MEDS Graph: Event-Centric Graph Models

• MEDS Graph is a family of graph-based representations that explicitly encode event and entity relationships using temporal, provenance, and interaction edges.
• It is applied in clinical informatics and beyond to improve data querying, predictive accuracy, and reasoning by capturing complex dynamics absent in flat data models.
• Recent frameworks leverage MEDS Graphs for effective medication recommendation, causal inference in LLM reasoning, and dynamic network analyses in non-medical contexts.

Searching arXiv for the cited papers and the term “MEDS Graph” to ground the article in current arXiv records. MEDS Graph is not a single formal object across the research literature. It denotes, or corresponds to, several graph-based representations whose common feature is an event- or entity-centric encoding of relations that are difficult to express in flat tables alone. In health AI, the term is most directly associated with source-faithful and event-centric graphs over electronic health records, visit–code graphs, heterogeneous patient–encounter–lab–medication graphs, and directed causal graphs used to probe LLM reasoning; outside clinical informatics, closely related abbreviations refer to time-varying communication graphs over Mobile Edge Devices and to morphing edge drawings in graph visualization (Kiiskinen et al., 21 Jun 2026, Marfoglia et al., 7 Jan 2026, Hettige et al., 2019, Mao et al., 2019, Rosenbaum et al., 2024, Nguyen et al., 2023, Misue et al., 2019).

1. Terminological scope and recurring structure

Across the cited works, a MEDS Graph typically makes entities first-class graph objects and places the analytical burden on edges rather than on implicit joins or sequence indices. Depending on the domain, nodes may represent patients, visits, encounters, codes, labs, medications, imaging events, XML fragments, diseases, or communication agents; edges may encode temporal order, containment, co-occurrence, provenance, causal relations, or neighborhood communication (Liu et al., 2018, Kiiskinen et al., 21 Jun 2026, Hettige et al., 2019, Mao et al., 2019, Rosenbaum et al., 2024, Nguyen et al., 2023).

Context	Nodes	Edge semantics
EHR event graphs	patients, visits, events, codes, XML fragments	temporal, containment, provenance
Clinical recommendation graphs	encounters, labs, medications, diseases, events	co-occurrence, interaction, conditional relation
LLM reasoning graphs	medical concepts or disease knowledge nodes	direct causality or retrieval links
MEC communication graphs	Mobile Edge Devices	undirected time-varying communication links

A recurring pattern is that the graph is introduced to recover structure that standard representations obscure. PatientEG states that existing relational data models on electronic medical records lack explicit and accurate semantic definitions of medical activities and their temporal relations, leading to inconvenience of query construction and inefficiency of query execution where multi-table join queries are frequently required (Liu et al., 2018). MedGraph makes a parallel point for representation learning: binary visit vectors with a static vocabulary of medical codes fail to encapsulate rich attribute information and ignore time gaps between visits (Hettige et al., 2019). VISTA Architect makes the same criticism of direct long-context prompting and retrieval-augmented generation, arguing that these approaches often miss temporal relationships and incur high cost and latency from repeated raw-text processing (Kiiskinen et al., 21 Jun 2026).

This suggests a family resemblance rather than a single ontology: a MEDS Graph is best understood as a graph-based representation introduced when temporal, relational, or provenance structure is itself part of the computational problem.

2. Event-centric clinical data graphs

In the most literal clinical sense, MEDS Graph refers to an event-centric graph over longitudinal health data. MEDS-OWL starts from the Medical Event Data Standard, where each row is an event tuple

$$(\texttt{subject\_id}, \texttt{code}, \texttt{[time]}, \texttt{[value]})$$

and lifts it into RDF/OWL using classes such as meds:Subject, meds:Event, meds:Code, meds:DatasetMetadata, meds:SubjectSplit, and meds:SubjectLabel, with edges such as meds:hasSubject, meds:hasCode, meds:assignedSplit, and meds:parentCode (Marfoglia et al., 7 Jan 2026). The first release of MEDS-OWL comprises 13 classes, 10 object properties, 20 data properties, and 24 OWL axioms, and the associated meds2rdf library converts MEDS events into RDF graphs validated with SHACL (Marfoglia et al., 7 Jan 2026). On the synthetic neurovasc dataset, the resulting graph contains 1,330,351 RDF triples and 245,015 meds:Event individuals, with each meds:Event having on average 5.31 triples (Marfoglia et al., 7 Jan 2026).

