This article was written by Nicholas Mitsakos : Chairman and CEO at Arcadia Capital Group.
Artificial Intelligence, Correlation, and Failure
Artificial Intelligence identifies correlations far more frequently than causal relationships. Correlations show how certain phenomena go together, which can be quite useful in many circumstances – but not in medicine. Only causal links tell why the presence of one thing determines another. In medicine, no diagnosis can be effective if we don’t know what causes what so we will know what to block, prevent or eliminate. Medicine needs to understand how a system is working, how it will be affected and how it might evolve. AI is touted as medicine’s new effective tool. But it needs a lot of help. Correlation is not causation, and AI’s technology is especially good at understanding correlation; not so much at causation. Since medical treatments currently recommended by AI focus more on correlation, they will ultimately fail.
In medicine, a vast number of variables can be interlinked. Diagnosing diseases depends on knowing which conditions cause what symptoms; treating diseases depends on knowing the effects of different drugs or lifestyle changes. Untangling the answers to these complex questions – understanding interrelationships and how an overall system is impacted by various elements – is typically done via rigorous observational studies or randomized controlled trials.
A Causes B, B Causes C, But A Doesn’t Always Cause C
These studies and trials create a wealth of medical data, but these data are spread across different clusters and sets, residing in siloed repositories, leaving many questions regarding causality and interrelationships unanswered. If one data set shows a correlation between obesity and heart disease and another shows a correlation between low vitamin D and obesity, what’s the link between low vitamin D and heart disease? Finding out typically requires another clinical trial. Many times, these trials can be unconvincing often because of poor trial design, cohort identification, and improper targeting – and are wasteful and time-consuming.
How do we make better use of this vast, but piecemeal, information? Computers spot patterns quite effectively, but, as we’ve seen, that’s just correlation. A handful of algorithms can identify causal relations within single data sets. But focusing on single data sets ignores nature’s complexity, won’t reflect what’s really happening in the real world, and the results are not useful or implementable for true therapeutic effectiveness.
Looking at it Another Way
A new way of looking at data is needed. Artificial Intelligence gives us the hope that causal relationships and useful treatments can be developed much more efficiently, and effective therapies more quickly. However, artificial intelligence has many drawbacks, including working quite well with the defined data set, but not understanding interactions among different data sets, or interrelated relationships among these data. AI, in spite of the hype, cannot diagnose medical conditions better than a human doctor. In fact, “it may perform significantly worse” (The Lancet, 2018). AI is still inadequate at identifying causation.
This problem has not been solved, although interesting research is ongoing. The most promising techniques allow large databases of untapped medical data to be mined using new algorithms to potentially discover new causal links.
Send in the Cryptographers
A new technique, inspired by quantum cryptography, allows large medical databases to be tapped for causal links. This is a fundamental breakthrough in thinking, and this perspective has the potential to spot cause-and-effect, supercharge medical diagnoses, and use AI effectively.
What is promising is that, instead of focusing on machine learning and the algorithms that artificial intelligence develops, this new quantum cryptography approach uses a mathematical formula that, essentially, tries to prove that data and communication are intact – no unwanted third parties are eavesdropping or observing data. If medical data sets can be treated as conversations, and variables that influence those data sets are thought of as third-party eavesdroppers, the math of quantum cryptography can determine whether or not any causal effects exist.
This has been tested on known causal relationships with interesting results. AI could detect the presence of pathology but could not identify a causal link among two or more elements that are known to cause the pathology (in this case, a malignant tumor where the size and texture were known and identified as causal, AI could not distinguish a causal link). However, the technique using quantum cryptography algorithms appears to identify causal relationships between these variables. In this study, the algorithm’s results were compared to a clinical study and the results were comparable.
In essence, this can be a powerful research tool for the development of new therapies. Instead of running new randomized controlled trials, the software may be able to use existing data to generate new results. This is quite promising.
A New Idea and a Long Road
There is still a long way to go, but this fresh approach addresses a fundamental misconception and flaw in thinking about AI and its potential for medical diagnosis, research, and drug development. Initially, this and other approaches can be used to complement trials and can perhaps highlight potential causal links more effectively.
New drugs are approved on the basis of correlation far too often because more robust data is too time-consuming or unachievable. Yet, there is an immediate medical need, so these proposed treatments move forward. Approvals based on randomized controls and focused trials are often unconvincing regarding true causality when the data is examined thoroughly. AI combined with new algorithmic approaches, such as quantum cryptography, may enable us to identify causation, and therefore much more effective treatments, more quickly, and cost-efficiently.