EpiNet — the Epistemic Network toolkit

EpiNet is Vahtian’s research-methods demonstrator for graph-shaped biomedical data. It produces calibrated, caveated, provenance-rich analysis reports. It is not clinical decision support. CiteVahti checks claim-evidence support in manuscripts. EpiNet checks whether graph-shaped analytical claims survive basic methodological stress tests. Both share the same brand principles: human review, provenance, caveats, and no unsupported claims.

EpiNet is a transparent node/edge network and feature-space analysis toolkit for graph-shaped datasets. The name reads Epistemic Network: the core question is not just "what does the model predict" but how well-founded each call is — how contestable, how calibrated, how well it transports. You load entities and relationships from CSVs and it computes graph features, honestly evaluates an outcome model, finds shortest paths, clusters nodes by feature-space centroid, and scores how contestable each call is — with publication-quality figures and a model card. Epidemiology is one use case; the core is domain-neutral (driven through lung-cancer quality indicators, lung-nodule risk, and lymphoma subtyping).

Scope. This is a research and education demonstrator, not clinical or public-health decision support. Any model it produces must be validated on independent, outcome-linked data before it means anything clinically. See Scope and caveats.

What distinguishes EpiNet from a thin scikit-learn wrapper is that honest evaluation is the default path: calibration and uncertainty are reported with the headline metric, while label-permutation nulls and community-aware splitting are built into the same workflow when requested (and enabled by the Workbench's default plan), so a good score can be checked against leakage or chance. Developed as part of Vahtian; Apache-2.0 licensed.

Jump to: About · What it does · Install · Quick start · What it looks like · EpiNet Workbench · Documentation · Scope and caveats · Tests · Citation · License

About this project

EpiNet began a few years ago as a personal Python coding project and has grown into a research-methods demonstrator with one organizing principle: honest evaluation is the default, not an afterthought. Where most modelling code optimizes a headline number, EpiNet is built to resist fooling itself — a label-permutation null, calibration, bootstrap intervals, and community-aware splitting live in the same run record rather than separate notebooks; contestability makes each call inspectable; and the optional federated layer shares findings across sites without pooling records. The name reads as Epistemic Network: the question is not just what the model predicts, but how well-founded each call is.

EpiNet's contribution is methodological reproducibility, not a new predictor: the default model is a standard random forest, with logistic regression and optional XGBoost alternatives, and the value is the conservative, auditable evaluation workflow wrapped around the estimator so the same checks travel with every analysis. It remains a research and education demonstrator — not clinical decision support.

What it looks like

Screenshot — the contestability lens on a lymphoma cohort

The contestability lens on the lymphoma grey-zone example. Left: how far each call is from flipping, with the most-contested tail shaded. Right: value of information — the features that most drive boundary flips (CD10, cyclin D1, Ki67, …). The same lens runs on any cohort. More figures in examples/sample-outputs/.

What it does

• Graph features — degree, weighted degree, clustering, component size, isolate flag, optional betweenness/closeness/PageRank.
• Honest outcome model — selectable estimator over graph features + node attributes: RandomForest by default, scaled regularized logistic regression as an interpretable comparator, optional XGBoost (pip install "vahtian-epinet[xgboost]"), or an optional tiny PyTorch MLP (pip install "vahtian-epinet[torch]"). Every option uses the same evaluation harness: discrimination (AUROC, AUPRC), classification (balanced accuracy, MCC, F1), calibration (Brier always; slope/intercept for binary outcomes), bootstrap CIs, permutation importance, a label-permutation null, community-aware splitting, small-cohort warnings, a reproducibility provenance block, and a TRIPOD+AI-flavoured model_card.md.
• Shortest paths — from sources to target nodes or target-outcome nodes, with per-target coverage.
• Feature-space clustering — k-means centroids + per-node distance to each outcome-class centroid (Euclidean or Ledoit–Wolf-shrunk Mahalanobis).
• Contestability (--run-contest) — the closed-form smallest feature-space move that flips a node's nearest-centroid class, plus a per-feature value-of-information ranking. See docs/methods.md.
• Input normalization — maps common column aliases onto the schema before validation; never silently (every rename logged, raw + normalized hashed).
• Epistemic claim gates — graph semantics and timing are recorded so a similarity graph is not mistaken for observed social, clinical, or exposure relations. See docs/epistemic-science.md.
• Federated pipeline (optional) — reconstruct the scaler, centroids, and contestability from per-site aggregates only, behind a fail-closed governance gate. See docs/federated.md and docs/governance-and-consent.md.
• Baselines & external validation — compare graph features against a node-embedding baseline and a no-information floor under the same harness, and validate a model on an independent cohort. See docs/validation.md.
• LLMvahti (experimental) — the same contestability lens pointed at LLM-as-judge verdicts: a blinded-second-rater audit where the human rates first, with bootstrapped inter-rater agreement, judge-confidence calibration, and per-verdict flip-distance over rubric criteria. Run it with the epinet-llmvahti command (--human/--judge CSVs); see docs/llmvahti.md.

