Oxford – Recovery : are the data hiding the deaths?

  •  SOUTENEZ L'INDEPENDANCE DE FRANCESOIR, FAITES UN DON !  

Oxford – Recovery : are the data hiding the deaths?

Publié le 07/08/2020 à 12:07 - Mise à jour à 17:17
Pexels
PARTAGER :

Auteur(s): Valère Lounnas and Xavier Azalbert for FranceSoir

-A +A

ANALYSIS: Following various publications, the analysis of the data from Recovery, the British self-called flagship clinical trial, has led to the identification of critical inconsistencies. Hereafter we present these analyses in detail with two different approaches:

  • An analysis of the data published at key dates with many inconsistencies raising important questions,
  • A retro-engineering analysis of Recovery's mortality curves further reveals inconsistencies in the published data.

 

The simple analysis of published data, supplemented by that resulting from exchanges between the stakeholder, is more than worrying. It would merit a genuine comprehensive independent audit as the observed consequences are critical and raise questions about the management of this study..

Retro-engineering was used to support the analysis of data inconsistency. In addition, we found it insightful to calibrate the results of Recovery on the French AP-HP study. Using the French data, we can correctly estimate the standard of care and dexamethasone arms. With regards to the hydroxychloroquine arm, numerous inconsistencies appeared that would demonstrate excess mortality in the first few days. They could correspond to the presumed overdosed patients (2400 mg of Plaquenil on the first day of treatment and more than 9000 mg over 10 days).

 

Intentional or not, these are dead patients. This study calls for a genuine independent audit :  as it stands, the pre-publication of Recovery cannot be approved due to the too many inconsistencies.

Foreword

FranceSoir has already published two articles on hydoxychloroquine (HCQ) overdosing during the first day of treatment of patients with Covid-19 acute symptoms in the HCQ arm of the Recovery trial (1, 2).  In that trial, a loading dose of 2.4 g was administered at day 1,  in 4 separate intakes spread over 24h, followed by 800 mg/day the next 9 days. These doses are to be compared with the more reasonable ones administered in the hospitals of Assistance Publique de Paris (600 mg at day 1 followed by 400 mg the next 9 days) (3). The interested reader can access these two articles wherein we strongly affirmed that aged and hospitalized patients, weakened by the Covid-19 disease, were exposed to potentially lethal overdoses of HCQ.

 

The Recovery trial was designed and led by two co-principal investigators, Pr Martin Landray and Pr Peter Horby,  who are both distinguished scholars from the University of Oxford. On June 6th, FranceSoir has interviewed Martin Landray who did not seem aware of the dose limiting toxicity (DLT) of HCQ (4).

 

We present here a thorough retro-engineering analysis of the preliminary results of the Recovery HCQ arm and its standard of care comparator (SOC). Our work is entirely based on the trial's pre-print document deposited on the MedRxiv internet archive site of the University of Cornell (5).

 

We regret that the preprint server of the University of Cornell is highjacked during a pandemic crisis for reasons that are more politically involved than scientifically grounded : the  use of hydroxychloroquine against the Covid-19 disease in the clinical setting remaining highly controversial despite a broad corpus of mounting evidences of its efficacy at least at an early stage of the infection and as prophylaxis (6, 7, 8, 9, 10) and probably also at a later stage when the inflammation reaction is important (11).

 

Analysis I - Overall the data in the trial present numerous inconsistencies of the most basis level

The table below presents, mimicking a giant Sudoku, the data from Recovery issued from their fragmented and incomplete publications. With the assistance of our British correspondents, we have obtained some exchanges under the Freedom of Information Act (FOI) existing in the United Kingdom.  They provide information on exchanges between various parties (Recovery Investigators, Data Monitoring Committee, MHRA). So, we analyzed these various exchanges and, as for a giant Sudoku, reconstructed the data table for various critical dates.

 

Several important dates:

  • May 22nd, 2020:  Date of the Lancet publication of the Mehra's study. A retracted study, that triggered numerous exchanges between the Data Monitoring Committee (DMC), a mandatory regulatory body set up for any trial in the UK), the Recovery team and the MHRA (equivalent to the French ANSM).
  • May 26:  Response of Recovery's Investigators to the DMC.
  • June 4:  The DMC asks Recovery to conduct a complete data analysis by unblinding the data. It is also the date of data analysis used for the preprinted published in MedRxiv in July.
  • June 5:  Recovery's piecemeal press release indicating that the HCQ arm is stopped because there is no effect and indicates a mortality rate.
  • June 8:  Recovery press release announcing the success of dexamethasone.
  • June 30:  Recovery video conference on the hydroxychloroquine arm (HCQ).

We reconstructed the table from all the documents provided, including some elements of language that highlight inconsistencies between the hydroxychloroquine (HCQ) and standard of care (SOC) arms). The main observations are summarized hereafter.

  • On May 22, the DMC asked Recovery investigators if they could confirm the HR-1.335 ratio (the ratio between the percentage of deaths between the HCQ and SOC arms of the Mehra et al. study) in the HCQ arm of the Recovery study.
     
    The response from Recovery's lead investigators by email speaks for itself: we recommend continuing inclusion because the DMC is satisfied that recovery's HR (risk ratio) is significantly different from the ratio of the Mehra al study. Surprising, but the exact figure of this ratio is nowhere to be found in the documents.
    Surely blinded by the HR of the Mehra study, the mortality rates in these two arms were not asked for by the DMC nor was it voluntarily disclosed by the investigators of Recovery. These proportions are known beyond any doubt as they are used to calculate the HR.  We will see later why this is so important. Moreover, it is totally implausible that this key data was not officially exchanged by mail since this is the information used for the decision is made... Would it be possible that the two parties have tacitly agreed that these figures should not be disclosed?
  • On June 4, 2020, the day of the retraction of the Mehra study, the MHRA requested that the confidentiality of the data should be lifted, and that the full analysis be done. The reason cannot be found anywhere.
  • On June 5, 2020, the lead investigators announced the discontinuation of Recovery's HCQ arm on the grounds that the drug showed no efficacy and that the mortality rate was identical to that of the standard care arm. It's over for hydroxychloroquine.
  • On June 8, 2020, Recovery announced that dexamethasone has a significant impact on the mortality rate HR = 0.89 versus the standard care arm (Column 8 June in the final document in appendix).

 

Inconsistencies in both the data and the analysis:

  • As of June 4, the death toll in the HCQ and SOC arms is different. The initial release, on June 5, gives only the ratios of 25.7% for HCQ and 23.5% for SOC. It allows to calculate the deaths at 396 and 736 for each of the respective arms. On June 30, 2020, in a conference, these figures amounted to 411 and 766 respectively, while in the preprint they are 418 and 768. All these different figures for the same data cut date of June 4th.
  • However, in the preprint, the Kaplan-Meier curve reports 436 deaths in the HCQ arm and 809 in the SOC arm. A notable inconsistency within the same document.
  • In reconstructing the data for May 22, we found in exchanges with the DMC that there were 1,026 deaths for 4515 randomized respectively in the HCQ arm (1515 patients) and SOC (3060 patients). But nowhere can we find the details of the deaths between HCQ and SOC arms!
  • Between May 22 and June 4, the inclusion rate in the trial increased by 4.84%, the rate of patients included in the HCQ arm is 3.08%. Incoherent, therefore, because the DMC had asked to continue the inclusion of the party members because there was no concern.
  • Between 22 May and 4 June, the number of deaths in the HCQ arm rose from 1026 to either: 1132 or 1177 or 1206 or 1245 depending on the data source used (and again this at the same data cut date June 4th)! The number of deaths in this arm increased by 21.35% between these two dates while inclusion increased by only 3%. Hard to understand even if considering that the death can come from the preceding days.
  • Another point that catches our attention is the similarity between the mortality rate of the SOC arm at June 4th (25%) compared to that of the SOC arm in the dexamethasone part of the study (25.7%).
  • So, these too many inconsistencies can only cast doubt on the data. And the exact death toll in the HCQ arm is still missing as of May 22.
  • The Kaplan-Meier curve published in MedRxiv will provide the answer. By calculating the death toll after 13 days using a triangulation calculation, we find 659 deaths in the SOC arm on the 13th day (the difference between May 22nd and June 4th). Subtracted from to the 1026 total deaths (data uncovered in the FOI obtained correspondences), we therefore have 367 deaths in the HCQ arm. Mortality ratios would therefore be 24.22% (HCQ) and 21.54% (SOC) with an HR = 1.12.  (Precision, the computation was made on the SOC arm due to its higher stability, computation may differ by 1 or 2 units depending on the triangulation method used. It is also adjusted for the inclusion data provided in the Recovery presentations.

 

Significant differences and problems

  • This HR =1.12 ratio of is statistically different from the ratio of the Mehra study under a bilateral test with p = 0.01. Were Recovery and the DMC right to continue this clinical trial ? Such a ratio should have worried both the Investigators and the DMC.
  • However, the Mehra study was flawed, and between May 22nd and June 4th no one asked what was going on or sought further information withing Recovery. Should we think.
  • The DMC and MHRA, probably blinded by the HR = 1,335 announced by the Mehra study, did not look any further. The fact that the ratio HR = 1.12 was different (from the 1.335) was enough to carry on, but that may not be the only reason as there is also a critical political factor to continue with this test.
  • Here, a key question arises: Is there a difference in effect between the HCQ and the SOC arms on May 22nd? This difference should have attracted the attention of the experts. These figures are necessarily known to both parties since they are needed for the ratio to be calculated.
  • By doing the statistical test between the HCQ mortality rate (24.22%) and the SOC mortality rate (21.54%), we find a statistically significant difference (IC 95% p = 0.02).
  • On May 22nd, the Chief Investigators of Recovery knew if there was a significant difference between the HCQ arm and the SOC arm. This information did not reach the DMC. Nor did the DMC ask in the communication exchanges. Any biostatistics expert would have asked considering the 1.12 HR ratio.

 

In addition, the mortality analysis (436 deaths or 27.9% in the HCQ arm and 809 or 25.6% in the SOC arm) from the Kaplan-Meier data gives an HR of 1.09, however this is not what the graphical curve represents. The HR = 1.09 provides a statistically significant difference with a unilateral statistical test at 95% with p = 0.0481. This further evidence implies that the seriousness of the results can only be questioned.

 

Therefore, the inclusion of patients in the HCQ arm continued while the study was flawed from the outset and mother prudence would have called to verify the data intrinsic to the study. So, with initial incompetence on the dosage and apparent willingness to hide information afterwards appear to have been the ins and outs of this study, it will most likely remain in the annals of a “what not to do trial”. Only a thorough audit could confirm this. Nevertheless, this information combined with the overdose show that on May 22nd the apparent excess mortality in the HCQ arm should have alerted the Chief Investigators. However, highlighting this information would have triggered a DMC and MHRA investigation as the HCQ would have been shown to be more harmful than SOC. Hence raising concerns and bringing the attention most likely on the excessive dosage used.

In addition, we draw the attention to the fact that on May 23rd, Recovery's investigators wrote that they were waiting for a new study on HCQ pharmacokinetics:

Responding to an earlier request from MHRA, we will soon introduce an amendment to the protocol for authorizing a subanalysis in RECOVERY to study the pharmacokinetics of hydroxychloroquine in more detail, and also to study the pharmacogenomics of hydroxychloroquine and azithromycin on cardiac arrhythmias. The aim is to identify risk factors for drug-induced arrhythmias and fever.

 

We maintain our accusations that the Chief Investigators withheld significant information.

 

Analysis II Retro engineering, Mortality data in the preprint document are incomplete and inconsistent

We note that concerning the toxicity of the HCQ arm during the first 48 hours of treatment administration, the preprint document is elusive and of exaggeratedly little informative value. This is despite questions raised by press articles on the possibility of too high a dosage of HCQ during the first day of treatment (FranceSoir) and other journals.

The Recovery authors response is : “We did not observe excess mortality in the first 2 days of treatment with hydroxychloroquine, the time when early effects of dose-dependent toxicity might be expected.”  They also indicate that no cardiac toxicity was reported from patient files at the time of the preliminary results were deposited in MedRxiv.

 

Artistic blur shrouds the reasons why a reduced number of patients have indeed received the treatment in the HCQ arm? Indeed, the text mentions that only 1395 pts have received HCQ in the 1516 patients in intention to treat, among 1561 patients randomized in the HCQ arm, who were therefore supposed to receive the treatment !... The sole quantitative information that may be potentially indicative of a toxicity related to treatment overdosing is the treatment inter quartile range (IQR) from 3 to 10 days. This means that a quarter of the patients who have received at least one dose of treatment, that is to say 349 pts, have stopped treatment within the first 3 days ! The only causes of treatment stops are theoretically : toxicity, patient refusal or the patient death, or possibly its  hospital discharge! The latter reason seems little probable, in this instance, since the median time to hospital discharge reported in the Assistance Publique – Hôpitaux de Paris (AP-HP) study (3) was 9 days for patients receiving HCQ. No treatment stop for toxicity, nor for patient refusal, is mentioned in the Recovery preprint document deposited on MedRxiv.

 

Quantitative information on the day-by-day number of deaths can be inferred from the mortality curves. Fig. 1 below indicates the numbers of patients still alive at days: 7, 14, 21 and 28 following randomization, concomitantly to treatment beginning. Other days mortality can be graphically deduced directly from the curves (Fig. 1) quite reliably with a square ruler.

 

 

 

 

The values are collected in the following table. At first sight, they reveal several inconsistencies. First of all, the data of the numbers of patients at risk (i.e. patients still alive at a given day D following randomization) indicate 436 deaths at D28 in the HCQ arm, versus 809 in SOC arm, whereas the text mentions 418 deaths with HCQ and 788 with SOC, at 28 days, and referring to 1561 and 3155 as the total numbers of patients randomized in each arm ! Secondly, the curves graphical analysis shows numbers of deaths at 7, 14, 21 and 28 days diverging considering those that can be inferred directly from the numbers of patients at risk. Obviously, the Recovery curves (Fig. 1) is a picture montage where indication of patients at risk are elements borrowed from another version of the Recovery mortality plot. It would be interesting to know the plot corresponding this these numbers of patients at risk. Deceased patients have been removed from the curves in both arms. It would be interesting to know why?

 

From the mortality curves data (Table below) we can deduce that the median time to death and first quartile in the HCQ arm are 6.6 IQR [3.7 – 12.71 ] days and 6.9 IQR [3.7 – 12.47] days in SOC. Therefore, no difference can be deduced from the curves about the first quartile Q1 = 3.7 days.  If thi information was exact, why not giving them quantitatively instead of providing them in a very cryptic way, whereas it is a question of real importance? Was it to avoid discussing their implications?  Figures from the AP-HP study on the median time to death are 8.52 IQR[4.52 – 14.26] days for HCQ and 7.54 IQR[3.94 – 13.0] for the standard of care (SOC)... Therefore, figures from Recovery are consistent with those from AP-HP regarding SOC but fundamentally differ from AP-HP data regarding HCQ, with a shortening of the median time to death and of the first quartile Q1 by 2 days and 1 day, respectively!

 

 

Mortality curves retro-engineering reveals a mathematical inconsistency for the HCQ arm

 

To check the  consistency of the mortality  curves, we have retro-engineered them using an ad-hoc mathematical model described in appendices 1 and 2. The probability p(j) for a given patient to die at day j is a function of the probability that he/she dies the day before p(j – 1) weighed by a factor determining the probability evolution that varies from day to day in a complex manner. The purpose of the model was to reconstruct the probability evolution with relatively few adjusted parameters. Analysis of the deceased patients in the SOC arm shows that during the first two days the disease worsens considerably before reaching a peak the 3rd and 4th days.  This forces a form factor controlled by the parameters f1 and f2 with estimates of 0.2 and 0.8 for the SOC arm.  These parameters determine the delay to disease worsening at D1 and D2, respectively.  Thereafter, the odds to die statistically decrease day after day. We found suitable to select an exponential decay controlled by 4 parameters (A , Amax., B et α) to model this process. Their values are comprised in intervals (1. ≤ A ≤ 1.2; 2. ≤ Amax. ≤ 5.5; 14. ≤ B ≤ 37.; 1.2 ≤ α ≤ 1.8) allowing the reproduction of a large variety of curves.

 

Thus, if we know Pvalues, the initial probabilities for patients to die at day j=0 in the HCQ arm and the SOC arm, we can therefore deduce probabilities of the following days for each arm. Every day, the number of patients dying in each arm is determined according to a Monte Carlo process, insuring natural fluctuation of mortality. A dead patient is immediately removed from the pool of alive patients. Starting probabilities are inferred from the Recovery curves (cf. Appendices 1 and 2).

 

Appendices 1 and 2 give a detailed description of the mathematical model so that those who would like to check its validity or criticize it may do it. Starting < P0 >  values are calculated from patients at risk between D1 and D7. Optimal parameters for the SOC curve can be established manually by tries and errors and visual appreciation of the fit (SOC Fig. 3 with parameters of Appendix 2) or by computer power, testing systematically all parameters combinations as for SOC (Fig.2) and also HCQ (Fig. 3).  Both approaches provided an excellent fit of the SOC curve from D1 to D28 although the fit obtained by systematic search (Fig. 2) is optimal. Computerized search gives parameters that can differ largely from manually inferred values meaning that curves are not uniquely described by one single set of parameters.

HCQ mortality seems mathematically differing from SOC mortality

 

Although the model does not account for the whole reality of mortality curves, it allows useful deductions. Despite its fitting flexibility, the model does not allow us to match the HCQ mortality curve entirely correctly from D1 to D28.

 

Logically, based on the conclusions from the authors of Recovery that HCQ is neither toxic nor active against Covid-19, the parameters that describe the mortality curve of HCQ should only differ marginally from those of the SOC arm. We therefore kept the disease worsening parameters at D1 and D2 (f1 = 0.2 and f2 = 0.8 describing a weak worsening) identical to those of SOC and conducted an exhaustive search to adjust the 4 other parameters, by testing systematically more than 6 million possible combinations. However, we could not completely fit the CQ curve, from D1 to D28, as for SOC. Between D6 and D9 a non-negligible  difference appears between the best  modeled curve and the curve of Recovery (Fig. 3) !

Widening the window search  for the parameters B (10. < B < 14.) and α ( 1.8 < α  < 2.2) that should help control the level of the curve in the shoulder zone D5-D10 did not allow a better fitting, leading to quasi identical curves with the same depletion between D6 and D9. 

 

The mean mortality between D1 and D7 of the HCQ curve of Recovery (< P0 > = 0.0239) is 15% higher than that of the SOC arm (< P0 > = 0.0210). This led us to envisage the possibility that a difference exists in the worsening parameters f1 and f2. We have let these parameters vary (0.2 < f1 < 0.7 and 0.6 <  f2 < 0.9) together with the 4 other parameters, generating 108 million possible combinations, unsuccessfully.

 

Therefore, we decided to fit the curve only from D4 to D28 and let differences in mortality between  D1 and D3 not influence the fitting.  This procedure allowed f1 and f2 to take values uniquely considering their influence on mortality from D4 onward, by acting on the number of patients at risk available beyond D3. This as allowed us to obtain an improved fit in the region D6 to D9 but has created an upward shift, from D1 to D4, due to an excess of 15 deaths at D1 (23 instead of 8) with repercussion to  D2, D3 and D4 (Fig. 4). The corresponding parameters are: A = 1.09 , Amax. = 5.40, B = 14.5,  α = 1.53, f1 = 0.45 et f2 = 0.6.

Despite its sophistication, our model can reconstruct quite accurately the SOC mortality of Recovery but not of HCQ. We must conclude that the nature of the HCQ curve is mathematically different from SOC.

We point out that the mean probability of death <P0 > = 0,0239 for the HCQ arm (calculated from the Recovery data) is too high to allow a simple upward shift of 2% of the  curve, in particular in the region D1-3, without implying a marked modification of the shape of the curve.

It is not forbidden to hypothesized that two opposite effects, toxicity and activity, may be superimposed in the same curve, preventing its modeling with the same mathematical function as  for SOC.

We also note that the test for equality  of proportions, applied to the numbers of deaths at D1 between the HCQ model arm and the SOC arm of Recovery, indicates that an excess of 15 deaths at D1 would correspond to a significant difference of mortality compared to SOC.

Time to death IQR parameters of AP-HP indicate a potentially significant efficacy of HCQ

 

To ascertain the consistency of our model we have calculated the mortality curve of the HCQ arm with the probability parameters that can be deduced  from the time to death IQR data of the  retrospective study of Hôpitaux de l'Assistance Publique de Paris (AP-HP) published as well in MedRxiv. We have adjusted the probability parameter P0 in order to obtain an initial probability for HCQ consistent with the level of the SOC arm in Recovery (see Appendix 2). Other parameters were adjusted to obtain a median time to death of 8.4 days and a Q1 time of 4.5 days, getting closer as much as possible to the AP-HP values, while keeping the mortality ratio in a compatible range of 4% difference (as reported by AP-HP).  Importantly, we had to modify f1 and f2 to lower values, 0.11 and 0.44, respectively. We have thus obtained without difficulty the following curves while keeping parameters of SOC unchanged.

 

At 28 days the rate of mortality the simulated HCQ curve is 21.0% whereas it is 25.%  for SOC, which gives a ratio = 1,19 in good agreement with the 28-day mortality ratio of AP-HP between HCQ and SOC (1.23 = 21.9%/17.8%). The fact that the worsening parameters were reduced by 45% is consistent with a protective effect of HCQ at D1 and D2.

 

Reconstruction of Recovery mortality curves for dexamethasone and SOC

The model has allowed us to reconstruct accurately survival curves of Recovery for the  dexamethasone arm and the SOC arm with 1166 additional patients (Fig. 6a). The initial mean probability, calculated from the patients at risk (between D1 and D7) for SOC has slightly changed and became  <P0> = 0,0218, adjusted to P0 = 0,0246 for the model, other parameters remaining unchanged. Probability parameters for dexamethasone, calculated the same way, are:  <P0> = 0,0151, adjusted to P0 = 0,0172 (Fig. 6b) , with other parameters manually adjusted to : α = 1.68, B = 24., A = 1.15, Amax. = 2.65, f1 = 0,2 and f2 = 0,8 to best fit patients at risk at days 7, 14, 21, 28. 

Discussion

Certainly, this ad-hoc modeling study cannot account for the physical reality underlying the mortality curves. However, it can help provide valuable insight because it does capture a true aspect of reality. The starting probabilities are derived from true data. For instance, it allows us to bring to the fore a mathematical inconsistency of HCQ with respect to SOC, contradicting the affirmation that the treatment may be non-active and non-toxic and thus does not differ from SOC. This modeling problem may indicate that there is a real medically involved difference in the time evolution of the HCQ and SOC arms. Especially that we do obtain excellent results with dexamethasone and the AP-HP parameters injected in the model.  Automatically fitting the curves resulted in clearly distinct values of the parameter B between HCQ (14.92 days) and SOC (18.14 days), which can be seen as an intrinsic time relaxation parameter characterizing the decay of the probability of dying over the days. However, as we have shown, parameters sets are not uniquely defining curves and one should remain cautious about unrealistic interpretation of this parameter. However, this strengthens  our belief that the HCQ curve is a composite curve combining limiting toxicity and curating benefit.

 

Contrary to SOC mortality (and also the dexamethasone mortality curve) it is not possible to produce a set of parameters that allow the HCQ curve to be reproduced entirely from D1 to D28. Regarding the Recovery data, the modeled curve exhibits a depletion of mortality between D6 and D9 that we could not alleviate with exhaustive search of the parameters space. The depletion could however be alleviated by acting on the disease worsening parameters at D1 and D2. The worsening parameters have been deduced empirically for SOC. By modifying the number of patients at risk at D1 and D2 they have a repercussion from D3 to D28. By optimizing curve fit between D4 and D28, the parameter f1 for D1 takes a value 225% greater than that of SOC (f1 = 0.45 instead of 0.2). This would correspond to an excess of 15 deaths at D1 consistent with HCQ overdosing.

 

Furthermore, transposing the median time to death and IQR data from AP-HP to the model allows us to show the potential efficacy of HCQ, with a rate of mortality reduction of 18% identical to that reported by AP-HP in univariate analysis.  Regarding the HCQ arm of Recovery, this would correspond to a total reduction of 90 deaths, 85 of which in the first 7 days!  

