Big Pharma's worst nightmare

LLMs which can assist vaccine safety research

Summary:

• The investigation began after Helen Petousis-Harris, who had received funding from Pfizer and GSK according to her GVDN paper, dismissed vaccine safety concerns on Twitter.

• The analysis stems from the GVDN paper's use of a specific threshold (Lower Bound Confidence Interval > 1.5) to identify "prioritized safety signals" for vaccine adverse events.

• VAERS data from Japan shows 50 cases of ADEM (Acute disseminated encephalomyelitis) within 30 days of Pfizer vaccination, compared to an expected 10.89 cases from their 196 million procured doses.

• Three different LLMs (ChatGPT, Claude, and Google Gemini) independently calculated the Observed/Expected ratio and confidence intervals, all confirming this met the criteria for a "prioritized safety signal."

• Large Language Models now enable complex statistical calculations for vaccine safety analysis, making independent verification more accessible.

Recently there was an in-depth reply tweet by Jikkyleaks as a response to Helen Petousis-Harris tweet about why Turbo Cancers are BS.

Link to tweet

GVDN paper makes the Pfizer vaccine look relatively safe compared to the others

There was a bit more back and forth, and then we found out that Helen had taken money from Pfizer before (emphasis mine)

Anders Hviid reports a relationship with Novo Nordisk Foundation that includes: funding grants. Anders Hviid reports a relationship with VAC4EU that includes: consulting or advisory. Finnish Institute for Health and Welfare (THL) conducts Public-Private Partnership with vaccine manufacturers and has received research funding from Sanofi Inc. Petteri Hovi has been an investigator in these studies, but has received no personal remuneration. Helen Petousis-Harris has served on expert advisory boards and had speaking engagements for Pfizer and GSK. She has also received research funding from GSK. She has not received any personal honoraria. If there are other authors, they declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Source

I had read that GVDN paper before and remember thinking that it made the other vaccines look worse than the Pfizer vaccine.

Now, this could be true given that all the COVID19 vaccines were pretty dangerous, but the most interesting phrase in that paper was “prioritized safety signal”, which was used to discount some Pfizer safety signals using a Lower Bound Confidence Interval (LBCI) of over 1.5 (emphasis mine):

Statistically significant differences were also found for transverse myelitis (OE ratio = 1.91; 95 % CI: 1.22, 2.84) and ADEM (OE ratio = 2.23; 95 % CI: 1.15, 3.90) after a first ChAdOx1 dose. Bell’s palsy had an increased OE ratio after a first dose of BNT162b2 (1.05; 95 % CI: 1.00, 1.11) and mRNA-1273 (1.25; 95 % CI: 1.11, 1.39). There were also increased OE ratios for febrile seizures following a first and second dose of mRNA-1273 (1.36, 95 % CI: 1.02, 1.77 and 1.44, 95 % CI: 1.04, 1.95, respectively), and for generalised seizures following a first mRNA-1273 dose (1.15, 95 % CI: 1.10, 1.20) and a fourth BNT162b2 dose (1.09, 95 % CI: 1.04, 1.14). No increased OE ratios were identified following a third dose of any vaccine. The results are concordant with the OE ratios of homologous schedules; however, an increased OE ratio for generalized seizures following a homologous schedule of four doses of mRNA-1273 (1.33; 95 % CI: 1.07, 1.63) was identified (Fig. 1). These outcomes did not meet the threshold for a prioritised safety signal following vaccination.

Now that Large Language Models (LLMs) are able to easily do these confidence interval calculations, I wondered if I could use them to find problems with the Pfizer vaccine.

Acute disseminated encephalomyelitis

I noticed that ADEM - Acute disseminated encephalomyelitis - had a very low expected value (2 expected cases for 36 million doses¹ of Moderna vaccine)

So I looked at the number of cases in the Foreign Serious VAERS reports search engine².

There were over 50 cases of ADEM for Japan within the first 30 days of taking a Pfizer vaccine.

How to reproduce this:

Select Acute Disseminated encephalomyelitis from symptoms

Select Country = Japan

Select all numbers less than or equal to 30³ in the days to symptom onset

And select Pfizer as the vaccine as shown in the first image.

Click here to see the full list of ADEM cases from Japan

What is the O/E ratio for ADEM in Japan?

Japan secured 194 million doses, which would be the upper limit for the maximum number of Pfizer doses administered to its population.

