VAERS reports are things that meet criteria, not any random event in an EMR.
- 1% of the events their untested, unvalidated model which is not described from 13 years ago flagged matched clinical reports. There's no almost no modeling they could have done that could be causal, and no reason to think that they're not just flagging common and largely irrelevant things that correlate with vaccine administration.
A reporting protocol & corresponding
algorithms were developed to detect potential adverse event cases using diagnostic codes, and
methods were tested to identify prescriptions or abnormal laboratory values that might be
suggestive of an adverse effect. These algorithms were designed to seek both expected and
unexpected adverse effects.
We had initially planned to evaluate the system by comparing adverse event findings to those
in the Vaccine Safety Datalink project [...] the components under this particular Aim were not achieved.
- There is no reason to even suspect that the criteria their model uses are more relevant than the peer reviewed criteria used to set criteria for reporting. Overwhelming reporting systems with noise is not useful. Their report lists these as possible reactions, and even that is a reach.
Of these doses, 35,570 possible reactions (2.6 percent of vaccinations) were identified.
Based on a conference paper made public several years later, of 209 physician-reviewed events that were flagged by their algorithm (which may be different from the one referenced here), 193 were not vaccine related.
This is a grant report. The authors do not have any reason to say that their modeling was a failure, as this is not peer reviewed. They have every incentive to say that there is open problem with how things are reported and their solution works. You will note exactly 0 peer reviewed publications listed in this grant report. This was not a successful project.
Some amount of adverse reactions will not be reported, of course - that is the nature of any reporting system. I don't think this kind of project is a bad idea, but the estimate referenced is entirely unsupported. It's no better then a random guess