PatientEG provides a more explicitly temporal event graph. Built on the Simple Event Model, it defines five medical entities—peg-o:Patient, peg-o:Disease, peg-o:Drug, peg-o:Assay, and peg-o:Surgery—and five medical events—peg-o:HospitalizationEvent, peg-o:DiagnosisEvent, peg-o:DrugEvent, peg-o:AssayEvent, and peg-o:SurgeryEvent—linked by sem:hasActor and timestamp properties (Liu et al., 2018). Its temporal layer consists of peg-o:Before, peg-o:After, peg-o:Concurrent, peg-o:During, and peg-o:Overlap, and these relations are only established between events of the same patient (Liu et al., 2018). The published dataset contains 191,294 events, 3,429 distinct entities, and 545,993 temporal relations, and is exposed through a SPARQL endpoint for clinical research queries (Liu et al., 2018).

MedGraph moves from semantic representation toward learned embeddings. It treats visits and codes as nodes in an attributed bipartite graph $(V, C, E_{vc})$, where $(v_i,c_j)\in E_{vc}\iff c_j\in C_{v_i}$, and augments this structure with a temporal visit sequence $V \xrightarrow{t} V$ per patient (Hettige et al., 2019). Each node is represented by a Gaussian embedding

$\mathbf{z}_k = \mathcal{N}(\mu_k,\sigma_k^2),$

with similarity measured by the 2-Wasserstein distance and temporal dynamics modeled by a point process whose conditional intensity follows RMTPP-style parameterization (Hettige et al., 2019). MedGraph evaluates on Heart Failure and Chronic Liver Disease cohorts and outperforms state-of-the-art EMR embedding methods on readmission and mortality prediction; on CL mortality prediction, MedGraph achieves AUC 0.7415 and AP 0.7143, compared with AUC 0.7385 and AP 0.7123 for MedGraph(S, ¬T) (Hettige et al., 2019).

VISTA Architect uses the term MEDS Graph explicitly for the lowest, “source-faithful” layer of its graph database. In this design, “patients, encounters, notes, measurements, procedures, medications, imaging records, and other clinical entries become graph nodes, while temporal, visit-level, and source-document relationships become graph edges” (Kiiskinen et al., 21 Jun 2026). The graph is not intended to summarize the record, but to make the original EHR computationally addressable, with XMLFragment nodes acting as provenance anchors and SOURCED_FROM edges connecting higher-level abstractions back to source evidence (Kiiskinen et al., 21 Jun 2026). On a thoracic oncology cohort, the median graph size is 3,608 nodes per patient, provenance lookups return in under 0.01 s, and the system achieved 96.4% accuracy on 15 tumor board-salient variables across 1,180 patients (Kiiskinen et al., 21 Jun 2026).

A central distinction emerges here. In PatientEG and MEDS-OWL, the graph is the primary data model. In MedGraph and VISTA Architect, the graph is both data model and computational substrate for downstream learning or reasoning. This suggests that, in clinical informatics, MEDS Graph is best treated as an event-centric graph layer that preserves relations—especially temporal and provenance relations—that would otherwise be lost.