Source

I just asked the same question to three different LLMs - ChatGPT (v3), Google Gemini and Claude:

Japan procured 196 million doses of the Pfizer vaccine. Based on expected background rates, if we administer 36 million doses of any vaccine, we anticipate 2 cases of Acute disseminated encephalomyelitis if we monitor the patient for 42 days. Japan had 50 cases for the 196 million upper limit of doses administered. What is the observed/expected ratio (O/E ratio), and what is the confidence interval? If the Lower Bound Confidence Interval (LBCI) is more than 1.5, it is considered a "prioritized safety signal". Would ADEM be a prioritized safety signal? Please explain your thought process.

Response from Claude

Let me solve this step by step.

1) First, let's calculate the expected cases of ADEM for 196 million doses:

   * If 36 million doses → 2 cases

   * Then 196 million doses → (196/36) × 2 = 10.89 expected cases

2) Now we can calculate the O/E ratio:

   * Observed (O) = 50 cases

   * Expected (E) = 10.89 cases

   * O/E ratio = 50/10.89 = 4.59

3) For the confidence interval of a rate ratio, we can use the formula:

   95% CI = exp[ln(O/E) ± 1.96 × √(1/O)]

   * ln(4.59) = 1.524

   * 1/O = 1/50 = 0.02

   * √(0.02) = 0.141

   * 1.96 × 0.141 = 0.276

   Lower bound = exp(1.524 - 0.276) = exp(1.248) = 3.48

   Upper bound = exp(1.524 + 0.276) = exp(1.800) = 6.05

Therefore:

* O/E ratio = 4.59 (95% CI: 3.48 - 6.05)

Since the lower bound (3.48) is greater than 1.5, ADEM would be considered a prioritized safety signal.

This indicates that the observed rate of ADEM was significantly higher than what would be expected based on background rates, with even the most conservative estimate (lower bound) showing more than 3 times the expected rate.

Response from ChatGPT

To determine if Acute disseminated encephalomyelitis (ADEM) is a prioritized safety signal based on Japan's data, let's break down the steps to calculate the observed/expected ratio (O/E ratio) and the confidence interval, followed by checking if the lower bound of the confidence interval (LBCI) is more than 1.5.

Step 1: Calculate the Expected Cases

The background rate suggests 2 cases for every 36 million doses administered. Japan administered 196 million doses, so we can scale this background rate to calculate the expected cases:

Expected Cases=236×196

Expected Cases=2×19636=39236≈10.89

So, the expected number of ADEM cases after administering 196 million doses is approximately 10.89.

Step 2: Calculate the Observed/Expected (O/E) Ratio

The observed number of ADEM cases in Japan is 50. The O/E ratio is given by:

O/E Ratio=Observed CasesExpected Cases=5010.89≈4.59

This O/E ratio of 4.59 suggests that the observed number of cases is higher than expected.

Step 3: Calculate the Confidence Interval for the O/E Ratio

To calculate the confidence interval, we typically use the Poisson distribution for rare events like ADEM cases. The confidence interval for the observed count O can be approximated as:

CI=(OE−z⋅O‾‾√E,OE+z⋅O‾‾√E)

wherezis the z-score for the desired confidence level (typically 1.96 for a 95% confidence interval).

Plugging in our values:

1. Observed cases (O) = 50

2. Expected cases (E) ≈ 10.89

3. O/E ratio ≈ 4.59

4. Standard error = 50√10.89≈0.653

5. 95% CI = 4.59±1.96×0.653

Calculating:

LBCI=4.59−1.96×0.653≈4.59−1.28≈3.31

UBCI=4.59+1.96×0.653≈4.59+1.28≈5.87

Step 4: Interpret the Confidence Interval

The 95% confidence interval for the O/E ratio is approximately (3.31, 5.87).

Step 5: Determine if ADEM is a "Prioritized Safety Signal"

Since the lower bound of the confidence interval (LBCI) is 3.31, which is greater than 1.5, ADEM meets the criterion for being considered a prioritized safety signal.

Conclusion

Yes, ADEM would be considered a prioritized safety signal in this scenario because the LBCI of the O/E ratio (3.31) is greater than 1.5. This indicates that the observed cases significantly exceed what would be expected based on background rates, warranting further investigation.

Response from Google Gemini

Here's how to calculate the O/E ratio and confidence interval for ADEM following the Pfizer vaccine in Japan:

1. Calculate the expected number of ADEM cases:

• The background rate is 2 cases per 36 million doses.

