Post-COVID Evidence Corpus — Analytics Dashboard

Search strategy

Search scope and keyword families used to build the corpus. Disease and intervention terms are grouped by sub-category so it's clear which families are well-covered and which are sparse — empty or small sub-groups are direct candidates for the next search-refinement iteration.

Databases searched

Disease & condition terms (38 terms across 4 families)

Intervention, trial-design & readout terms (86 terms across 6 families)

Discard terms (20 terms)

Pipeline funnel

Where the corpus shrinks and where signal concentrates. The drop from raw to dedup is mostly cross-source overlap; the drop to linked indicates how many articles are tied to a registered trial.

Source coverage

Each article is counted under every source that returned it. A high multi-source overlap means the pipeline is robust; a high single-source share means dropping any one source would leave gaps. 1381 of 7570 articles (18%) appear in 2+ sources — these are the highest-confidence records. 184 preprints (medRxiv/bioRxiv) are included and shown in amber.

Temporal trends

Articles tend to lag trial registrations by 2–3 years. A peak in trial starts without a corresponding wave of publications is the dark-trial signal — studies registered but not yet reported.

Where are the missing publications?

Of 307 completed trials, 220 (72%) have no linked publication in our corpus. This is the most actionable gap: completed trials that should have results, where evidence may be sitting in journal pipelines, conference abstracts, or unpublished. Recruiting and active trials have no expected publications yet, so their high unlinked rate is normal.

Top 10 completed trials with no linked publication (sorted by completion date)

Disease scope

Coverage of Post-COVID, ME/CFS, post-viral, and related conditions across the corpus, split by trials and articles.

Readout & measurement modalities

Detection of digital, voice, patient-reported, and functional readouts in the corpus. 1311 of 8308 records mention at least one modality. Sparse groups are themselves a finding — they flag measurement strategies the current corpus barely covers, and may motivate the next round of search refinement.

Where conditions and measurements meet

Cross-tabulation: how many records mention both a given condition family and a given readout type. This is the view that says which measurement strategies are actually being applied to which patient populations. Empty cells flag concrete gaps to address in the next search iteration.

Evidence tier pyramid

Each trial is scored by trial status + phase + results bonus (+5 if results are available). Higher tiers represent more mature evidence. Tier 13 corresponds to completed Phase III/IV trials with results, assuming the configured status-priority scale.

Phase × Status heatmap

Trials are grouped by their highest detected CT.gov phase and current recruitment/status. Mixed-phase studies such as Phase I/II are assigned to the higher phase.Where trials concentrate. Heavy density in Recruiting + Phase II tells you the field is mid-trial; density in Completed + Phase III is where evidence actually lands.

Population & cohort characteristics

Who is being studied — sex eligibility, age inclusion windows, enrollment scale, and (where baseline data is posted) race/ethnicity breakdown of enrolled participants.

Trial design & eligibility

Sex eligibility, enrollment size, and age inclusion criteria across all 738 interventional trials. Enrollment values combine actual enrollment for completed/reporting trials and planned enrollment for ongoing trials, as reported by ClinicalTrials.gov.

Enrolled population demographics — from 39 trials with posted results

Mean enrolled age across reporting trials: 44.4 yrs · Mean female proportion: 63.8% female · Total participants in race/ethnicity data: 16,896. Bars show aggregate counts across all trials that posted baseline demographics.

Population stratification

Cross-tab of age population type × sex eligibility. Answers queries like female-only trials in working-age adults directly.

Intervention landscape (types)

CT.gov intervention type taxonomy, computed across 738 trials (each trial counted at most once per type). The largest categories: OTHER (249 trials, 34% — multi-modal rehabilitation, lifestyle, and complex-care interventions); DRUG (191, 26%); BEHAVIORAL (157, 21%). The combined weight of OTHER and BEHAVIORAL exceeds DRUG, meaning the KG schema must accommodate interventions that are not drug-with-dose entities. Trials with combination or multi-arm designs may appear in more than one type.

Intervention names extracted from ClinicalTrials.gov

This chart is data-driven directly from the CT.gov intervention_names field. Each bar is an intervention name as registered in ClinicalTrials.gov, not an expert-normalized class. Each trial contributes at most once per intervention name.

Most frequently tested drugs across Post-COVID trials

This chart restricts the denominator to CT.gov trials whose intervention_types include DRUG (191 trials). Bars count only intervention names whose paired CT.gov intervention_type is DRUG. Names are raw CT.gov labels, not expert-normalized classes. Each trial contributes at most once per drug name. Note: 12 trial rows had unequal numbers of intervention names and types; only positionally paired entries were counted.

Which treatments are being studied — and how many trials cover them?

Each bar shows how many Post-COVID trials in this corpus test a given drug or intervention class. The list was compiled from expert knowledge of the Post-COVID treatment landscape — covering antivirals, immunomodulators, anticoagulants, autonomic agents, and repurposed drugs — then matched systematically against trial records from CT.gov (191 drug trials total). This gives an at-a-glance picture of where research effort is concentrated, which therapeutic hypotheses are being actively tested, and where coverage is thin. Trials can match more than one class.

Autoantibody hypothesis coverage

Autoantibody terms are read from autoantibody_terms.yaml and matched against CT.gov trial fields and article titles/abstracts. Out of 738 trials, 14 (2%) mention autoantibody-targeting therapies, but only 1 (0%) explicitly mention autoantibody targets or mechanisms. 14 trials test potentially autoimmune-directed therapies without explicit target/mechanism measurement. Detected literature signal: 125 articles mention autoantibody-targeting therapies and 0 articles mention specific targets/mechanisms.

Very few trials explicitly mention autoantibody targets or mechanisms, so a separate target chart is suppressed. This sparsity is itself part of the evidence gap.

Sponsor analysis

Industry vs. academic vs. NIH split tells you who's funding Post-COVID research. The top-15 sponsors view shows whether the field is dominated by a few large programs (RECOVER) or fragmented across many small efforts.

Endpoint coverage (primary vs secondary)

Coverage of measurement categories across primary endpoints (738 trials) and secondary endpoints (610 trials). Categories with strong presence in both columns are core constructs the KG must support; categories that appear mostly as secondary are tracked alongside as comorbidity or safety signals. The split also tells you which constructs carry explicit hypothesis weight (primary) versus which are recorded for context (secondary).

Knowledge graph preview

Two views of the intervention → outcome relationships the corpus already supports. The network view (top) is a force-directed layout with intervention nodes colored by CT.gov type — drugs in pink, behavioral in teal, devices in blue, procedures in amber — and outcomes as navy diamonds. Hub interventions (Acupuncture, Exercise) emerge in the dense centre; peripheral interventions like Stellate Ganglion Block sit at the edge anchored to a single outcome. The adjacency matrix (bottom) shows every co-occurrence including the 28 weak (single-trial) connections — empty or pale cells are coverage gaps the KG can flag. Across 738 trials, the strongest edge is Acupuncture → Fatigue (6 trials). Comparator arms (placebo, usual care, sham) are excluded from the intervention column. The full KG holds every (intervention, outcome, evidence_weight) triple plus population, study, and citation links — these views are one slice.

Field-normalized impact (NIH Relative Citation Ratio)

RCR is available for 5541 articles (73%). Median RCR = 1.06. 2853 articles cite above their field average (RCR > 1.0); 1588 are high-impact (RCR > 2.0). The histogram is clipped at the 99th percentile (15.61) for readability; 56 articles exceed that display limit, with maximum RCR = 418.51. Unlike raw citation counts, RCR controls for both publication year and field, so a 2026 paper with 2 citations can outrank a 2022 paper with 50 if it's cited at a higher rate for its age and topic. Better signal for evidence weighting in a young corpus.

Metadata completeness

Where records exist but are missing fields they should have. Trials without results are the biggest signal — that's evidence registered but not yet usable for the KG.

AI-assisted curation — overview

GPT was used as an AI-assistance, to complement what was missing from the search keyword terms and the fetched clincial trials.

Verification verdicts

Number of AI-detected IDs