Install

pip install vahtian-epinet   # from PyPI: the package + the `epinet` command

From a source checkout (for development or the example extras):

pip install -e .            # installs the package + the `epinet` command
pip install -e ".[dev]"     # also pytest + ruff + hypothesis (for development)
pip install -e ".[lidc]"    # pylidc, for the LIDC-IDRI / LUNA16 examples
pip install -e ".[excel]"   # xlrd + openpyxl, for the TCIA diagnosis spreadsheets
pip install -e ".[xgboost]" # optional XGBoost estimator backend
pip install -e ".[torch]"   # optional tiny PyTorch MLP estimator backend

requirements.txt lists the core runtime dependencies if you prefer not to install the package.

Quick start

epinet \
  --nodes synthetic_nodes.csv \
  --edges synthetic_edges.csv \
  --outcome-column Outcome \
  --target-outcome 1 \
  --model random_forest \
  --output-dir epinet_outputs

(epinet ... is the installed console command; python -m vahtian.epinet.toolkit ... works identically without installing.) This runs graph-feature generation, an honestly-evaluated outcome model, and shortest-path summaries side by side.

Key outputs in epinet_outputs/:

• model_metrics.json — discrimination, classification, calibration, iteration_summary, bootstrap CI, permutation test, data warnings, provenance
• model_card.md — TRIPOD+AI-flavoured human-readable model card
• model_feature_importance.csv — permutation importance (± importance_std)
• node_features.csv, shortest_paths.csv, nearest_targets.csv, target_coverage.csv, provenance.json, run_summary.json
• plots/*.png — network, calibration, learning curve, metric stability, confusion matrix, and more (see docs/methods.md)

The data format is documented in Data-format.md.

EpiNet Workbench

Local CSV-to-report interface — plan → run → optional UI

EpiNet Workbench is a local CSV-to-report interface. It does not replace the command line and it is not AutoML: it never searches models or chases a metric. It writes an analysis.yaml file and runs the same EpiNet engine as the CLI, so every interactive run is reproducible without the interface.

# 1. Plan: profile the data, infer a schema, write a reviewable config
epinet-workbench plan --nodes synthetic_nodes.csv --edges synthetic_edges.csv \
  --outcome Outcome --output analysis.yaml

# 2. Run: execute the plan and write a portable result bundle
epinet-workbench run --config analysis.yaml

# 3. UI (optional): the same plan/run behind a local Streamlit workbench
pip install -e ".[ui]"
epinet-workbench ui

The config — not hidden UI state — is the source of truth. The five-screen UI (Data → Schema → Plan → Run → Report) only ever builds an analysis.yaml and calls the same runner.

Supported modes

1. Single CSV — feature-space analysis for an ordinary table (no network inference claimed unless you opt into a similarity graph).
2. Nodes + edges — the full graph pipeline on real graph-shaped data.
3. Development + validation — runs external validation by default; the mode for publishable work.

Safety gates block runs that would produce nonsense (no/single-class outcome, ID used as a feature) and warn on the rest (suspected outcome leakage selected as a feature, missing validation cohort, identifier-looking columns). A too-small positive class downgrades model training to a descriptive report rather than fabricating metrics.

