Looking forward and seeking a solution hypothesis to the present critical care research crisis highlighted by the recent REMAP CAP, what are the thoughts of adding symbolic causal modeling with DAGs or some other such symbolic method as a requirement for funding of an RCT?
On X I provided the evidence that comparison of DAGS would have revealed a disparate block of the arrow from immune suppression to increased bacterial load AUC (in the time domain) vs viral load AUC. (AUC here refers to the viral or bacterial load over time)
Past (and present) oversimplified analysis is that immunosuppression induced by treatment of community acquired pneumonia with corticosteroids would reduce injurious inflammation in both bacterial and viral pneumonia. Based on this simplified view researchers think they can lump both community acquired bacterial and viral under the synthetic syndrome of Community acquired pneumonia (CAP). It is this type of oversimplified thinking with no formal casual modeling (eg no DAGs) which has dominated in critical care clinical experimentation science. However these studies have not been reproducible for decades.
Here are some quotes from X. First from Grok to show the two edged scalpel of corticosteroids.
“Mechanism: Corticosteroids suppress the immune system by reducing inflammation and cytokine production, which can impair the body’s ability to clear viral infections like influenza. This suppression may prolong viral shedding, especially in the lungs.”
Now think about what you know about antivirals vs antibiotics. You may not know that Antivirals are quite unreliable but antibiotics are quite reliable for a broad range of community acquired viral and bacterial pulmonary pathogens, many of which are deadly.
Now draw two DAGs one for viral pneumonia and one for bacterial pneumonia.
Note the sequential arrows from corticosteroids (C) to immunosuppression (I) to increased viral AUC (V) to increased mortality (M) .
C → I → V → M
Note in the bacterial DAG that the analogous sequential arrows from corticosteroids to immunosuppression to increased bacterial AUC (B) to increased mortality is blocked by antibiotics at the
C-> I → [blocked] B-> M
This shows the pitfall of substituting a hypothesis for formal causal modeling. Obviously you can’t lump these pneumonia (as they did in REMAP CAP) in a disease agnostic RCT as the % mix of viral vs bacterial pneumonia would affect the ATE rendering multiple RCT non-reproducible.
The disparate effects of the antibiotics vs antivirals blocking vs not blocking respectively, the increased AUC of the pathogen count in the time domain is something that was overlooked.
Comparison of the two DAGS discloses this. This would have been warning enough to simply do the viral RCT separate from the bacterial RCT. Indeed the viruses will need to be separated as we learned to our sorrow that lumping severe COViD pneumonia into ARDS ventilator protocols (as had been done with influenza pneumonia) resulted in significant harm on a worldwide scale before the critical care docs at the bedside revolted. A deeper understanding of all of this might have prevented the idea that COViD pneumonia could just be lumped in. In other words the culture of blind lumping of different diseases by guessed threshold set triage would have been eliminated by self correction of science long ago.
Of course even these DAGs are oversimplified but they are perfectly illustrative. These DAGs are easy to do but they require a depth of knowledge of the potentially relevant physiology which brings the statistician and PI into a deeper discussion of constraining physiological paths.
Here a marked difference is easy to see with these DAGs but easily missed without them. The formalization of causal modeling, as opposed to the standard worded gestalt, is the process required to assure the public’s funds are well spent and potential harms which may be hidden by lumping are well defined.
We are so careful with formal processes to prevent “Swiss cheese” penetrating mistakes in the hospital. We should apply the same formal protections to the process of research design.
The synergy between formal symbolic causal modeling and RCT disclosed the subtle Langmuir apical error which had been missed for decades and which rendered those RCT as pathological experimentation science.
Now consider an RCT for corticosteroids for septic shock which is defined by a guessed threshold set and comprised lumping at a supermassive scale with hundreds of different infections treated with antibiotics, anti fungal and antivirals lumped by a guessed threshold set in standardized (mandated for grants) PettyBone sepsis RCTs.
The many DAGs which would have been required to determine if the lumping of these different infections for testing a single treatment using a RCT was valid would have shown that these lumping RCT would NOT expected to be reproducible.
Some may call a massively lumping (PettyBone) RCT a decades old standardized pathological design of experiment but I call it a “RCT mimic” because it should not have the purple robes of a real RCT which fooled ALL of us and the RCT facade is still fooling most critical care science leaders and grant reviewers who decide the fundable standard research methodology of the field.
(At one time back in the 20th century we were all fooled because they were labeled as RCT)
Here you see formal causal modeling is a means to add much needed rigor to help determine the DoE. Had causal modeling been required to acquire the grants in critical care in the past, it might have saved critical care science decades of wasted research and careers and massive wasted resources and unnecessary patient harm. Even if the approach is Bayesian the same symbolic causal modeling would be required.
So there was a desire to move to solutions. This of course is only a “solution hypothesis”. What do others think of this proposed solution to improve critical care experimentation science.
Hope you’ll get better feedback from others on the site.