Drug-disease Graph: Predicting Adverse Drug Reaction Signals via Graph Neural Network with Clinical Data

• Drug and disease embeddings learned with Skip-gram and categorical code embeddings jointly form node features before GNN-based ADR prediction.The graph combines same-type relationships based on embedding distances with drug-disease relationships based on conditional probabilities.Source: Figure 1 · p. 2, §1 · p. 3
• Post-marketing ADR signals matter because spontaneous reporting systems suffer from underreporting, selective reporting, and missing drug-usage data.Clinical databases such as EHR and healthcare claims data have therefore become important alternative sources.Source: §1 · p. 2
• Graph structure is motivated by polypharmacy and multiple drug-disease relations that may be obscured by disconnected representations.The model predicts whether drug and disease nodes have an ADR relation from learned node representations.Source: §1 · pp. 2–3
• The study constructs a heterogeneous Drug-disease graph from 12 years of NHIS-NSC healthcare claims covering one million people.Drug and disease nodes represent prescription and diagnosis codes derived from the claims data.Source: §1 · pp. 2–3
• AUROC and AUPRC performance improved over alternative machine-learning algorithms, while the model also identified ADR candidates absent from established databases.The approach uses clinical claims data and was designed to supplement existing ADR resources.Source: §1 · p. 3

• Pre-marketing studies use chemical structures and biological properties, including graph-based biomedical information, to predict side effects of drug candidates.Source: §2 · p. 3
• Post-marketing ADR detection traditionally relies on voluntary reports and disproportionality analysis, while newer studies use EHR and claims data with standard machine-learning methods.Source: §2 · p. 3
• Existing clinical-database research had not explored graph neural networks for predicting post-marketing ADR reactions.Source: §2 · p. 4

3.1 Problem Formulation

• The task predicts whether a drug-disease pair represents a potential ADR signal using labels for known side-effect relationships.Source: §3.1 · p. 4

3.2 Code Embedding Learning with Skip-gram Model

• Drug and disease code sequences from patient records are embedded with Skip-gram to capture temporal proximity between codes.The method uses separate prescription and diagnosis sequences and a context window of 16 codes in each direction.Source: §3.2 · p. 5

3.3 Drug-disease Graph Construction

• The heterogeneous graph combines drug and disease nodes with homogeneous and heterogeneous edges, using proximity-based and category-based node representations.Initial node vectors concatenate Skip-gram embeddings with multi-hot representations from ATC and ICD-10 code hierarchies.Source: §3.3 · pp. 5–6, Definition 3 (Node Representations) · p. 6

Definition 3 (Node Representations)

• Homogeneous edge weights encode temporal proximity through embedding distances, whereas drug-disease edge weights use conditional prescription probabilities.Edge-forming thresholds are subsequently used to control graph sparsity.Source: Definition 3 (Node Representations) · pp. 6–7

3.4 A GNN-based Method for Learning Graph Structure

• The GNN aggregates weighted neighbor features through nonlinear layers, using either fixed degree-normalized weights or learned attention weights.The fixed weighting network is called GCN, while the attention-based network is called GAT.Source: §3.4 · p. 7
• A bilinear layer uses final drug and disease embeddings to predict ADR labels, with training minimizing cross-entropy loss.Source: §3.4 · p. 8

4.1 Data preprocessing

• The dataset contains 607 drugs, 556 diseases, and 28,746 positive drug-side effect pairs, with an equal number of randomly sampled negative pairs.Positive labels come from SIDER, while graph nodes include all drugs and diseases from NHIS-NSC.Source: §4.1 · p. 8, §4.2 · p. 8

4.2 Experimental Settings

• Disease classes are kept disjoint across training, validation, and test sets, using 80%, 10%, and 10% splits to reduce leakage and make inference more difficult.Disease class is defined by the first three ICD-10 digits; low- and high-threshold graphs are also constructed to examine graph sparsity.Source: §4.2 · pp. 8–9

4.3 Evaluations

• 0.746 average test accuracy is achieved by adrGCNhigh for infrequent drug-ADR pairs, indicating usefulness for detecting rare symptoms.SIDER defines rare or post-marketing ADRs as reported with frequencies under 0.01.Source: §4.3 · p. 10
• Graph-based approaches surpass all non-graph-based approaches, with the best AUROC achieved by GCN using the low edge-forming threshold.The evaluation compares logistic regression, a fully connected neural network, vanilla GCN, GAT, and edge-type-specific adrGCN variants across low- and high-threshold graphs.Source: §4.3 · pp. 9–10

4.4 Newly-Described ADR Candidates

• High-probability pairs labeled negative by SIDER are filtered against baseline neural-network predictions, and clinical experts confirm that some are genuine ADRs.The method selects predictions above 0.97 and reports the resulting pairs in Table 3.Source: §4.4 · p. 10, Table 3 · p. 11
• Several discovered pairs involve symptoms rather than diseases, including cardiac murmur, abnormal reflex, edema, and tetany, with drug-specific clinical rationales supporting their ADR relevance.Examples connect dasatinib with cardiac murmurs, hydroxycarbamide with abnormal reflex, and bisphosphonates such as alendronic acid with tetany.Source: §4.4 · p. 10

5 Conclusion

• The proposed graph-based approach constructs a graph from large-scale healthcare claims data to capture relations among drugs and diseases.The authors describe this as their central approach for ADR detection.Source: §5 · p. 11
• The model improves drug-ADR prediction and identifies pairs absent from the established ADR database, with clinical experts validating the graph-based findings.The study currently makes inferences within the labeled dataset and plans to extend discovery to unlabeled data.Source: §5 · p. 11

A Healthcare claims dataset

• NHIS-NSC contains longitudinal diagnoses, prescriptions, and procedures for one million sampled patients observed from 2002 to 2013.The database is established by South Korea’s national health insurance system.Source: A Healthcare claims dataset · p. 13
• Statistics computed across drug-disease combinations in NHIS-NSC are used to determine edges in the drug-disease graph.The claims data therefore supplies the relation information used in graph construction.Source: A Healthcare claims dataset · p. 13
• Drugs and diseases are encoded as ATC and ICD-10 codes, yielding 1,201 medication codes and 1,872 disease codes.These code representations provide the graph’s clinical entities.Source: A Healthcare claims dataset · p. 13

B Adverse drug reaction dataset

• SIDER contains 139,756 drug-side effect pairs covering 1,430 drugs and 5,868 side effects, extracted from public ADR information using natural language processing.The source identifiers are converted from STITCH compounds and UMLS concepts into ATC and ICD-10 codes.Source: B Adverse drug reaction dataset · p. 13
• The converted SIDER dataset contains 562 ATC drug codes and 931 ICD-10 side-effect codes with hierarchical categorical structure.These codes align the labeled ADR data with the clinical graph representation.Source: B Adverse drug reaction dataset · p. 13