In a recent interview with The New England Journal of Medicine, Dr. Adewole Adamson, Dermatologist, Health Services Researcher, and Assistant Professor of Internal Medicine at Dell Medical School, University of Texas explained what machine learning (ML) is and the promises it brought to medicine through analyzing a vast amount of data points and generating its own decision making or prediction algorithms. Dr. Adamson said whilst ML assist medical professionals to work more efficiently, there are certain inherent limitations, particularly in the diagnoses of early stage cancers.

Absent of “gold standard”

He explained most ML algorithms are trained via a process called supervised learning, whereby computers are presented with thousands or even millions of images labeled by different human pathologists. The algorithm will then learn to categorize these images into “cancer” or “not cancer” on the basis of patterns (i.e., shapes, colors, edges) rather than developing their own systems of judgement.

Often, for early stage cancers, there are many disagreements amount pathologists on whether an image or a pathological observation should be accounted for cancer because clinically defined cancer is also accompanied by symptoms. Some of these abnormalities may meet pathological definition of cancer but may not result in symptoms or death. In view of this variability, there is no absolute benchmark to determine if a ML algorithm had indeed make an accurate categorization because in the eyes of some pathologists, it may be a false positive.

Resulting in unnecessary treatments

Although ML algorithms may replicate or even outperform human in categorization or making predictions, it may not necessary lead us closer to the truth, since it lacks the mean to ascertain whether an abnormality will eventually evolve into cancer or death. Nevertheless, as Dr. Adamson pointed out, ML is a powerful techniques capable of detecting even the smallest anomaly which comes with unknown clinical significance, resulting in further treatments that may or may not be necessary. As such, there is a need to deploy this valuable technology with care.

With that, he suggested adopting a three-category framework. Instead of putting images into “cancer” or “not cancer”, Dr. Adamson proposed “total agreement that’s cancer”; “total agreement that’s not cancer” and “limited”. He believe this framework will offer a more efficient and honest categorization as uncertainties are highlighted right away to clinicians and they can spend more time on these borderline or tricky cases. In the long run, the framework can also facilitate the field forward, as professionals better define what is true cancer that will designate to shorten someone’s life or cause undesirable symptoms.

Author Bio

Hazel Tang A science writer with data background and an interest in the current affair, culture, and arts; a no-med from an (almost) all-med family. Follow on Twitter.