3. Graphs for medication recommendation, imputation, and pharmacovigilance

A second major usage of MEDS-like graph structures appears in medication recommendation and related prediction tasks. MedGCN constructs a heterogeneous graph $G=(\mathcal{V},\mathcal{E})$ with four node types—patients, encounters, lab tests, and medications—and three clinically meaningful bipartite adjacencies: $A_{E\times P}$, $A_{E\times L}$, and $A_{E\times M}$ (Mao et al., 2019). Encounter–lab edges are weighted by normalized lab values, while medication recommendation is cast as multi-label classification at encounter nodes and lab test imputation as regression on the encounter–lab matrix (Mao et al., 2019). The MedGCN propagation rule aggregates messages separately from patients, labs, medications, and encounters, and the joint objective combines binary cross-entropy for medications with masked mean squared error for labs as a form of cross regularization (Mao et al., 2019). On NMEDW, MedGCN achieves LRAP $0.7588 \pm 0.0028$, MAP@2 $(V, C, E_{vc})$0, and lab imputation MSE $(V, C, E_{vc})$1; on MIMIC-III, it achieves LRAP $(V, C, E_{vc})$2, MAP@2 $(V, C, E_{vc})$3, and MSE $(V, C, E_{vc})$4 (Mao et al., 2019).

ALGNet defines a medication-centered graph pair rather than a full patient graph. It builds an EHR co-occurrence graph $(V, C, E_{vc})$5 over medications, with weighted edges reflecting co-prescription counts, and a DDI graph $(V, C, E_{vc})$6 over the same medication vocabulary, with binary edges for known drug–drug interactions (Nguyen et al., 2023). Light graph convolution is applied separately to each adjacency, and the resulting embeddings are fused as

$(V, C, E_{vc})$7

This memory graph is then queried by an augmentation memory network together with visit-history memory (Nguyen et al., 2023). On the MIMIC-III subset used in the paper, ALGNet reaches DDI rate 0.0791, Jaccard 0.5176, F1 0.6729, and PR-AUC 0.7714, with better Jaccard, F1, and PR-AUC than the compared baselines (Nguyen et al., 2023).

BernGraph redefines the graph construction problem for binary EHRs. Each patient is represented by a binary event vector $(V, C, E_{vc})$8, and each medical event $(V, C, E_{vc})$9 is treated as a Bernoulli random variable with estimated mean $(v_i,c_j)\in E_{vc}\iff c_j\in C_{v_i}$0 (Piao et al., 2024). For patient $(v_i,c_j)\in E_{vc}\iff c_j\in C_{v_i}$1, the initial node feature for event $(v_i,c_j)\in E_{vc}\iff c_j\in C_{v_i}$2 is

$(v_i,c_j)\in E_{vc}\iff c_j\in C_{v_i}$3

so that zero entries also propagate non-zero learning signals (Piao et al., 2024). Edge weights are conditional probabilities $(v_i,c_j)\in E_{vc}\iff c_j\in C_{v_i}$4, and the resulting patient-specific event graphs are processed with E-GraphSAGE (Piao et al., 2024). On MIMIC-III, BernGraph reports Jaccard 0.5887, F1 0.8459, PRAUC 0.8442, and AUROC 0.9632, outperforming baseline methods that use secondary information (Piao et al., 2024).

The “Drug-disease Graph” for adverse drug reaction detection uses yet another heterogeneous construction. Drug nodes are ATC-coded prescriptions, disease nodes are ICD-10-coded diagnoses, homogeneous drug–drug and disease–disease edges are weighted by Gaussian kernels over Skip-gram embeddings learned from longitudinal code sequences, and heterogeneous drug–disease edges are conditional probabilities $(v_i,c_j)\in E_{vc}\iff c_j\in C_{v_i}$5 (Kwak et al., 2020). A bilinear decoder scores ADR signals using final node embeddings from GCN or GAT layers, trained against SIDER-derived labels (Kwak et al., 2020). The best reported model, $(v_i,c_j)\in E_{vc}\iff c_j\in C_{v_i}$6, reaches AUROC $(v_i,c_j)\in E_{vc}\iff c_j\in C_{v_i}$7 and AUPRC $(v_i,c_j)\in E_{vc}\iff c_j\in C_{v_i}$8, and the graph also surfaces high-scoring drug–disease pairs not present in SIDER (Kwak et al., 2020).

