by Dr. Steve Poulin

Machines walkingComputers tutoringCars drivingAlgorithms diagnosing.  As we outlined in Part One of this series, each of these examples highlights some of the amazing progress that we have made using artificial intelligence.  In Part Two, we discussed how these artificial intelligence systems work.  Beyond understanding what artificial intelligence is and how it works, we must also understand its pros and cons.

To truly appreciate the pros and cons of artificial intelligence, we’ll explore the ideas through the lens of using algorithms for medical diagnoses.  These algorithmic diagnoses hit upon the crux of the debate over machine learning.  This is a task that has been faithfully performed by well-trained medical practitioners for centuries.  How does this change when we hand the reins over to machines?

First, let’s talk about the pros of using artificial intelligence techniques for medical diagnoses:

  • There is no limit on the number of factors we can use.
    When humans diagnose conditions, there is a limit to the amount of information that we’re able to ingest at any given point.  Machines are not limited by these constraints.  They can process any number of factors that they are given.
  • They can be far more accurate.
    Medical practitioners follow very specific rules for diagnoses.  In the article, the ABCD mnemonic for diagnosing melanomas is highlighted - “Melanomas are often asymmetrical (“A”), their borders (“B”) are uneven, their color (“C”) can be patchy and variegated, and their diameter (“D”) is usually greater than six millimetres (sic)”.  This is essentially a series of 4 if-then statements.  For artificial intelligence techniques, there is no limit on the number of factors that can be included in these rules.  This means that we can include as many if-then statements as the algorithm determines we may need.  More importantly, the weights created for each factor very precisely represent their relative importance in the rules.  By generating far more if-then statements, artificial intelligence techniques offer much greater accuracy for a diagnostician.
  • Rules can be quickly applied.
    In medicine, the ability to quickly diagnose an ailment is of utmost importance for long-term patient health.  For example, being able to quickly diagnose a stroke or identify cancer early would have major long-term benefits for patients.  Machine learning techniques can quickly apply these very complex rules to generate a diagnosis, which is critical in a field in which delays may have serious health consequences.

While each of these is wonderful, we must also understand the relative cons of using artificial intelligence techniques:

  • Algorithms can take time to run.
    Although machine learning represents one of the most sophisticated methods for predictive analytics, one of its drawbacks has been the amount of time it can take to run.  However, this issue has become less relevant as computer processing speeds have dramatically increased.  Nevertheless, the volume of data available for predictive analytics has also grown, which requires more efficient machine learning algorithms.
  • We may “overfit” the data.
    One of the most serious risks of using machine learning for predictions is a problem known as overfitting.  This term is used to describe the development of a predictive model that is too specific its subjects, and therefore does not generalize well to new subjects.  Overfitting is a serious problem if the data used to develop models is biased.  For instance, 96 percent of participants in modern studies on disease genetics have been people of European descent, which means that models built from these studies would not generalize well to people of African or Asian descent.
  • The algorithms are a “black box”.
    With so much data being ingested and so many rules being generated, artificial intelligence techniques often leave us asking “why?”  Artificial intelligence systems can build the models on their own, but do not tell us how the models have been built.  Indeed, the models can often appear to be “all knowing but perfectly impenetrable.”  This flies in the face of what we might expect from a medical diagnosis, where a skilled physician is able to reassure us about the whys just as much as the whats.

While these issues are serious, solutions are being developed that mitigate them.  For example, in response to how long algorithms may take to run, “big data” applications such as Apache Spark have been developed.  These systems can process data from distributed file systems such as Hadoop, and analyze data in-memory rather than moving data in and out of a hard drive. This has drastically improved processing speed, with more improvements to come.

Despite the need for caution, machine learning offers an alternative to the traditional predictive analytics methods that can incorporate greater complexity.  Even with these issues, we believe that the benefits of artificial intelligence techniques far outweigh the risks.  With the attention it is receiving from developers involved in predictive analytics, artificial intelligence techniques will only become more prevalent as they become more efficient and accurate.  We predict that artificial intelligence will make its way into all facets of our lives, far beyond machines walking or algorithms diagnosing.

AuthorRyan Harrington