Among the countless medical conferences I've attended over the years, one presentation stands out. Called "Stump the Professors," it featured a panel of top medical school professors who were challenged to figure out the diagnosis in several tough clinical cases.
One by one, residents would file up to the podium and describe their case, outlining signs and symptoms, lab results, and pathology findings. One thing that stands out in my mind is the lighthearted banter and laughter among the participants as professors struggled to figure out some of these esoteric disorders. The residents relished the fact that they really were stumping the professors.
Among the expert panelists was one professor who remained relatively quiet until the end of each case, only occasionally asking questions about some lab report or symptom. But when all the facts were in, he would nail virtually every diagnosis. What an amazing gift.
Fast forwarding to today and the age of clinical decision support software (CDSS) and IBM's Watson computer, the question that comes to mind is: Can health IT duplicate this kind of diagnostic reasoning, or at least facilitate the process?
Computer-generated diagnostic tools, which are usually part of a larger CDSS, come in a few basic flavors. Many consist of three main components: a knowledge base, an inference engine, and a communication mechanism.
The knowledge base contains reasoning guidelines, including "if/then" rules. For instance, one rule may state: If drug X is taken with drug Y, alert the physician for a possible harmful interaction. The inference engine then combines the rules with each patient's data--in this example, it would be the list of medications he's taking. The communication mechanism provides a way to give the results of the data analysis to the clinician, who then decides what action to take.
Other diagnostic programs rely not on a knowledge base, but on some form of artificial intelligence--neural networks, for instance--to find patterns in the patient's data and associate them with a list of diagnoses.
Among the players in the field of diagnostic software are Simulconsult and Dxplain. Typically, these programs use probabilistic algorithms and pattern-matching to create a list of diagnoses ranked from most likely to least likely disorders. The software generates the list of diseases by looking at each sign, symptom, and lab finding to determine how strongly each of the physical findings is associated with the disease under consideration. For example, how probable is it that a patient with carrot-orange skin is suffering from carotene toxicity.
Many of these programs also incorporate a huge database that condenses clinical research from major clinical journals and clinical practice guidelines to ensure that the final results are best practice.
Taking A Closer Look
The software developed by Simulconsult, which incorporates statistics-enhanced pattern matching, offers a good example of the impressive reasoning skills of these diagnostic tools.
The example provided by Simulconsult starts with a patient presenting with kyphosis, an abnormal curvature of the spine. The initial information screen asks for the patient's age (8 months), gender, and family history. Then it asks the physician to input the most unusual or prominent physical finding.
Once kyphosis is entered in the search engine, the system wants more details: Is the kyphosis accompanied by scoliosis? The clinician observes that there is no scoliosis, enters that fact, and the program then asks: When did the kyphosis begin? A drop-down menu lets him choose its onset.
Once that initial data is entered, the program offers several possible differential diagnoses, listed in a chart that shows their relative probability in an easy-to-read graph. At the end of this first run in this example, achondroplasia--a common form of dwarfism--tops the list.
Now the software gets really sophisticated. The physician has to add more physical findings to narrow the possible diagnoses. This is the juncture at which the program will outthink many clinicians who can't bring to mind the long list of possible findings that might accompany kyphosis in an 8-month-old baby who is not suffering from scoliosis.
The program itself provides a list of possible physical signs to look for. In this case, it includes macrocephaly--an enlarged head--which the baby does have. Once the doctor clicks on enlarged head and chooses onset at about 6 months of age, the graph listing all the differential diagnoses changes dramatically, with mucopolysaccharidosis Type I, an inherited metabolic disorder, jumping to the top of the graph.
The program takes the clinician through a series of other physical findings and eventually suggests a blood test to confirm the diagnosis. Quite impressive.
How Well Accepted?
So why aren't diagnostic aids like Simulconsult more widely used? For one thing, this particular program confines itself to genetic, neurological, and metabolic diseases, which narrows down considerably the millions of combinations of physical findings and diseases any one patient may suffer from.
But the obstacles go beyond that one. Many diagnosis-support systems are standalone applications that don't work well with electronic health record systems. Busy clinicians often object to the extra work required to locate patient data and input it into a separate program.
Assuming that these obstacles could be overcome, how effective are diagnostic programs overall? In a review published in 2005 in the Journal of the American Medical Association that included 10 trials of such applications, four of the studies found the programs made a difference because they helped diagnose cardiac ischemia in the emergency room; reduced unnecessary hospital or coronary care admissions; improved detection of mood disorders in a posttraumatic stress clinic; or sped up the diagnosis of bowel obstruction.
So evidently these programs work--at least some of the time. There are two reasons for their selective success rate, according to the JAMA report. When clinicians are automatically prompted to use the software through the workflow process, the programs produce better results. And in general CDSS programs that are home grown--developed by hospitals and other care providers--seem to be more successful, evidently because there are physician champions in-house to encourage others to get on board.
Thinking back to that Stump The Professors presentation, it's hard to imagine artificial intelligence experts developing a diagnostic support program that could duplicate the supercomputer inside that one medical professor's head.
But at the same time, it's hard to imagine a physician who would not benefit from having a user-friendly database of relevant diseases, linked to all the millions of combinations of signs, symptoms, and lab findings. Once clinicians venture past the diagnosis of common, easily recognizable disorders, it seems almost impossible to do all the mathematical calculations needed to find that proverbial needle in a haystack.
Paul Cerrato is editor of InformationWeek Healthcare.