN for EHR cohort, PhysioNet 2012, and the eICU-CRD Demo. On PhysioNet's 1,600-patient test set, it reached 0.8309 AUROC and 0.4764 AUPRC. On the 498-stay eICU test set, it reached 0.8579 AUROC and 0.5058 AUPRC.

The APACHE IVa comparison needs careful interpretation

On the 336 eICU stays where a vendor prediction was available, APACHE IVa achieved 0.9043 AUROC. MedKG-Signal's 0.8579 AUROC across the broader eICU test set is equivalent to 94.9% of that reported score. This is not evidence that MedKG-Signal replaces APACHE IVa. The models use different variable sets, and the comparison is not based on an identical evaluation subset.

The research value is the trade-off being explored: approaching a professionally curated clinical baseline while returning graph-grounded evidence paths that a conventional severity score does not provide.

The losing result is scientifically useful

Signal-only achieved 0.9041 AUROC on the synthetic Synthea community cohort, compared with 0.7344 for MedKG-Signal. That outcome was expected because the synthetic target was driven deterministically by LOINC observations. Adding text and graph context could not create new signal that was absent from the data generation process.

This result is a useful warning against architectural maximalism: the right model is the one that matches the information structure of the task. A multimodal graph system should earn its additional complexity through measurable lift, stronger traceability, or both.

Why evidence-grounded explanations matter

A fluent explanation can be more dangerous than no explanation when its clinical claims are unsupported. MedKG-Signal addresses this by measuring evidence precision, unsupported-claim rate, path validity, and the number of evidence paths returned per prediction—not just predictive discrimination.

The experiments report 100% evidence precision, zero unsupported claims, and three grounded paths per prediction for MedKG-Signal across all four cohorts. These are pipeline-level explanation metrics, not proof that clinicians will find every path useful. Prospective clinician review is still required to establish actionability, usability, and safety.

Even with that limitation, the design principle is transferable: if a clinical AI system cannot show which observations, relationships, and source records support a claim, the claim should not appear as established fact.

From a research result to a responsible deployment

A hospital deployment would require much more than a strong retrospective benchmark. At minimum, teams should plan for:

• external validation on the intended patient population and care setting;
• calibration checks across demographic, clinical, and operational subgroups;
• explicit clinician review, override, and escalation workflows;
• immutable logging of inputs, model versions, retrieved evidence, and final decisions;
• monitoring for data drift, broken graph links, and changing clinical practice;
• privacy, access-control, retention, and data-residency controls;
• human-factors testing to determine whether explanations help or distract clinicians.

For a broader implementation framework, see Trustworthy Medical AI: Explainability, Safety, and Audit Trails. The same source-attribution patterns also apply to auditable deep-research agents in other regulated domains.

Limitations and next steps

The current study is promising but bounded. The primary code-driven cohort combines empirical code vectors with synthetic signal phenotypes and a learnable synthetic mortality outcome. The eICU analysis uses the 2,492-stay demo rather than the full credentialed database. The paper also reports an additional 50–100 milliseconds of graph-encoding latency compared with early fusion.

The next meaningful tests are full multi-site validation, waveform-matched evaluation, paediatric and non-US cohorts, subgroup analysis, and prospective clinician-in-the-loop studies. Respiratory and arterial blood-pressure waveforms would also broaden the phenotype layer, provided the ontology and validation process expand with them.

The practical takeaway

Multimodal clinical AI should not mean placing more features into a larger black box. The goal is to combine complementary evidence while preserving the chain from observation to prediction to explanation.

MedKG-Signal offers one blueprint: encode each modality, represent medical relationships explicitly, promote signal phenotypes into the graph, and block explanatory claims that cannot be grounded. Its benchmark results justify further study—not clinical use—and its boundary-case result is a reminder to prefer simpler models whenever they fit the data better.