These systems differ in graph granularity—patient-centered, encounter-centered, medication-centered, or event-centered—but share a common computational move: graph structure is used to expose relationships that pure sequence models or flat vectors do not directly encode.

4. Graph-guided medical reasoning and multimodal retrieval

A further development is the use of MEDS-like graphs as reasoning or retrieval substrates for language and vision-LLMs. MedG-KRP externalizes an LLM’s biomedical reasoning as a directed graph

$(v_i,c_j)\in E_{vc}\iff c_j\in C_{v_i}$9

where each node is a medical concept and each directed edge $V \xrightarrow{t} V$0 means “$V \xrightarrow{t} V$1 directly causes $V \xrightarrow{t} V$2” (Rosenbaum et al., 2024). The graph is constructed in two stages: recursive node expansion to identify concepts that cause or are caused by a root medical concept, and exhaustive edge refinement over all ordered node pairs (Rosenbaum et al., 2024). Human reviewers then score the resulting graphs for accuracy and comprehensiveness, while a BIOS-based comparison computes precision and recall using short paths in an external biomedical knowledge graph (Rosenbaum et al., 2024). Across 60 generated graphs, GPT-4 has the best human-reviewed mean accuracy and comprehensiveness, at 3.37 and 3.23, whereas PalmyraMed has the best BIOS-aligned precision and recall, approximately 0.243 and 0.033 (Rosenbaum et al., 2024). The divergence between human review and KG alignment is a direct result reported by the paper.

Med-GRIM uses Graph-RAG rather than causal graph extraction. Its DermaGraph dataset organizes 50 dermatological conditions as disease nodes with child nodes for symptoms, treatment, and preventive measures, plus inter-disease edges based on textual similarity (Madavan et al., 20 Jul 2025). Each disease node stores multimodal and textual embeddings derived from BIND, a BLIVA-based encoder extended with a True Transformation Layer and sub-query projections to obtain a denser joint embedding space (Madavan et al., 20 Jul 2025). Retrieval uses a hybrid similarity score

$V \xrightarrow{t} V$3

followed by graph expansion to neighbors and a second-stage response filter that asks symptom questions and retains conditions with $V \xrightarrow{t} V$4 (Madavan et al., 20 Jul 2025). On DermaGraph QA, Med-GRIM with Graph-RAG reaches 83.33% accuracy and 0.81 semantic-BERT, exceeding Med-GRIM with plain RAG at 78.7% and 0.81, and prompt engineering raises BIND-based Med-GRIM from 80.7% to 83.33% (Madavan et al., 20 Jul 2025).

These two papers instantiate complementary roles for graphs in model-centric medical AI. In MedG-KRP, the graph is the output of reasoning and the object of evaluation. In Med-GRIM, the graph is an external knowledge substrate used to constrain reasoning at inference time. A plausible implication is that MEDS Graph can function either as a representation of medical knowledge or as an interface for auditing model-generated knowledge.

5. Communication and mobility graphs under the MEDS abbreviation

Outside health data modeling, the abbreviation MEDS also appears in networking and mobility. In CrowdCache, MEDs are Mobile Edge Devices, and the relevant graph object is the sequence of undirected time-varying communication graphs

$V \xrightarrow{t} V$5

over the MED set $V \xrightarrow{t} V$6 (Nguyen et al., 2023). The neighbor set is

$V \xrightarrow{t} V$7

so each node has a self-loop (Nguyen et al., 2023). The graph sequence is assumed to be $V \xrightarrow{t} V$8-connected, Metropolis weights produce symmetric doubly stochastic mixing matrices $V \xrightarrow{t} V$9, and decentralized gradient play updates local estimates and actions using only neighbor communication (Nguyen et al., 2023). The paper’s main theorem states that, for a sufficiently small constant step size, Algorithm 1 converges linearly to the unique Nash equilibrium, and the authors emphasize that the game itself does not depend on the communication graph; the graph only constrains the information structure (Nguyen et al., 2023). In simulation, the MED graph is a proximity-based, time-varying geometric graph built from campus coordinates, with $\mathbf{z}_k = \mathcal{N}(\mu_k,\sigma_k^2),$0 in the base case and scalability tests up to $\mathbf{z}_k = \mathcal{N}(\mu_k,\sigma_k^2),$1 (Nguyen et al., 2023).

