HODC-AES: Higher-Order Drug-Combination Adverse-Event Signal Benchmark

Release version: v1.0.0
Release date: 2026-05-08
Paper status: Submitted to International Conference on Computing Advancements (ICCA)

Title

HODC-AES: A Reliability-Aware Benchmark for Higher-Order Drug-Combination Adverse-Event Signal Prediction in Computational Pharmacovigilance

Authors

Asif Elahi, Nur Hossain Mahruf, Raihan Kabir, Anowar Hossen Farvez

Affiliations

1. Asif Elahi — Dhaka International University, Bangladesh
2. Nur Hossain Mahruf — Dhaka International University, Bangladesh
3. Raihan Kabir — Department of Computer Science and Engineering, University of Asia Pacific, Bangladesh
4. Anowar Hossen Farvez — Department of Computer Science and Engineering, University of Asia Pacific, Bangladesh

Corresponding author

Asif Elahi
Email: asifelahi488@gmail.com

Overview

HODC-AES is a reliability-aware benchmark for higher-order drug-combination adverse-event signal prediction in computational pharmacovigilance.

The benchmark is built from public FAERS quarterly reports and supports evaluation under:

• higher-order drug-combination settings,
• random and temporal validation,
• contrastive negative protocols,
• calibration analysis,
• top-k alert-budget evaluation,
• temporal generalization-gap analysis.

This benchmark treats FAERS-derived labels as reported pharmacovigilance signal candidates, not as confirmed causal clinical drug-drug interactions.

Main release statistics

• Positive signal candidates: 973,898
• Generated contrastive negatives: 1,500,000
• Negative protocols: random, drug-count-matched, frequency-matched, one-drug-replacement, hybrid-hard
• Known-positive overlap among generated negatives: 0
• Negative label issues: 0
• Cross-protocol duplicate negatives: 8
• Unique drug combinations: 165,878
• Unique reaction terms: 12,311

Splits

The benchmark uses:

1. deterministic random 70/15/15 train/validation/test split,
2. temporal split:
   • training: 2014Q3 to 2021Q4
   • validation: 2022Q1 to 2022Q4
   • testing: 2023Q1 to 2024Q3

Baseline

The conference baseline uses a leakage-controlled text representation with:

• HashingVectorizer,
• linear SGD classifier with logistic loss,
• balanced positive/negative training caches,
• calibration using Platt scaling and isotonic calibration.

Main reported results

Raw model results:

Split	AUROC	AUPRC	Brier	ECE
Random	0.7104	0.5605	0.2169	0.0927
Temporal	0.6010	0.4665	0.2519	0.1229

Temporal-minus-random raw gaps:

Metric	Gap
AUROC	-0.1094
AUPRC	-0.0940
Brier	+0.0350
ECE	+0.0302

Repository and archive

• GitHub repository: https://github.com/asifelahii/HODC-AES
• Zenodo DOI: 10.5281/zenodo.20090199
• Zenodo URL: https://zenodo.org/records/20090199

Package types

GitHub-slim package

Recommended for GitHub. Contains documentation, final tables, final figures, release manifest, checksums, citation file, and license notes.

Large benchmark data files are excluded from the slim GitHub package.

Zenodo-full package

Recommended for Zenodo archival release. Contains benchmark data, generated negatives, prediction files, final tables, final figures, documentation, release manifest, checksums, citation file, and license notes.

Intended use

This benchmark is intended for research on reliability-aware computational pharmacovigilance, especially evaluation of higher-order drug-combination adverse-event signal models under temporal shift and alert-budget constraints.

Not intended use

This benchmark should not be used as direct clinical evidence of causal drug-drug interactions. FAERS is a spontaneous reporting system, and reports may be affected by under-reporting, duplicate reporting, missing information, reporting bias, and temporal changes in reporting behavior.

License

• Software/code: MIT
• Dataset/documentation: Creative Commons Attribution 4.0 International (CC BY 4.0)

See LICENSE, DATA_LICENSE.md, and DISCLAIMER.md.

Citation

Please cite the associated paper and the archived Zenodo release once the DOI is available.

@misc{hodc_aes_2026,
  title        = {HODC-AES: A Reliability-Aware Benchmark for Higher-Order Drug-Combination Adverse-Event Signal Prediction in Computational Pharmacovigilance},
  author       = {Elahi, Asif and Mahruf, Nur Hossain and Kabir, Raihan and Farvez, Anowar Hossen},
  year         = {2026},
  version      = {v1.0.0},
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.20090199},
  url          = {https://zenodo.org/records/20090199}
}