Scientific claims check. Every run distils its diagnostics into plain-language claim gates — written into the model card, claims_check.json, and the HTML report:

• Permutation null — signal above null vs signal not detected.
• Split sensitivity — random vs community-aware split, to expose a headline that leaned on leakage between connected cases.
• Baseline floor — does the model beat the no-information baseline? The margin is measured paired per split (model and floor share the same splits), with a Nadeau–Bengio-corrected interval, so the verdict is three-way: beats floor, at floor, or not resolvable at this n when the interval straddles the line and the data cannot yet say.
• External validation — run or not, and how far performance transported.
• Graph semantics — whether the edge definition and edge timing license a graph-shaped predictive or epidemiological claim.
• A standing "do not claim clinical utility unless…" caveat, generated into every report and not removable by theming.

Branded HTML report. epinet-workbench run also writes a self-contained, offline index.html into the bundle (summary, caveats, claims check, model card, metrics, calibration, permutation null, baselines, contestability, plots, provenance, CSV downloads) — the portable artifact to share or print to PDF. A theme block in analysis.yaml sets the brand, title, logo, colours, and plot palette (slide between the colourblind-safe Wong and the Vahtian Sentinel palettes) — but the caveats, claims check, and provenance are always rendered:

reporting:
  brand_name: "Vahtian / EpiNet"
  report_title: "EpiNet Analysis Report"
  logo_path: null
  primary_color: "#5E4F99"
  accent_color: "#8273C0"
  plot_palette: "wong"   # or "vahtian"

Outputs (the downloadable result bundle): index.html, analysis.yaml, model_metrics.json, model_card.md, claims_check.json, split_comparison.json, provenance.json, node_features.csv, node_contestability.csv, model_feature_importance.csv, baseline_comparison.csv, external_validation.json (publication mode), publication-quality plots/, and environment.txt.

Scope: research and education only. Not clinical decision support.

Documentation

Docs index — methods, examples, federated, governance, validation

• docs/methods.md — evaluation design (iterative evaluation, permutation null, community-aware splitting), the diagnostic figures, the contestability theory, and methodological boundaries.
• docs/examples.md — worked examples: shortest paths, the CiteMatch evidence graph, feature-space clustering, the pulmonary-nodule cohort, real LIDC-IDRI, and the Nordic lung-cancer quality-indicator network.
• docs/federated.md — the federated fit, federated contestability, the registry adapter, and the sealed-egress model.
• docs/governance-and-consent.md — what the governance gate enforces vs what remains a policy/legal responsibility (explicitly non-legal).
• docs/validation.md — representation baselines (incl. a node-embedding comparison) and external validation: does the model transport?

Each worked example also has a builder script and a walkthrough under examples/*_usecase.md; the federated and governance pipelines have runnable demos under examples/federated_* and examples/governance_*.

Scope and caveats

What EpiNet does not do, and what to add before any real-world use

The model is intentionally simple. It does not infer causality, outbreak dynamics, clinical risk, or intervention effects. Network features can be useful descriptors, but they can also encode sampling bias, measurement bias, and structural confounding. Use the outputs as exploratory evidence, not as decisions.

Before using EpiNet for health, education, welfare, employment, or public-sector decisions, add: domain-specific data validation; directed/temporal assumptions; uncertainty and sensitivity checks; external validation on independent outcome-linked data; privacy and governance review; and human review of any operational recommendation. The bundled cohorts are synthetic or small and selection-biased — see the per-example limits in docs/examples.md. For clinical prediction, align reporting with TRIPOD+AI; for AI interventions, the bar moves toward prospective evaluation (e.g. CONSORT-AI).

Tests and linting

How to run the tests and the linter

python -m unittest discover -s tests   # or: pytest  (adds the hypothesis property tests)
ruff check .

GitHub Actions runs both on every push and pull request across Python 3.10–3.12 (.github/workflows/tests.yml).

Citation

How to cite EpiNet

If you use EpiNet, please cite it via CITATION.cff (GitHub's "Cite this repository" button generates APA/BibTeX from it).

License

Apache License 2.0. See LICENSE and NOTICE. (Releases up to and including 0.4.2 were published under the MIT License and remain available under those terms; Apache-2.0 applies from 0.4.3 onward.)