A related but distinct non-clinical meaning appears in route optimization for electric vehicles. There, MEDs are Mobile Energy Disseminators, and the relevant graph is the road network $\mathbf{z}_k = \mathcal{N}(\mu_k,\sigma_k^2),$2 augmented with MED rendezvous points $\mathbf{z}_k = \mathcal{N}(\mu_k,\sigma_k^2),$3, static charging stations $\mathbf{z}_k = \mathcal{N}(\mu_k,\sigma_k^2),$4, and binary edge labels $\mathbf{z}_k = \mathcal{N}(\mu_k,\sigma_k^2),$5 indicating when an EV follows a MED and receives dynamic wireless charging (Kosmanos et al., 2017). The objective minimizes total travel time including driving, waiting, and charging, while constraints enforce path structure and state-of-charge feasibility (Kosmanos et al., 2017). The paper reports that the MED-enabled system is approximately 2×, 3×, and 4× faster than the SCS-only system at the three demand levels considered (Kosmanos et al., 2017).

These papers make clear that “MEDS Graph” cannot be assumed to be medical. In CrowdCache, it is a dynamic D2D communication topology. In EV routing, it is an augmented transportation graph with mobile charging nodes. The overlap is purely terminological.

6. Boundary cases, misconceptions, and disambiguation

Several misconceptions arise because identical or near-identical abbreviations are used for different graph constructs. First, in CrowdCache the MED graph is not a payoff graph or dependency graph: each MED’s utility depends on all actions through a global price function, but the graph only specifies who can exchange information with whom during decentralized computation (Nguyen et al., 2023). Second, in VISTA Architect the MEDS Graph is explicitly not a summary layer; it is source-faithful and exists so that higher-level abstractions can always be traced back to exact notes, labs, and reports (Kiiskinen et al., 21 Jun 2026). Third, in BernGraph the graph is not temporal in the visit-sequence sense; it is built from cohort-level Bernoulli means and conditional probabilities over binary event outcomes, with patient specificity encoded in node features rather than explicit visit nodes (Piao et al., 2024).

A further source of ambiguity is that MED in graph visualization means Morphing Edge Drawing, not a medical graph. In that literature, a graph $\mathbf{z}_k = \mathcal{N}(\mu_k,\sigma_k^2),$6 is drawn so that each edge $\mathbf{z}_k = \mathcal{N}(\mu_k,\sigma_k^2),$7 is represented by a time-varying morphing function $\mathbf{z}_k = \mathcal{N}(\mu_k,\sigma_k^2),$8, derived from a partial edge drawing function $\mathbf{z}_k = \mathcal{N}(\mu_k,\sigma_k^2),$9 and a ratio function $G=(\mathcal{V},\mathcal{E})$0 (Misue et al., 2019). The resulting 1/4-SHMED reduces the reading-time penalty of static 1/4-SHPED for adjacency checks: the difference between 1/4-SHPED and 1/4-SHMED is statistically significant with $G=(\mathcal{V},\mathcal{E})$1, although no significant accuracy difference is found (Misue et al., 2019). This is a graph drawing technique, not a health data representation.

The literature therefore supports a narrow editorial rule: MEDS Graph should always be disambiguated by domain and schema. In clinical informatics, it most often denotes an event-centric, provenance-aware, or relation-aware graph over longitudinal health data. In recommender systems and pharmacovigilance, it denotes heterogeneous graphs over encounters, medications, labs, diseases, or events. In LLM evaluation and Graph-RAG, it denotes causal or retrieval graphs over medical concepts. In networking and visualization, the same letters refer to unrelated graph objects with entirely different semantics.