The principle of parsimony Firstly, however, we have to consider how we may measure the usefulness of a method of diagnosis. That it should be sufficiently accurate goes almost without saying, but I would like to suggest that brevity (or, to be more specific, parsimony) is the second most important criterion by which a diagnostic process should be judged.
The great Johann Karl Friedrich Gauss, for example, when challenged as to how he had arrived at the conclusion of one of his exquisitely brief theorems, would reply loftily, "When a beautiful cathedral is built, who wants to see the scaffolding?" His proofs were indeed so short that they were often disbelieved.
Is diagnosis necessary? If we pursue the virtue of parsimony sufficiently ruthlessly we reach the interesting conclusion that under certain circumstances diagnosis is a pointless or meaningless diversion in the therapeutic process, which is, after all, the one that interests the patient. As a first example, take the question of whether a patient who has acute pain in the abdomen has appendicitis or Meckel's diverticulitis. Provided that the affected structure is in its usual position the precise diagnosis has no effect whatever on the management of the patient. It is the decision to perform a laparotomy through an incision in the right iliac fossa that matters.
A man aged 45 who has a blood pressure of 130/87 mm Hg.
The question
"Does this patient have hypertension?" is pointless,
as it assumes that patients either have or do not have hypertension, whereas all that they have is different degrees of increased blood pressure. The key questions are not diagnostic but prognostic-namely, is this patient at increased risk of death, stroke, or other complications, and if so will the benefits of antihypertensive treatment or the search for a directly treatable cause, or both, outweigh the costs? Finally, consider the value of psychiatric diagnosis, a subject on which whole books have been written.' I recall as an undergraduate being alternately fascinated and astonished by the teaching of William Sargent. Into his outpatient department would come a man enshrouded by an almost palpable aura of gloom; we would be told rather unnecessarily that he was depressed. Next through the door would come a woman who looked as if she had just won a parliamentary election, even though she was anxious about the possibility of there being a recount. "Ah, another classical example of depression," would declare the master. It took a brave student indeed to ask how two such different people could possibly be suffering from the same disease. "Simple, the lady has atypical depression," we would be told. I have to confess that my nerve failed me at this point, and I dared ask no more. Ultimately, the truth dawned. Patients who had classical endogenous depression responded to drugs that were self evidently antidepressant. If patients who had other symptoms responded to antidepressant drugs then they must be depressed. Simple, really. However illogical the argument, Sargent was trying to maintain that what mattered was to recognise which patients would respond to different forms of treatment, rather than to argue over the truth of the diagnosis with which their problem was labelled.
We may look on our therapeutic objective as killing the tubercle bacillus or closing the ventricular septal defect, but this has no value unless we make the sick patient better. Furthermore, this disease centred approach breaks down with something like rheumatoid arthritis, as it is at present not possible to separate the disease from the patient. Our only therapeutic objective is to make the sick patient better, so this may or may not include "naming" the disease.
The elucidation of the causative, structural and functional changes may not come in any particular historical order, but the paradigm has two characteristics: first, it is expected or at least hoped the relations will be specific (unique cause, unique structural and functional changes belonging to one syndrome); second, as knowledge progresses, the defining process is 'pushed to the left' in the sequence given above. In other words, a disease will not be allowed to remain in syndromal terms if it can be explained or defined in functional terms; a functional syndrome will not be left in these terms if it can be characterised structurally, and 'cause' takes priority overall."
Algorithmic approaches to diagnosis are exemplified by programs for evaluating acid-base disorders8 and comatose patients.9 If an algorithm is sufficiently simple there is no point in computerising it; it is quicker and simpler to follow a printed version of the original algorithm.
THE BLUNDERBUSS APPROACH The blunderbuss approach is the traditional method taught to medical students. They take a detailed history, examine the patient from top to toe, and then order every test that could conceivably have some bearing on the problem. Not until all the information is to hand do they try and work out what is the matter with the patient. This is done by fitting the pattern of abnormalities found either to textbook descriptions of diseases or to their own database of diseases in patients whom they have previously seen. Computer applications The nearest automated approach to the blunderbuss method is that of database comparisons.5 An interactive search is made of a large database of information on patients, looking for those who match the particular patient under consideration. Usually a match is first sought on a limited list of features, with the result that a rather large and inhomogeneous subset of matching patients is obtained. The number of features to be matched is then increased, and the subset usually becomes smaller. When a matching subset is obtained in which the disease diagnosed in all patients is the same the new patient is assumed to have the same disease and can then be added to the database. This method requires the accumulation of a large database free from errors, which is very expensive, yet there is no real concept within it of a list of possible diagnoses, each having a different probability of being true. Probably the most suitable application is inthe diagnosis of rare syndromes,' where there is a real problem of human memory and collation of small snippets of information from diverse sources.