 

When the authors of Recovery affirm that their trial shows that HCQ is neither toxic nor active it is about a language abuse, playing on the convention on the interpretation of the p-value translating the  statistical power of a trial. The significance threshold is arbitrarily set to 0.05 which means 5 % of chances to mistakenly conclude that the two mortality curves are different. In Recovery, the reported p = 0.18, between the HCQ and SOC arms, tells us objectively that there are 18% of chances to be wrong (and thus 82% of chances to be right) in affirming that HCQ has resulted in more deaths than SOC. All that is dreadful for the two co-principal investigators of Recovery who seem not to realize that the 5% significance threshold is a subjective convention that leads to schizophrenia. Quoting their own words we find : “Patients allocated to hydroxychloroquine were less likely to be discharged from hospital alive within 28 days (60.3% vs. 62.8%; rate ratio 0.92; 95% CI 42 0.85-0.99)”, ...  and some readers have their eyes wide open on the curves and the preprint lack of quantitative data.  

 

We think that Recovery authors must be more transparent regarding the fact that 268 patients (349 – 81 from accumulated deaths at D3) in the HCQ arm have stopped treatment between D1 and D3! What are the reasons for these treatment stops that are certainly not hospital discharge? There is no reason to think that the reported interquartile of treatment duration, IQR = [3 to 10] days cannot be applied as well to the deceased patients, meaning 104 deaths at 3 days and not 77 as indicated by the mortality curve of HCQ. It would have been a more directly relevant information. The fact that 272 patients ave stopped treatment within 3 days from treatment randomization is a strong indication, if not an evidence, of an excess toxicity.

 

The authors of Recovery claim that there was no HCQ overdosing in Recovery, taking as an evidence the  comparison with the  Brazilian trial of Borba et al. (12) where chloroquine was administered as a base, a form of the molecule which would correspond according to them, to a substantially higher dose of HCQ as a salt (sulfate). In fact, 800 mg of HCA sulfate correspond to 620 mg of HCQ base. Therefore, in the Brazilian trial it is clear that the equivalent of 1.55 g of HCQ sulfate was administered every day (including the first day) rapidly leading to death by cardiac failure, and subsequently to a law suit. But, a higher dose of 2.4 g was administered the first day in Recovery!  The accumulated dose in the first and second days overcomes that of the Brazilian trial, with 3.2 g of HCQ against the equivalent dose of 3,1 g. At D3 the accumulated dose is thus 4 g in Recovery against 4.5 g in the Brazilian trial. In fact, this is a very similar dose in the first 3 days !  The following days, the HCQ dose was indeed substantially lower and may have resulted in a lethal toxicity, somehow going unnoticed because taking another form than sudden arrhythmia and cardiac failure (see Conclusion statement) and masking the potential benefit of HCQ. 

 

Conclusion

We continue to think that overdosing of hydroxychloroquine in Recovery has certainly led to an excess toxicity in the first 3 days of treatment, with many treatment stops and a number of toxic deaths still to be elucidated. Identifying them may be a complicated task since HCQ overdosing result in acute respiratory failure just like Covid-19! The authors of Recovery are very elusive about this reality. They skillfully admit that : “...those (patients) not on invasive mechanical ventilation at baseline were more likely to reach the composite endpoint of invasive mechanical ventilation or death (29.8% vs. 26.5%; risk 44 ratio 1.12; 95% CI 1.01-1.25)” and add : There was no excess of new major cardiac arrhythmia.  How can they know it since they specify further in the text: We did not collect information on physiological, electrocardiographic, laboratory or virology parameters.”?

 

Finally, we believe our retro-engineering study of Recovery demonstrates that, unfortunately, the in-retrospect study of AP-HP, with only 365 patients having received HCQ, lacked power to demonstrate benefit in multivariate analysis, whereas irresponsible overdosing has led Recovery to failure although the HCQ arm was well dimensioned with 1561 patients.  

 


 

Appendix 1 : mathematics of the mortality curve retro-engineering

Appendix 2 : parameters used for mortality curves fitting

 

Complete table

 

Auteur(s): Valère Lounnas and Xavier Azalbert for FranceSoir


PARTAGER CET ARTICLE :


Oxford – Recovery : are the data hiding the deaths?

Annonces immobilières

Newsletter


Fil d'actualités Société




Commentaires

-