Tuesday, May 03, 2016

Will the computer someday replace the physician as diagnostician?

With the growing enthusiasm over Watson and other forms of high technology decision support has come the nutty idea that computers may eventually surpass clinicians in the diagnostic process. Taking that idea to its full extent, in such a world the role of doctors would be restricted. The need for clinicians would be gone though we would still need providers to navigate the EMR and coordinate care (essentially secretarial duties), do procedures and maintain a “human touch” in healthcare through education, counselling and other types of social interaction. Could this ever come to pass?

It has already been the subject of an experiment, the conditions of which gave the idea the best possible chance to work in two ways. First, the experiment was conducted in what is arguably one of the most mechanistic and formulaic areas of diagnostic medicine. Second, it's been going on, repeated time and time again with generation after generation of software “improvement,” for decades. I am referring, of course, to computerized interpretation of electrocardiograms. Despite being given every conceivable chance it has failed. From a recent review on the topic:

The use of digital computers for ECG processing was pioneered in the early 1960s by two immigrants to the US, Hubert Pipberger, who initiated a collaborative VA project to collect an ECG-independent Frank lead data base, and Cesar Caceres at NIH who selected for his ECAN program standard 12-lead ECGs processed as single leads. Ray Bonner in the early 1970s placed his IBM 5880 program in a cart to print ECGs with interpretation, and computer-ECG programs were developed by Telemed, Marquette, HP-Philips and Mortara. The “Common Standards for quantitative Electrocardiography (CSE)” directed by Jos Willems evaluated nine ECG programs and eight cardiologists in clinically-defined categories. The total accuracy by a representative “average” cardiologist (75.5%) was 5.8% higher than that of the average program (69.7, p less than 0.001).

Those results don't say much for the cardiologists either but that's a topic for another discussion.  In a green journal editorial in 2012 Dr. Joseph Alpert cited additional research from the 1970s:

In 1976, I was involved in one of the earliest evaluations of 5 competing computer programs that interpreted electrocardiograms (ECGs).1 At that time, computer interpretation of ECGs was just beginning to make its way into hospitals in the United States and abroad. Dr Arthur Hagan and I evaluated the accuracy of the different computer interpretations compared with our own experienced analysis of more than 100 ECGs with various well defined abnormalities.

The results were illuminating. The computer interpretations were often wrong, particularly with respect to arrhythmia identification. Furthermore, the different computer ECG readings from the 5 programs often were surprisingly different. The conclusion of this early study was that computers were not as accurate in reading ECGs when compared with experienced cardiologists. We suggested that all computer-read ECGs should be over-read by an experienced physician. In the end, this study showed that the overall accuracy score for the computer ECG programs was approximately 80%, and as already noted, the computer was particularly poor on arrhythmia interpretation.

Of note, Alpert cites no improvement in over 30 years. Again from the editorial:

This is still the situation today with all ECGs with computer diagnoses over-read by an experienced physician, usually a cardiologist. Of note, when I am the over-reading cardiologist in our hospital, I still find that the computer reading of the ECG is incorrect approximately 20% of the time.

Because we often rely on the ECG to supply the critical data to guide decision making in very ill patients, this is unacceptable. And it hasn't improved in decades. These numbers were derived using artificial conventions. The results would certainly be even worse against more nuanced standards based on subtle ECG patterns.

Alpert suggests the reason for such poor results:

What is the reason that the most sophisticated computer ECG interpreting software makes so many mistakes? I think the answer lies in the remarkable and extensive capacity of the human brain to recognize visual patterns. This capacity is the reason that a person with minimal prior instruction can recognize a van Gogh painting without looking at the accompanying label. The distinctive style of van Gogh is easily recognized by the highly complex visual pattern recognition system of our central nervous system... Today, we apply this ability in a variety of areas, including athletic endeavors, police investigations, aesthetics, and many other venues, including the interpretation of ECGs.

Based on this explanation and the lack of progress over time it would appear unlikely that the computer will supplant the clinician in ECG interpretation let alone in other areas of diagnostic evaluation that are far more complex and less mechanistic.

No comments: