Detecting post-acute COVID-19 vaccination syndrome in pharmacovigilance data: a pre-registered analysis of v-safe free text

Abstract Introduction. Post-acute COVID-19 vaccination syndrome (PACVS) is a chronic multi-system condition that pharmacovigilance systems code one symptom at a time. CDC's v-safe collected free-text answers at 3, 6 and 12 months after a dose; one prior analysis used its structured answers, none its free text. Objective. To test, under a pre-registered protocol, whether long-term v-safe free text shows a coherent, persistent and under-captured multi-domain symptom pattern. Methods. Four questions were registered (osf.io/gnfpd) before any long-term text was read: clustering of nine candidate symptom domains, early onset and persistence, under-capture and dispersed coding in VAERS, and burden. Text was classified by a frozen keyword lexicon plus a language model and validated against CDC's MedDRA coding, with a confirmatory gate of precision and recall of at least 0.85. Results. Only fatigue passed the gate, so the primary analysis could not be run; every result is the registered nine-domain sensitivity analysis. Of 7,819,251 entries, 1,595,121 were long-term, and of 8.62 million person-doses with a long-term survey, 87,229 (1.01%) met the case definition. Domain pairs co-occurred 1.25 times as often as chance (p = 0.001) in every dose and manufacturer stratum and in 1.06 million VAERS reports. Nulls matched on how much each person wrote, or fixing each person-dose's own domain count, removed that aggregate excess but left a core in which cognitive symptoms co-occurred with sleep 2.3 to 2.7 times and with fatigue 2.1 to 2.3 times more than chance, in women and in men. The registered onset test was uninformative (28% early against a schedule-dependent 71% expected). Symptoms recurred at a later survey in 47 to 60% of cases, depending on how cases qualified, against 1.5 to 10.5% of comparison groups. Cases reached VAERS slightly less often than people with named acute diagnoses (rate ratio 0.94, 95% CI 0.90 to 0.98), and their reports were coded to a narrower, not wider, spread of terms. Conclusion. Of person-doses with a long-term v-safe survey, 1.01% describe a multi-domain pattern whose cognitive, fatigue, sleep and dysautonomia core clusters beyond chance after adjustment for verbosity. When the symptoms began could not be established, and the data cannot attribute the pattern to vaccination rather than to intercurrent illness or selection into long-term response. The classifier did not meet its pre-registered accuracy standard, although a blind single-coder audit of 306 disagreements found it right in 86% of its unmatched positives against 60% for CDC's coding; the findings are hypothesis-generating.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-16
DOI
https://doi.org/10.21203/rs.3.rs-11721193/v1
Primary Topic
Pharmacovigilance and Adverse Drug Reactions
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Detecting post-acute COVID-19 vaccination syndrome in pharmacovigilance data: a pre-registered analysis of v-safe free text

Matthew T. J. Halma
4 links
Research Square
Pharmacovigilance and Adverse Drug Reactions
preprint

Detecting post-acute COVID-19 vaccination syndrome in pharmacovigilance data: a pre-registered analysis of v-safe free text

Matthew T. J. Halma
preprint en

Abstract

Abstract Introduction. Post-acute COVID-19 vaccination syndrome (PACVS) is a chronic multi-system condition that pharmacovigilance systems code one symptom at a time. CDC's v-safe collected free-text answers at 3, 6 and 12 months after a dose; one prior analysis used its structured answers, none its free text. Objective. To test, under a pre-registered protocol, whether long-term v-safe free text shows a coherent, persistent and under-captured multi-domain symptom pattern. Methods. Four questions were registered (osf.io/gnfpd) before any long-term text was read: clustering of nine candidate symptom domains, early onset and persistence, under-capture and dispersed coding in VAERS, and burden. Text was classified by a frozen keyword lexicon plus a language model and validated against CDC's MedDRA coding, with a confirmatory gate of precision and recall of at least 0.85. Results. Only fatigue passed the gate, so the primary analysis could not be run; every result is the registered nine-domain sensitivity analysis. Of 7,819,251 entries, 1,595,121 were long-term, and of 8.62 million person-doses with a long-term survey, 87,229 (1.01%) met the case definition. Domain pairs co-occurred 1.25 times as often as chance (p = 0.001) in every dose and manufacturer stratum and in 1.06 million VAERS reports. Nulls matched on how much each person wrote, or fixing each person-dose's own domain count, removed that aggregate excess but left a core in which cognitive symptoms co-occurred with sleep 2.3 to 2.7 times and with fatigue 2.1 to 2.3 times more than chance, in women and in men. The registered onset test was uninformative (28% early against a schedule-dependent 71% expected). Symptoms recurred at a later survey in 47 to 60% of cases, depending on how cases qualified, against 1.5 to 10.5% of comparison groups. Cases reached VAERS slightly less often than people with named acute diagnoses (rate ratio 0.94, 95% CI 0.90 to 0.98), and their reports were coded to a narrower, not wider, spread of terms. Conclusion. Of person-doses with a long-term v-safe survey, 1.01% describe a multi-domain pattern whose cognitive, fatigue, sleep and dysautonomia core clusters beyond chance after adjustment for verbosity. When the symptoms began could not be established, and the data cannot attribute the pattern to vaccination rather than to intercurrent illness or selection into long-term response. The classifier did not meet its pre-registered accuracy standard, although a blind single-coder audit of 306 disagreements found it right in 86% of its unmatched positives against 60% for CDC's coding; the findings are hypothesis-generating.

Research Square
Open Source Medicine Foundation (EE)
Good health and well-being
Pharmacovigilance and Adverse Drug Reactions
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon