I argue that pathogenic microorganisms are contributing to virtually every major chronic disease today. Though it will some time before publishing a completed hypothesis espousing as much, and longer until EMR conducts trials relating to the empirical treatment of chronic infection, readers will get introduced to this idea piecemeal through our article series. Today I will examine one facet of this hypothesis – namely antibiotic side effects. In my article on the safety of dietary supplements, I discussed the Jarisch-Herxheimer Reaction (JHR). JHRs are a kind of reaction that an infected person may have in response to an anti-infective therapy. Such reactions are an indication that a person is carrying an infectious disease. In the present article, I examine patterns in reported drug reactions that seem to distinguish antibiotics as an entire class of drugs from non-antibiotic drugs. I hypothesize that these patterns are best explained as being a result of JHRs related to undiagnosed chronic infection. I further hypothesize that this is the case even with a number of dietary supplements.
Adverse Reactions to Antibiotics Hint at Undiagnosed Chronic Infection
Say a person goes to the doctor with a urinary tract infection. The doctor prescribes antibiotics. After taking the antibiotics, the patient unexpectedly experiences fever and chills. The doctor then potentially reports this reaction to the relevant authorities whose job it is to monitor real-world adverse drug reactions (ADRs). There are thousands of types of ADRs across all drugs that get reported in this manner. An ADR isn’t proof that there is a causal relationship (i.e. that the reaction is actually a drug side effect), but the more of a specific type of ADR that gets reported, the more it raises the suspicion. In the above example, many might assume that the fever and chills could be due to the patient being allergic to the drug, coincidence, or some other explanation. But I ask the question – could many ADRs reported for antibiotics actually have little to do with things like the toxicity of the drug or allergy? Could they actually be JHRs that give us a clue that people are carrying undiagnosed chronic infectious disease? In summary, that’s what my preliminary analysis suggests.
Selecting Drugs for the Calculations
I identified 10 of the most commonly prescribed non-antibiotic drugs in the US. Then I identified 7 of the most commonly prescribed antibiotics in the US. I required that no two drugs belong to the same class. For example, there could not be two statins drugs or two fluoroquinolones, etc. Usage data for these drugs in the US was obtained from studies that used the National Health and Nutritional Examination Survey (NHANES) for the 2-year cycle of 2011-2012.1,2 In these studies, usage was measured by percentage of US population that used a specific drug within the last 30 days. For 4 out of 7 antibiotics, I could not actually find usage data. However, because usage data was available for grouped classes of antibiotics, I was able to select these 4 antibiotics based on them being the most commonly recognized antibiotic within their respective classes.
The 10 non-antibiotic drugs this yielded were simvastatin, lisinopril, levothyroxine, metoprolol, metformin, hydrochlorathiazide, omeprazole, amlodipine, albuterol, and citalopram. The 7 antibiotic drugs this yielded were ciprofloxacin, amoxicillin, cephalexin, clindamycin, metronidazole, azithromycin, and doxycycline.
Obtaining Adverse Drug Reaction Data
I used the World Health Organization’s public ADR database called VigiAccess3. This is a publicly available summary version of their full database called VigiBase. For the purposes of my rough calculations, I only needed the data in VigiAccess. It allows you to search by any drug name and it will tell you how many ADRs have ever been reported for that drug. You can look at the total ADRs or specific ADRs. You can also break down the total number of ADRs (but not specific types) by year, or by geographic location (but not both simultaneously). Data was accessed in February 2017.
I also accessed data from the FDA’s reporting system in order to repeat the same calculations from two different sources to see if my results were consistent. Until specifically stated, the data discussed in this article refers to VigiAccess.
While data from VigiBase would have been the most reliable, I resorted to using some surrogate data that could be calculated from the data available in VigiAccess. The query required to obtain the data from VigiBase would have cost nearly $5,000, which was prohibitive.
Sorting by Highest Proportional Reporting Ratios Yields ADRs that Look like Infection
There is a statistic called proportional reporting ratio (PRR) that tells something about how often a specific ADR is reported for a given drug, relative to another drug.4 For example, if drug A has received 1,000 total ADRs, and 20 of them are for coughing, then we know that the proportion of adverse events of coughing for this drug is 20/1000=0.02. Then say if drug B has received 1500 total ADRs, and 45 of them are for coughing, then we know the proportion of adverse events of coughing for this drug is 45/1500=0.03. Then we take the ratio of these two proportions and get 0.03/0.02=1.5. Since this ratio turns out to be greater than 1, this tells us that drug B seems to have a larger proportion of ADRs for coughing than does drug A. The higher this ratio, the greater the suspicion of there being a causal relationship between that drug and the specific type of ADR.
I compiled a list of all possible ADRs across all 17 drugs. This yielded well over 3,000 types of ADRs. Then I pooled together the counts of each specific ADR – as well as the total ADR counts – within each drug group respectively (i.e. non-antibiotics and antibiotics). Then I eliminated any type of ADR which did not represent at least 1% of the total number of ADRs for at least one of the two pooled drug groups. This cut the list down to a total of just 61 ADRs. Then I calculated the PRRs for these 61 types of ADRs. The results are summarized below in Table 1:
|25||Abdominal pain upper||0.72||1.38|
|26||Pain in extremity||0.62||1.62|
|38||Acute kidney injury||0.43||2.30|
|47||Therapeutic response decreased||0.25||4.04|
|55||Drug dose omission||0.10||9.67|
|56||Product quality issue||0.10||10.08|
|57||Blood glucose increased||0.07||14.27|
As you scroll down this table, any ADRs that are above the red line happen in greater proportion with antibiotics, relative to non-antibiotics. And any ADRs below the red line happen with greater proportion with non-antibiotics, relative to antibiotics.
As we can see, antibiotics yield a greater proportion of ADRs such as rashes, anaphylaxis, hives (urticaria), itching (pruritus), fever, chills, and so on. Now this is where potential author bias may come into play, but when I look at these top ADRs for antibiotics, they strike me as being reminiscent of infectious disease. Fever and chills are perhaps the most reminiscent. Note that “hypersensitivity” could also involve a fever.
Examples of ADRs that non-antibiotics yield a greater proportion of are suicide, low sodium (hyponatremia), weight gain, cough, hypertension, hypotension, myalgia, constipation, fatigue, and depression. When looking at this list, the first thing we must do is discard anything that requires a substantial length of time to develop. For example, you would not expect to gain weight from a 10-day course of antibiotics. That is a type of ADR that is only relevant to drugs that are meant to be used long-term. We also have to discard other irrelevant items like “drug dose omission”, etc. After discarding such items, we have to ask ourselves, “could I have just as easily suspected these were related to infection had I been blinded to drug group?” Though in theory, virtually any symptom imaginable could be caused by a JHR, I think the answer is “no”.
In other words, the kinds of ADRs antibiotics yield initiate suspicion that warrants a closer look.
Computing a Pseudo-Rate Ratio Suggests a 5-Fold Difference in ADR Rates
The statistics in Table 1 are only relative to the total number of ADRs received for a given group of drugs. In other words, they in no way say anything about the absolute population sizes that lead to those ADRs. That would be needed in order to calculate incidences and relative risk (RR). And now I incorporate such data. In general, this is an invalid thing to do with a spontaneous reporting system. But I will argue that in this case, it is actually a useful indicator that helps us generate a hypothesis. The reason being is that the signal I apparently detect appears to be so large that the caveats and biases of such an action cannot render the signal as being artefactual.
I already described my sources for percent drug use in the last 30 days in the US for the 2-year cycle of 2011-2012, for the 17 drugs in question. I limited my calculations to the year 2011, and hence I assumed that this 2-year data was a fair representation of just the year 2011 as well. Trends in drug use suggested that this was a fair assumption. I took the 2011 US population to be 312 million. Multiplying the two gives estimated numbers of person-years of use for non-antibiotics and antibiotics.
I filtered total number of ADRs for each drug by the year 2011. For each drug, I then scaled down these 2011 worldwide ADRs by multiplying by a geographic percentile for the “Americas”. This served as a surrogate count of total ADRs for each specific drug, for just the year 2011, for just the USA. (Caveats such as this are discussed later.)
Then I pooled the ADR counts for the drugs within each of the two respective classes together (antibiotics and non-antibiotics) and calculated a “pseudo”-rate ratio. I also calculated an extreme-case pseudo-rate ratio by looking at only the antibiotic with the lowest ADR incidence compared to the non-antibiotic with the highest ADR incidence.
