An integrated multimodal temporal-graph framework for rare pharmaceutical supply-chain disruption early warning with explainable AI and permissioned governance
Rare but high-impact disruptions in pharmaceutical supply chains can propagate across multi-echelon networks, threatening medicine availability and operational continuity. Early warning is challenging because positive examples are scarce, network dependencies evolve over time, relevant evidence is multimodal, and regulated multi-partner settings require auditable review. This study evaluates an integrated node-day framework for predicting whether a qualifying disruption condition will occur within a 14-day future horizon. A Temporal Graph Network (TGN) representation is fused with operational, Internet of Things (IoT)/cold-chain, de-identified textual, and news/social signals. Partner-as-task meta-learning addresses cross-partner heterogeneity, while P2 further incorporates quality-controlled conditional variational autoencoder (cVAE) augmentation restricted to the Train partition. Post-hoc feature-contribution diagnostics, fidelity-qualified local surrogates, and joint temporal graph-context occlusion provide non-causal evidence for human review. A permissioned Hyperledger Fabric proof of concept records provenance and integrity metadata for protected off-chain artifacts. Evaluation used 36 months of confidential real-world data from a 16-partner network under a leakage-controlled chronological protocol with partition-boundary target embargoes, Train-only fitting, and real-only validation and test partitions. Across five matched seeds, the complete P2 model achieved a mean Test PR-AUC of 0.848, Recall@40 of 0.903, and Cost 10 of 3269.4, where Cost 10 = 10FN + FP. Relative to the TGN-only baseline, P2 increased PR-AUC by 0.101 and reduced Cost 10 by 31.3%. In a prespecified ten-run paired analysis, the PR-AUC advantage of P2 over the TGN-only baseline remained significant after Holm correction (adjusted p = 0.0078). At the final P2 configuration, the matched P1–P2 comparison increased mean Test PR-AUC by 0.02070 and reduced Cost 10 by 408.0 penalty units, while Recall@40 changed only modestly (+ 0.00223). Diagnostic, ablation, and sensitivity analyses were consistent with complementary predictive value across the evaluated cumulative configurations, without isolating causal effects of individual components. The findings support an association-based early-warning and auditable review system for the studied network; prospective external validation is required before broader deployment.
Authors
- maral mirza mohammadi (ORCID: https://orcid.org/0009-0009-9933-1128)
- Athena Omidian (ORCID: https://orcid.org/0009-0002-0284-6272)
- Asiye Omidian (ORCID: https://orcid.org/0009-0008-0855-8320)
- Hanie Sadat Ghodrati (ORCID: https://orcid.org/0009-0007-6945-1984)
- Zohreh Keivan (ORCID: https://orcid.org/0009-0000-2438-0659)
- Ladan Mohebian (ORCID: https://orcid.org/0009-0008-7539-5277)
- Ali Bayat (ORCID: https://orcid.org/0009-0004-9813-5356)
Institutions
- New Mexico State University (US)
- The Graduate Center, CUNY (US)
- University of Nevada, Reno (US)
- University of Cincinnati (US)
- Iran University of Science and Technology (IR)
- University of Cincinnati Medical Center (US)
- McMaster University (CA)
Publication Details
- Journal
- Discover Artificial Intelligence
- Published
- 2026-11-17
- DOI
- https://doi.org/10.1007/s44163-026-02606-0
- Primary Topic
- Supply Chain Resilience and Risk Management
- Type
- article
- Field-Weighted Citation Impact
- 0.00