MohsinElis/biomedical-disease-mapping

Representation versus Geometry: A Controlled Factorial Benchmark for Unsupervised Biomedical Disease Knowledge Mapping

Official repository and open-science artifacts for the submission to Arabian Journal for Science and Engineering (AJSE) (Springer Nature).

Overview

This repository hosts the canonical multi-disease PubMed corpus (14,064 records), factorial benchmark matrices, inferential statistical test outputs, information-theoretic hierarchy metrics, blinded clinical adjudication schemas, and the publication-ready IEEEtran manuscript.

Key Scientific Findings

1. The Lexical Parity Paradox: Sparse TF-IDF achieves topic coherence (NPMI 0.1158, 116 topics) statistically indistinguishable from 768-d domain-adapted BioBERT embeddings (NPMI 0.1154, 80 topics) under HDBSCAN density clustering (Mann-Whitney U = 4836.5, p = 0.6156, bootstrap 95% CIs [0.0965, 0.1345] vs [0.0931, 0.1378]).
2. Selective Coverage Trade-Off: HDBSCAN achieves high separation (Silhouette 0.4786) by setting aside 30.5% (TF-IDF) and 42.4% (BioBERT) as noise. When KMeans is evaluated under matched cardinalities ($K=80, 116$), its NPMI coherence rises to 0.1166 and 0.1133 with 100.0% document coverage.
3. Information-Theoretic Structural Hierarchy: HDBSCAN micro-clusters partition five broad clinical macro-categories with 97.4% purity and 96.0% information gain ($H(\text{Macro}|\text{Micro}) = 0.092$ bits vs 2.317 bits base entropy), while a 5.1% cross-cutting fringe quantitatively captures biological multimorbidities (gut-brain axis, systemic inflammation).
4. Clinical Concept Grounding: 91.4% of extracted keywords match verified MeSH descriptors, supported by end-to-end case traces and a formalized blinded adjudication protocol.

Master Benchmark Summary

Representation	Clustering Method	Topics	Coverage	Silhouette (2D)	DB Index	NPMI Coherence	ARI (NCBI)	NMI (NCBI)
TF-IDF	HDBSCAN	116	69.5%	0.4786	0.6034	0.1158	0.0184	0.1658
BioBERT	HDBSCAN	80	57.6%	0.4225	0.7068	0.1154	0.0139	0.1429
BioBERT	KMeans ($K=5$)	5	100.0%	0.4112	0.8124	0.0823	0.0376	0.0811
TF-IDF	KMeans ($K=5$)	5	100.0%	0.3708	0.8426	0.0780	0.0625	0.1196
TF-IDF	KMeans ($K=80$ Matched)	80	100.0%	0.4113	0.7256	0.1166	0.0165	0.1459
TF-IDF	KMeans ($K=116$ Matched)	116	100.0%	0.4267	0.6976	0.1133	0.0127	0.1476
Bag-of-Words	LDA ($K=5$)	5	100.0%	---	---	0.1518	0.0541	0.0832

Repository Organization

• q1_research_upgrade/: Complete upgraded research deliverables:
  • paper/: Upgraded IEEEtran LaTeX source (main.tex) and compiled 8-page PDF (main.pdf).
  • reports/: Complete audit reports (Stages 0 through 10, claim audits, issue registers).
  • metrics/: Master benchmark matrix, hierarchy validation, topic diversity.
  • statistics/: Inferential tests (Mann-Whitney U, t-test, bootstrap CIs).
  • annotations/: Blinded clinical expert protocol, rating schema, case study traces.
  • data_manifests/: Canonical PMID manifests, attrition flows, dedup audits.
  • reproducibility/: Reproducibility run manifest, scripts, and guide.
• pubmed_dataset.csv: Canonical 14,064-abstract corpus.
• TF-IDF_Results/: 11 CSVs and 11 PNGs from the sparse frequency pipeline.
• BioBERT_Results/: 11 CSVs and 11 PNGs from the BioBERT transformer pipeline.
• Coherence_Results/: NPMI topic coherence evaluations and distribution plots.
• Manuscript_Results/: Macro-level clinical categories and Micro-level top-10 topics.