Causality in Medicine: Moving Beyond Correlation in Clinical Practice

A growing body of research suggests it’s time to abandon outdated ideas about how to identify effective medical therapies.

The statement "correlation is not causation" emphasises the distinction between two statistical concepts, suggesting that the presence of a correlation between two variables does not necessarily imply a causal relationship between them. The acceptance of this statement by researchers is widely acknowledged and justified. The sequential occurrence of event A and event B does not imply a causal relationship between A and B. The findings of an observational study using a sample size of 1,000 adults indicate a potential association between high amounts of vitamin C intake and a reduced likelihood of developing lung cancer. However, it is important to note that this study does not establish a causal relationship between the nutrient and cancer prevention. It remains plausible that the observed association could be influenced by a confounding variable, which may have influenced both the intake of vitamin C and the occurrence of lung cancer. In essence, individuals who consume high quantities of vitamin C may exhibit a decreased susceptibility to developing lung cancer due to their heightened health consciousness relative to the general population. Consequently, they are more inclined to abstain from smoking, thereby mitigating their risk of developing this particular form of cancer.

This example serves to demonstrate the presence of confounding variables, which are potential circumstances that can lead to a misleading perception of a cause-and-effect relationship that may not actually exist. The rationale behind the preference for interventional trials, such as randomised controlled trials (RCTs), over observational studies is in their greater reliability in establishing causal relationships. However, it is crucial to note that within the realm of clinical care, there exist numerous treatment procedures that lack evidence from randomised controlled trials (RCTs). Likewise, numerous risk factors are linked to diverse diseases; however, establishing a definitive causal relationship between these risk factors and the onset of the diseases can often pose challenges. 

Randomised controlled trials (RCTs) continue to be widely regarded as the gold standard in the field of medicine. However, their practicality is often hindered by several factors. Firstly, the cost associated with conducting RCTs is often prohibitively high. Secondly, it would be considered unethical to conduct an RCT that exposes patients to potentially harmful risk factors and compares them to a control group that does not face such risks. Lastly, the inclusion and exclusion criteria typically For example, researchers often opt to exclude individuals with concurrent medical issues, a practise that has the potential to introduce bias and hence impact the findings of the study.

One potential approach to tackle this issue involves embracing imperfect evidence and employing a reliability scale or continuum to assess the efficacy of different therapies, distinguishing those that are deemed worthwhile from those that are not. The proposed scale exhibits a gradual increase in the strength of evidential support from the left to the right throughout the continuum. 

In the absence of randomised controlled trials (RCTs), it is plausible to contemplate the utilisation of observational research, such as case-control and cohort trials, as a means to substantiate the implementation of a certain therapeutic intervention. Although observational studies might potentially be misleading due to the omission of confounding variables, there exist epidemiological criteria that enhance the credibility of these imperfect investigations:

• A heightened degree of association or correlation between two variables tends to provide stronger indications of a potential cause-and-effect link compared to a lesser association.
  
  
• Temporality. The purported consequence should adhere to the presumed cause, rather than vice versa. It would be illogical to propose that the development of tuberculosis (TB) is attributed to exposure to Mycobacterium tuberculosis if all instances of the infection were observed prior to patients' exposure to the bacteria.
  
  
• There is a discernible dose-response association between the purported cause and its corresponding effect. As an illustration, in the event that researchers discover a correlation between a blood lead level of 10 mcg/dl and minor learning problems in children, a blood lead level of 15 mcg/dl and moderate deficits, and a blood lead level of 20 mcg/dl and severe deficits, this continuum of effects provides more support for the argument of causality.
  
  
• The inclusion of a physiologically plausible mechanism of action that establishes a causal relationship enhances the strength of the argument. The occurrence of lead poisoning has been associated with neurological impairment resulting from oxidative stress and various other biochemical pathways, as supported by available research.
  
  
• The repeatability of study findings is enhanced when separate investigators are able to replicate the results obtained by a previous group of investigators. This replication serves to provide additional evidence for the cause-and-effect relationship under investigation.

