Paper Database

Interactive database of the papers in the literature review, plus a catalogue of open-source ECG feature extractors. It has four views:

• Papers: filter by model category, year, evaluation category, evaluation metric, loss function, task, conditioning, data type, data format, number of leads, sampling rate, segment length, dataset, data access, code availability and publication type. Click a paper to see each metric it uses, how it was used, and any fairness, robustness or privacy analysis.
• Metrics by evaluation category: pick Fidelity, Diversity, Utility, Privacy, Fairness, Robustness, Clinical validity, Human / expert, Conditional consistency or Efficiency to list the metric families in that category, the number of papers using each, and whether ECGEN-Eval implements it. Click a metric to list the papers and how each one used it. The filters on the Papers tab also apply here.
• Feature extractors: open-source ECG feature tools usable from Python (and the published reference measurements in PTB-XL+). Filter by category, licence, input type, lead support and by the features you need; click a tool for its accepted inputs, sampling-rate handling, minimum duration, native output names and the API that produces each feature.
• SQL console: run read-only SQL against the database.

Quick links: Fidelity · Diversity · Utility · Privacy · Fairness · Robustness · Clinical validity · Feature extractors

How the data was produced

Every paper was read in full text. Each metric was recorded as the paper names it, then mapped to a canonical metric family and one or two evaluation categories. Some fields were judgement calls, so check the original paper before citing a detail. Filters within a group match any selected value by default; click match ANY to switch the group to match ALL.

The SQLite database

The page opens the database directly in your browser with sql.js (SQLite compiled to WebAssembly). Nothing is sent to a server. Use the SQL console above for queries the filters can't express.

• ecgen_literature.sqlite: the database (open it with sqlite3, DB Browser for SQLite, pandas, …)
• papers.csv: one flat row per paper, for spreadsheets
• feature_extractors.csv: one row per feature extractor with its input specification
• feature_support_matrix.csv: canonical feature × extractor matrix (x = the tool returns that feature)

import sqlite3
con = sqlite3.connect("ecgen_literature.sqlite")
con.execute("SELECT * FROM metric_usage").fetchall()

Schema

The schema stores filter dimensions as data, so new categories, metric families or whole new facets are added as rows rather than new tables:

Table / view	Contents
papers	One row per paper: bibliographic data, model family, task, datasets, leads, sampling rate and segment length, verbatim loss_functions and eval_metrics, descriptions of any fairness / robustness / privacy analysis, notes and takeaway for ECGEN-Eval
evaluations	One row per metric or protocol a paper uses: name as in the paper, canonical metric family, and how it was used
evaluation_categories	Evaluation categories of each evaluation (Fidelity, Diversity, Utility, Privacy, Fairness, Robustness, Clinical validity, Human / expert, Conditional consistency, Efficiency)
authors, paper_authors	Authors and their order on each paper
facets	Filter dimensions, e.g. category, year, eval_category, metric, loss, task, conditioning, data_type, data_format, leads, sampling_rate, segment_length, dataset, data_access, code, venue_type
terms	Values of each facet. For metrics, parent_id is the default evaluation category and ecgen_eval names the matching ECGEN-Eval metric key; categories have a description
paper_terms	Which terms apply to which paper
paper_facets (view)	One row per (paper, facet, term), the easiest table to filter on
evaluation_details (view)	One row per (evaluation, category) with paper, metric name, family, ECGEN-Eval key and how it was used
metric_usage (view)	Papers per (evaluation category, metric family), with ECGEN-Eval coverage
meta	Build date, scope and source

Feature-extractor catalogue (independent of the paper tables):

Table / view	Contents
extractors	One row per tool: category, links, install command, licence and licence class (permissive, copyleft, none, noncommercial, proprietary), version, language and runtime, paper, method, notes
extractor_inputs	What the tool accepts: signal types, input kinds, array shape, multi-lead support, sampling-rate argument / default / supported range / recommended minimum, minimum duration, amplitude units, built-in preprocessing, output format and level
extractor_input_kinds, extractor_signal_types	One row per input kind (raw_1d, raw_multilead, wfdb_record, csv, rpeaks, rr_intervals, …) or signal type, for filtering
feature_vocab	Canonical feature vocabulary (fiducial points, intervals, amplitudes, axes, HRV, signal quality, …). ecgen_eval_column links a feature to ECGEN-Eval's summary_df column
extractor_features	One row per (tool, canonical feature, API): the tool's native output name(s), the function that produces them and the level (beat, lead, record, window)
extractor_evidence	Source files, docs and papers read to verify each entry
extractor_feature_details (view)	Tool + feature + native name + API in one row
feature_coverage (view)	For each canonical feature, how many tools provide it and which
ecgen_feature_support (view)	For each ECGEN-Eval feature column, every tool that can provide it, with its input requirements

Example: privacy metrics and the papers that use them.

SELECT family, metric_name, paper_id, year, how
FROM evaluation_details
WHERE category = 'Privacy'
ORDER BY family, year DESC;

Example: papers that use both an embedding FID and a Turing test.

SELECT p.id, p.year, p.title
FROM papers p
WHERE p.id IN (SELECT paper_id FROM paper_facets WHERE facet = 'metric' AND term = 'FID / embedding Fréchet')
  AND p.id IN (SELECT paper_id FROM paper_facets WHERE facet = 'metric' AND term = 'Expert / Turing test');

Example: permissively licensed tools that return both QRS onset and T offset (needed for QT) and how they take the sampling rate.

SELECT x.name, x.license, i.input_kinds, i.multilead, i.fs_param, i.fs_supported
FROM extractors x JOIN extractor_inputs i ON i.extractor_id = x.id
WHERE x.license_class = 'permissive'
  AND x.id IN (SELECT extractor_id FROM extractor_features WHERE feature_id = 'qrs_onset')
  AND x.id IN (SELECT extractor_id FROM extractor_features WHERE feature_id = 't_offset');

Rebuilding

The database is generated by literature_review_claude/build_db.py from the review's papers_index.csv and the per-metric records in structured_extraction.jsonl. The extractor catalogue comes from literature_review_claude/feature_extractors/ (feature_vocab.csv and tools.jsonl; see its README for the record schema):

python literature_review_claude/build_db.py         # papers with full text
python literature_review_claude/build_db.py --all   # also include abstract-only entries