This clinical approach is woefully unimaginative, cumbersome, and extravagant (as opposed to parsimonious). When applied to laboratory tests it wastes money, not only because many of the tests originally ordered are unnecessary but also because the more tests that are ordered the more likely it is that, by chance, one or more will turn out to be "abnormal" and start a wild goose chase of further tests to investigate the chance abnormality.
Thirdly, from the point of view of computer modelling (and this applies to expert systems in general) the question has to be asked whether the objective should be to mimic clinicians or to use computers to do the things that clinicians cannot do in their heads (such as multivariate analysis). I understand the motivation of those working with computers who wish to mimic the brain, but I as a doctor want a system that will do better than the best clinician. A machine that simply does what a clinician (even a superb clinician) does is simply not a very attractive proposition
speech recognition by computers became successful, when statistical analysis was chosen leaving the liguistic rules system.
The ultimate in statistical diagnosis is to make no prior assumptions about diagnostic categories but to let the data speak for themselves. This corresponds to syndromic diagnosis by computer. For obvious reasons this approach has been mostly applied to the diagnosis ofmental illness, in which there is considerable doubt about the validity of traditional diagnostic categories. On the whole, however, these multivariate methods (factor and cluster analysis) have proved to be disappointing.'
BRITISH MEDICAL JOURNAL VOLUME 295 21 NOVEMBER 1987
Conclusions No explanation of human diagnostic logic so far conceived has been entirely satisfactory, though study of the alternative models is extremely instructive. Similarly, no method of diagnosis helped by computers has been shown consistently to be superior to all others. This is an exciting field of research precisely because it is so wide open. The validation of approaches by artificial intelligence to diagnosis has been particularly scanty-either non-existent or based on fewer than 20 patients. It is essential that comparisons of alternative diagnostic aids' "Is2 should be carried out as stringently as are those at present for new therapeutic aids such as drugs. FJM is supported by the Vanderbell and British Heart Foundations.
The great Johann Karl Friedrich Gauss, for example, when challenged as to how he had arrived at the conclusion of one of his exquisitely brief theorems, would reply loftily, "When a beautiful cathedral is built, who wants to see the scaffolding?" His proofs were indeed so short that they were often disbelieved.
Is diagnosis necessary? If we pursue the virtue of parsimony sufficiently ruthlessly we reach the interesting conclusion that under certain circumstances diagnosis is a pointless or meaningless diversion in the therapeutic process, which is, after all, the one that interests the patient. As a first example, take the question of whether a patient who has acute pain in the abdomen has appendicitis or Meckel's diverticulitis. Provided that the affected structure is in its usual position the precise diagnosis has no effect whatever on the management of the patient. It is the decision to perform a laparotomy through an incision in the right iliac fossa that matters.
A man aged 45 who has a blood pressure of 130/87 mm Hg.
The question
"Does this patient have hypertension?" is pointless,
as it assumes that patients either have or do not have hypertension, whereas all that they have is different degrees of increased blood pressure. The key questions are not diagnostic but prognostic-namely, is this patient at increased risk of death, stroke, or other complications, and if so will the benefits of antihypertensive treatment or the search for a directly treatable cause, or both, outweigh the costs? Finally, consider the value of psychiatric diagnosis, a subject on which whole books have been written.' I recall as an undergraduate being alternately fascinated and astonished by the teaching of William Sargent. Into his outpatient department would come a man enshrouded by an almost palpable aura of gloom; we would be told rather unnecessarily that he was depressed. Next through the door would come a woman who looked as if she had just won a parliamentary election, even though she was anxious about the possibility of there being a recount. "Ah, another classical example of depression," would declare the master. It took a brave student indeed to ask how two such different people could possibly be suffering from the same disease. "Simple, the lady has atypical depression," we would be told. I have to confess that my nerve failed me at this point, and I dared ask no more. Ultimately, the truth dawned. Patients who had classical endogenous depression responded to drugs that were self evidently antidepressant. If patients who had other symptoms responded to antidepressant drugs then they must be depressed. Simple, really. However illogical the argument, Sargent was trying to maintain that what mattered was to recognise which patients would respond to different forms of treatment, rather than to argue over the truth of the diagnosis with which their problem was labelled.