• Japan administered 196 million doses.

• Expected cases = (196 million doses / 36 million doses) * 2 cases = 10.89 cases (approximately 11 cases)

2. Calculate the observed/expected ratio (O/E):

• Observed cases = 50

• Expected cases = 10.89

• O/E ratio = 50 / 10.89 = 4.59

3. Calculate the 95% confidence interval:

We'll use the Poisson distribution to calculate the confidence interval since we're dealing with rare events (counts). There are several ways to calculate the CI for a Poisson distribution; a simple method uses the square root of the observed events for estimation:

• Lower bound ≈ Observed cases - 1.96 * sqrt(Observed cases) = 50 - 1.96 * sqrt(50) = 50 - 13.86 = 36.14

• Upper bound ≈ Observed cases + 1.96 * sqrt(Observed cases) = 50 + 1.96 * sqrt(50) = 50 + 13.86 = 63.86

Now, we need the confidence interval for the O/E ratio, not just the observed cases. We can achieve this by dividing the confidence intervals for the observed cases by the expected number of cases:

• LBCI = 36.14/10.89 = 3.32

• UBCI = 63.86/10.89 = 5.86

Therefore, the 95% confidence interval for the O/E ratio is approximately 3.32 to 5.86.

4. Determine if ADEM is a prioritized safety signal:

The LBCI (3.32) is greater than 1.5. Therefore, ADEM would be considered a prioritized safety signal following Pfizer vaccination in Japan based on these calculations.

Thought Process Summary:

The observed number of ADEM cases was substantially higher than what would be expected based on background rates. The calculated O/E ratio quantifies this difference (4.59), and the confidence interval (3.32-5.86) suggests that this elevated risk is statistically significant (it doesn't include 1, which would represent no difference). Because the LBCI exceeds 1.5, this meets the pre-defined criteria for a prioritized safety signal, indicating that further investigation and monitoring of this potential association are warranted.

Important Considerations:

• This is a simplified calculation. A more precise confidence interval for a Poisson distribution, especially with larger observed counts, might use other methods, leading to slightly different results. Specialized statistical software is often employed for these calculations.

• Correlation does not equal causation. While the elevated O/E ratio suggests a possible link between the Pfizer vaccine and ADEM, it doesn't definitively prove that the vaccine caused the increased ADEM cases. Other factors could be at play. This calculation simply flags a potential safety concern that requires further investigation.

• Background rates are not always precise, and may vary between populations. The accuracy of the expected case calculation depends on the reliability of the background rate used.

• "Prioritized safety signal" is a pre-defined criterion based on risk assessment principles; the 1.5 threshold is a pragmatic choice to trigger further investigation, not an absolute indicator of causality.

As expected, Google Gemini goes the extra mile on behalf of Pharma :-)

Conclusion

Anyone with some basic skills in data science could have created the VAERS data subsets which I used for the calculations. However, the tricky part would have been the calculations for the O/E ratio including the confidence intervals and such. But we can see that LLMs can not only do the calculations, they can also explain the calculations. And using multiple LLMs help us cross verify that the LLMs are not hallucinating⁴ any numbers :-)

As Large Language Models improve, it will become easier and easier to do all these cross verifications, and it is going to get harder and harder for Big Pharma to obfuscate their findings⁵.

1. In the Results section, they state the total number of Moderna doses - “The total vaccinated population across all sites comprised 99,068,901 individuals. Most vaccine recipients were in the 20–39 and 40–59-year age groups (Table 1). In total, 183,559,462 doses of BNT162b2, 36,178,442 doses of mRNA-1273, and 23,093,399 doses of ChAdOx1 were administered across all the sites in the study periods. The highest numbers of doses were administered in France (120,758,419), followed by Canada – Ontario (32,159,817) and Australia – Victoria (15,617,627).“

2. Reminder: it is based on a copy of the VAERS CSV file I downloaded for November 2022, right before the purge

3. Yes, this means we are ignoring a few cases which could have potentially had days to symptom onset between 31 - 42

4. By the way, hallucination is already quite rare when it comes to step by step calculations like these. But as I have demonstrated, you can send the same question to multiple LLMs just to be safe.

5. And just to be very clear, I am not saying they are obfuscating anything in the GVDN paper. Japan is not even a part of the GVDN network. But we can use their own analysis and apply it to other data subsets to see if the vaccine safety signals are internally consistent across regions.