The results are summarized in Table 2 below:
|Pseudo-Rate Ratio (P-RR)b||5.14|
These calculations suggest that the average antibiotic leads to an estimated 5-fold number of ADRs per person-year as compared to the average non-antibiotic drug. And remember, this isn’t just a single antibiotic outlier being compared to a group of non-antibiotic drugs. This is a pooled group of antibiotics. This means that unless someone can come up with an example of a very dramatic reporting bias or caveat that would specifically afflict antibiotics, this large discrepancy is likely primarily due to the anti-infective functions of the drugs themselves. After all, what else do they have in common? Hence we hypothesize that the antibiotics are interacting with undiagnosed chronic bacterial infection, resulting in Jarisch-Herxheimer reactions, which are then getting reported as ADRs.
If we just look at the apparently least dangerous antibiotic (doxycycline) and compare it to just the apparently most dangerous non-antibiotic (omeprazole), we get a rate ratio of 1.04. So this suggests that the tendency for antibiotics to yield more ADRs than non-antibiotics may be across the board, and not just due to an outlier skewing the results.
Caveats and Biases Unlikely to Explain It
Calculating rate ratios in this manner is scarcely a valid technique owing to many caveats. ADRs have an estimated 94% median rate of under-reporting.5 Now if this under-reporting rate were perfectly constant for every ADR type, and for every drug, then it wouldn’t bias the calculation of a rate ratio. However, there might be reasons why this under-reporting rate could differ according to the drug. For example, if an antibiotic causes a visible rash, because the doctor is likely to see it, it has an increased chance of being reported as an ADR. In comparison, if a blood pressure medication causes fatigue, and the patient doesn’t even mention it, it is less likely to get reported. And indeed, by my calculations, skin-related ADRs constitute about 32% of all ADRs with antibiotics, compared to only 9% with non-antibiotics. Such a likely reporting bias would contribute towards over-estimation of my “pseudo”-rate ratio. Yet even so, this is not nearly enough to explain away a 5-fold difference, even if we were to make the absurdly conservative assumption that 100% of visible skin issues get reported. In any case, I use the term “pseudo” to acknowledge that this isn’t a reliable estimate.
Note that antibiotics are generally prescribed for short periods of time. Non-antibiotics are generally taken continuously. My prescription usage data was based on count of those who used a drug within the last 30 days. This means an antibiotic could have been taken for say only 1 to 10 of those 30 days, yet still have been counted as much as a drug that was taken the entire month. This is a factor that potentially makes my pseudo-rate ratio a very conservative estimate, since antibiotics have far fewer total doses with which to yield ADRs, as compared to non-antibiotics. This point cannot be overstated. Also, short-term use of drugs may give users less time to gain confidence that they are experiencing an adverse event, and hence they may be less likely to report it to their doctors as compared to drugs used long-term. Doctors may also have less opportunity to learn how to identify adverse events with short-term drug use. Short-term drug use also decreases the chance of a concurrent drug precipitating a reaction, thereby misleadingly associating the short-term drug with that reaction as well. All of these factors of treatment duration contribute to underestimation of antibiotic adverse drug reactions.
Also, recall that for 4 out of 7 antibiotics I only had prescription drug usage data for their entire classes grouped together. I opted to use those class numbers as upper-bound estimates of the true usage of these 4 individual antibiotics. It is likely that these antibiotics represented that vast majority of total use in their respective classes. In any case, to the extent that this bounded data biases my calculations, it again makes my pseudo-rate ratio a conservative estimate.
Another consideration is that the prescription usage data all provided 95% confidence intervals. I used the reported point estimates. For example, the point estimate and 95% CI for percentage adult users for all antibiotics grouped together in one study we cited was 4.2(3.7-4.9) with p<0.001. Worst case, this uncertainty still can’t explain the 5-fold difference in ADRs. Also, the two studies I sourced usage data from had slight discrepancy in estimated antibiotic usage (despite both using the same data from NHANES) because one looked at children and adults, and the other looked at only adults. However, this could not have affected my estimates much.
Also, an issue is that VigiAccess does not let you filter by just the USA. The smallest region I could choose was the “Americas”. It is not clear how much, or in which direction, this biased the calculation.
One last consideration is that perhaps the 10 non-antibiotics I selected was not a large enough sample. Future investigation should pool a larger number of drugs.
Although no conclusions can be made, when you look at the various caveats and biases that would either increase or decrease the estimated rate ratio, it is difficult to dismiss an estimated 5-fold difference in ADR incidence. In fact, overall, it seems this could very well be a severe underestimate.