We may look on our therapeutic objective as killing the tubercle bacillus or closing the ventricular septal defect, but this has no value unless we make the sick patient better. Furthermore, this disease centred approach breaks down with something like rheumatoid arthritis, as it is at present not possible to separate the disease from the patient. Our only therapeutic objective is to make the sick patient better, so this may or may not include "naming" the disease.
The elucidation of the causative, structural and functional changes may not come in any particular historical order, but the paradigm has two characteristics: first, it is expected or at least hoped the relations will be specific (unique cause, unique structural and functional changes belonging to one syndrome); second, as knowledge progresses, the defining process is 'pushed to the left' in the sequence given above. In other words, a disease will not be allowed to remain in syndromal terms if it can be explained or defined in functional terms; a functional syndrome will not be left in these terms if it can be characterised structurally, and 'cause' takes priority overall."
Algorithmic approaches to diagnosis are exemplified by programs for evaluating acid-base disorders8 and comatose patients.9 If an algorithm is sufficiently simple there is no point in computerising it; it is quicker and simpler to follow a printed version of the original algorithm.
THE BLUNDERBUSS APPROACH The blunderbuss approach is the traditional method taught to medical students. They take a detailed history, examine the patient from top to toe, and then order every test that could conceivably have some bearing on the problem. Not until all the information is to hand do they try and work out what is the matter with the patient. This is done by fitting the pattern of abnormalities found either to textbook descriptions of diseases or to their own database of diseases in patients whom they have previously seen. Computer applications The nearest automated approach to the blunderbuss method is that of database comparisons.5 An interactive search is made of a large database of information on patients, looking for those who match the particular patient under consideration. Usually a match is first sought on a limited list of features, with the result that a rather large and inhomogeneous subset of matching patients is obtained. The number of features to be matched is then increased, and the subset usually becomes smaller. When a matching subset is obtained in which the disease diagnosed in all patients is the same the new patient is assumed to have the same disease and can then be added to the database. This method requires the accumulation of a large database free from errors, which is very expensive, yet there is no real concept within it of a list of possible diagnoses, each having a different probability of being true. Probably the most suitable application is inthe diagnosis of rare syndromes,' where there is a real problem of human memory and collation of small snippets of information from diverse sources.
This clinical approach is woefully unimaginative, cumbersome, and extravagant (as opposed to parsimonious). When applied to laboratory tests it wastes money, not only because many of the tests originally ordered are unnecessary but also because the more tests that are ordered the more likely it is that, by chance, one or more will turn out to be "abnormal" and start a wild goose chase of further tests to investigate the chance abnormality.
Thirdly, from the point of view of computer modelling (and this applies to expert systems in general) the question has to be asked whether the objective should be to mimic clinicians or to use computers to do the things that clinicians cannot do in their heads (such as multivariate analysis). I understand the motivation of those working with computers who wish to mimic the brain, but I as a doctor want a system that will do better than the best clinician. A machine that simply does what a clinician (even a superb clinician) does is simply not a very attractive proposition
speech recognition by computers became successful, when statistical analysis was chosen leaving the liguistic rules system.
The ultimate in statistical diagnosis is to make no prior assumptions about diagnostic categories but to let the data speak for themselves. This corresponds to syndromic diagnosis by computer. For obvious reasons this approach has been mostly applied to the diagnosis ofmental illness, in which there is considerable doubt about the validity of traditional diagnostic categories. On the whole, however, these multivariate methods (factor and cluster analysis) have proved to be disappointing.'
BRITISH MEDICAL JOURNAL VOLUME 295 21 NOVEMBER 1987
Conclusions No explanation of human diagnostic logic so far conceived has been entirely satisfactory, though study of the alternative models is extremely instructive. Similarly, no method of diagnosis helped by computers has been shown consistently to be superior to all others. This is an exciting field of research precisely because it is so wide open. The validation of approaches by artificial intelligence to diagnosis has been particularly scanty-either non-existent or based on fewer than 20 patients. It is essential that comparisons of alternative diagnostic aids' "Is2 should be carried out as stringently as are those at present for new therapeutic aids such as drugs. FJM is supported by the Vanderbell and British Heart Foundations.
No comments:
Post a Comment