The fulfilment of these requirements implies a causal relationship in observational studies, however, the establishment of causality can be achieved by the application of a statistical technique known as causal inference. The technique, pioneered by Judea Pearl, Ph.D., recipient of the 2011 Turing Award, is widely regarded as revolutionary by prominent intellectuals and is expected to have significant ramifications in the field of clinical medicine, as well as for the advancement of AI and machine learning. At the recent Mayo Clinic Artificial Intelligence Symposium, Adrian Keister, Ph.D., a senior data science analyst at Mayo Clinic, asserted that causal inference represents a potentially significant advancement in the scientific method, comparable in importance to the emergence of modern statistics, if not surpassing it.

Causal inference, in its conceptual framework, involves the transformation of claims expressed in natural language into mathematical statements, facilitated by the introduction of novel operators. Although this may appear intimidating to anyone lacking a strong background in statistics, it is fundamentally similar to our use of math as a means of communication. The mathematical expression "fifteen times five equals seventy five" can be represented as "15 x 5 = 75". In the present scenario, x functions as an operator. If the new mathematical language of causal inference were to be employed to depict an observational research assessing the correlation between a novel pharmaceutical intervention and a rise in individuals' longevity, it might be represented as follows: The expression P (L|D), where P represents probability, L represents longevity, D represents the drug, and | denotes the conditioning operator, signifies the likelihood of lifespan given the drug.

In contrast, an interventional study, specifically a randomised controlled trial (RCT), would be expressed as X causing Y if the probability of Y given the intervention do(D) is greater than the probability of Y in the absence of the intervention. In this context, the do-operator signifies the act of implementing the intervention, such as administering the drug under controlled conditions. The aforementioned formula represents a method of asserting that the administration of medicine X yields the outcome of longer life, provided that the observed effects of the intervention surpass the likelihood of achieving longer life without the treatment, specifically denoted as the probability in the placebo group, P(Y).

This novel methodology also use causal graphs to illustrate the association between a confounding variable and a hypothesised cause-and-effect relationship. By employing this particular graph, it becomes possible to depict the practical application of the technology within a real-life context. Examine the correlation between smoking behaviour and the development of lung cancer. Over the course of several decades, there has been an ongoing debate among statisticians and policy makers on the causal relationship between smoking and cancer, mostly due to the reliance on observational evidence to establish this link. The graph would exhibit a visual representation similar to the one depicted.

Figure 1:

The variable denoted as G represents a complicating factor, such as a genetic propensity, whereas S represents smoking and LC represents lung cancer. The inference being made is that if a third variable is responsible for both individuals smoking and developing cancer, it is not necessarily possible to definitively establish a causal relationship between smoking and lung cancer.  Pearl and his colleagues made a significant finding that entails the identification of an intermediary element in the link between smoking and cancer. This discovery enables the establishment of a causal relationship between the two variables through the utilisation of mathematical computations and algebraic transformations. As depicted in Figure 2, the tar deposits found in the lungs of smokers serve as an intermediary component.To gain a more comprehensive comprehension of the principles behind causal inference, it is highly recommended to delve into Judea Pearl's seminal work, "The Book of Why." The resource offers a comprehensive elucidation of causal inference in a simplified and accessible manner. To further explore the topic, one may consider referring to the book titled "Causal Inference in Statistics: A Primer."

If causal inference had been available throughout the 1950s and 1960s, it is plausible to suggest that the argument put up by tobacco industry lobbyists could have been effectively countered. Consequently, this could have potentially resulted in the prevention of a significant number of premature deaths, amounting to millions of lives. The aforementioned approach exhibits significant promise as we embark on its application to prediction algorithms and other digital tools based on machine learning. 

Figure 2:

To gain a more comprehensive comprehension of the principles behind causal inference, it is highly recommended to delve into Judea Pearl's seminal work, "The Book of Why." The resource offers a comprehensive elucidation of causal inference in a simplified and accessible manner. To further explore the topic, one may consider referring to the book titled "Causal Inference in Statistics: A Primer."

If causal inference had been available throughout the 1950s and 1960s, it is plausible to suggest that the argument put up by tobacco industry lobbyists could have been effectively countered. Consequently, this could have potentially resulted in the prevention of a significant number of premature deaths, amounting to millions of lives. The aforementioned approach exhibits significant promise as we embark on its application to prediction algorithms and other digital tools based on machine learning. 

Bagikan Artikel ini