Grouping Related ADRs Together Yields 25 to 50-Fold Differences in ADR Rates
With spontaneous reporting systems the same or similar reactions could have dozens of different names. Individually, they may all look like they don’t count for much, but if we call them all the same thing, some numbers could be really big. I noticed that skin-type ADRs had a variety of names for simply rash, itching, and hives. I also noticed that a number of reactions of a potentially severe nature seemed to happen disproportionately with antibiotics. I searched the entire list of roughly 3,400 ADRs and created two ADR groups that merged the counts of the following ADRs:
- Skin-type ADRs – Rash, pruritus, urticaria, rash macro-papular, rash erythematous, erythema, rash puritic, rash macular, erythema multiforme, rash papular, pruritus generalized, rash macular, erythema nodosum, rash morbilliform, generalised erythema, rash vesicular, drug eruption, purpura.
- Severe-type ADRs – hypersensitivity, drug hypersensitivity, anaphylactic reaction, anaphylactic shock, anaphylactoid reaction, serum sickness, serum-like sickness.
For each drug, surrogate counts for each specific type of ADR, for just the year 2011, for just the USA, were calculated, just as before. Then, as was done in Table 2, I calculated pseudo-rate ratios for these merged ADRs. The results are summarized below in Table 3:
|Pseudo-Rate Ratios (P-RR)|
|Extreme-case P-RR, Skin-Typec||2.11|
|Extreme-case P-RR, Severe-Typec||2.24|
The grouped skin-type ADRs occur with an estimated 48-fold incidence in the average antibiotic, relative to the average non-antibiotic. And with the grouped severe-type ADRs, there is a 27-fold incidence for antibiotics, relative to non-antibiotics. For all the same caveats and biases I already discussed, these numbers are also likely to be underestimates. Also, note that the probable reporting bias I described relating to rashes and other visible skin signs does not apply here, as such a bias would affect both drug groups equally. Overall, these are very large ratios that make a case for a causal relationship.
I again calculated extreme-case pseudo-rate ratios, which indicated that – with respect to these specific ADR subgroups – even the apparently least dangerous antibiotic still has over twice the ADR rates as the apparently most dangerous non-antibiotic (again this turned out to be doxycycline and omeprazole). Hence, because these types of ADRs tend to occur disproportionately with all the antibiotics I examined, and because the ratios are too large to be explained by reporting biases and caveats, the only logical remaining explanation is that the functional anti-infective similarities of the drugs must be involved. Hence I posit the presence of undiagnosed chronic infection.
Another ADR group that would be worth investigating is anything related to fever. These could be under names like pyrexia, hyperpyrexia, hypersensitivity, serum-like sickness, influenza, influenza-like illness, and I don’t know how many other names. Also, chills might be another ADR type to try and group together.
It would also be worth investigating the correlation between having a member of one ADR group appear simultaneously with a member of another ADR group within the same report. For example, do pyrexia and rash tend to appear together, etc?
ADRs for Dietary Supplements Also Raise Suspicion of Infection
Just take a look at a selection of the ADRs that ascorbic acid turns up through VigiAccess:
It’s the same exact story. The same kinds of ADRs we see with antibiotics are topping the list and constituting a large percentage of the overall total. And this time I haven’t even made an effort to pool similar types of ADRs together. Though one thing to note is that the total number of ADRs (5,404) has a caveat, so my computed proportion can’t be taken at face value. Every material in the VigiAccess database lists a total number of ADRs at the top. But it you sum up all the individual counts of ADRs for a drug, you will always get a total that is larger than the total at the top. In the case of ascorbic acid, you get a total of about 10,000. That’s a roughly 2-fold difference. For the drugs I’ve examined, this factor was pretty much always between 1.5 and 2.0. Correct me if I am wrong, but I am presuming this discrepancy factor is because a single ADR that a doctor submits could have multiple symptoms on it. The total at the top may be the total number of submissions, but the average submission will list more than 1 type of reaction. And in case some are wondering, duplication bias is something completely different from this, and doesn’t explain this discrepancy factor.
Now some may be thinking, “How can a little old vitamin C supplement lead to such hypothetical Jarisch-Herxheimer reactions? Do you expect me to believe that?” Well, what other explanation is there? Vitamin C must either be directly anti-infective, or else stimulate the immune system to fight off a chronic infection. Vitamin C has been reported to treat a variety of infectious diseases. Research in this regard is summarized in my article on measles vaccination. One thing to note is that despite popular belief, vitamin C deficiency or depletion is common, with incidences ranging from 5% to 25% even in the United States and the UK, and as high as 74% in a country like India.6–8
Some might say that people take vitamin C when they have colds and flu, therefore the colds and flu cause fever and chills, and not the vitamin C. In other words, maybe the ADRs are explained by the context of prescription/usage, rather than by the properties of the anti-infective. Well, if people are smart enough to know they have a virus coming on – and hence start vitamin C – they would be smart enough to know that fever and chills are expected symptoms, making it unlikely that they (much less a doctor) would misattribute them to vitamin C side effects. Also, rash, itching, and hives do not commonly occur with colds and flu. Similarly, I don’t think the context of prescription can explain my findings with prescription antibiotics.
