Clustering

Are clinical trials designed to sell drugs? 

Drugs are expensive products.  The profits they produce are protected by laws that license their availability. The laws are designed to protect  consumers from danger and fraud.  Many drugs are clearly dangerous. ( It is not an accident that same Hebrew word means herb, poison and medicine) Oncologists use very dangerous drugs and their approval is done in an atmosphere of desperation. 

Clinical trials prove the safety and efficacy of drugs. Patients are heterogenous along many dimensions that impact on safety, efficacy, tolerability of compounds in trials.  

The criteria for safety and efficacy depend upon the disease. Many different entities can be called the same disease. If the criteria  for efficacy are low enough, selecting a subset  of patients with a more responsive (form of the) disease may prove sufficient efficacy for approval. Since the responsive subset was not identified, the more general use of this medicine in a more representative population will not show the same promising result. 

Great pains are taken to assure that experimental and control groups are matched, but the parameters that separate responders from failures may be subtle.  Thus, limiting a trial to patients who have an adequate "functional status" may increase the chance that the medicine is given to a responsive patient. When the drug is approved, it may be limited to higher performing patients ... and that limit may be enforced by payers... and the limitation may not be valid.  Alternatively, the population identified by the surrogate marker of functional status may exclude others who would benefit, but their identity is not ascertained. 

How should it be:

Phase 2 trials should include clustering. As many details that pertain to diagnosis ( and its ambiguities) should be cataloged, along with as many parameters that pertain to tolerance, along with readily available parameters ( blood tests, urine, VS, especially height and weight, etc) .  Parameters of response, beyond RECIST or other "objective" parameters, should be recorded.  The data should be used in a dimension reduction, difference finding, fuzzy logic ML program., like UMAP in R

The outcome could identify subgroups who respond and others that do not. Comparing responding and non-responding subgroups in several trials could help redefine the disease set and point to  a direction for investigation or management of non-responders... as well as the lucky ones. 

I would be surprised if this is not already done by sponsors of clinical trials, but it may often not be in their interest to release such information, since it could limit sales. If this is happening, it is a crime.