Searching through other dietary supplements might turn up similar results to vitamin C, though data is often sparse. A casual look at ADRs for both vitamin D3 and iodine made me suspect pooling skin-type ADRs would yield the same kind of finding. These are both potentially anti-infective or immune modulating. Mega-doses of iodine use to be used as an antibiotic. Also, sodium bicarbonate (baking soda) turns up a high proportion for fevers and chills. Baking soda is antifungal and is claimed to effectively treat Candida infection. Interestingly, glucose also turns up a fair amount of fever and chills. I have argued elsewhere that this could be due to glucose exacerbating undiagnosed Candida infection.
Here is another thought. Though concerns over the safety of dietary supplements are legitimate, they have often been exaggerated. And here is another reason why that may be. I presently wish to suggest that many adverse reactions to such supplements may actually be JHRs. In other words, the materials themselves may not be toxic, but may be rather killing a chronic infection. Perhaps instead of only looking at such reports as evidence of possible harm, we should also be using such data to screen for potentially effective anti-infective therapies.
FAERS Data Looks Similar
Because using VigiAccess required the unfortunate introduction of surrogate data, I sent an FOI request to the FDA (FDA Adverse Events Reporting System9) for the drugs in question, for the year 2011, for only the USA. I repeated the same calculations. This yielded a pseudo-rate ratio of about 3.2, which is smaller than my previous estimate of 5.14, but still not easy to explain away. For merged skin-type ADRs and merged severe-type ADRs, the pseudo-rate ratios where both around 12, which is still quite large.
I found in the FDA data that skin-type ADRs and severe-type ADRs were much more prevalent with amoxicillin as compared to other antibiotics. With the VigiAccess data, amoxicillin was more or less an average antibiotic in these regards. Perhaps this might be explained by greater attention to hypothetical penicillin allergy in the USA. I suggest that perhaps adult penicillin allergy could sometimes be a misdiagnosis, and that such symptoms might sometimes be better explained by my hypothesis of a Jarisch-Herxheimer reaction.
Although using the FAERS data arguably makes more sense than using VigiAccess (owing to not needing to introduce surrogate data), it cannot be automatically assumed to be so. For example, even while FAERS contained a higher total ADR count than the surrogate data derived from VigiAccess, the merged skin-type and severe-type ADR groups represented a much smaller proportion of the total ADRs in FAERS, as compared to in VigiAccess. This shows the difficulty in making comparisons.
The FDA erroneously omitted data for azithromycin from my request. So these rough calculations were based on 6 antibiotics, with the 7th antibiotic imputed. This was done by assuming azithromycin’s proportion of the total ADRs in FAERS was the same as in VigiAccess. In any case, this couldn’t dramatically affect the calculation.
Antibiotic Side Effects – Closing Discussion
I wish to note that since antibiotics are generally used short term, the types of side effects they yield will obviously tend to have a more rapid onset compared to non-antibiotics. And because JHRs would be expected to tend to have more rapid onsets than other types of side effects, Bayesian probability would suggest this in and of itself increases the probability that the average antibiotic ADR is actually a JHR. This makes me wonder about things like “rapid onset” neuropathy with fluoroquinolones.
I also wonder about all the people who say they’ve been “floxed” by fluoroquinolones. Indeed, I found ciprofloxacin to have the highest ADR incidence of all antibiotics I analyzed. Could undiagnosed infection be much of the real explanation for floxing? Or is it just that people are more likely to report ciprofloxacin side effects due to greater public awareness?
If someone is infected, a specific antibiotic may not have activity against that infection. And even if it does, it may still not precipitate a JHR. Hence, we could just be seeing the tip of the iceberg in terms of indicating the presence of infection.
In any case, I think this analysis is highly suggestive of the presence of undiagnosed chronic infection as an explanation for potentially the majority of antibiotic side effects. I hope some researchers will take interest in this hypothesis and investigate it more formally. The same kind of investigation should also be done with antivirals, antifungals, and antiparasitics.