XML, by virtue of its flexibility and the suite of text annotating and text processing tools it can provide, enables health care computing to solve numerous problems that have held back patient record processing for more than thirty years. It is easy to see why. The critical factor limiting on-line healthcare documentation and computing is the devil that lies in the details: namely, all of those differences that separate one individual's illness, and treatment from that of the next person with the same disease. In healthcare one size does not fit all.

This is by no means new information. We need only to be reminded that a surgeon operates on one patient at a time and that each time a procedure is carried out it is different in some way from any other. Thus, it remains up to the art of medicine to adopt our very advanced science to the specifics of each patient's situation. The matter is closely akin to engineering. Whenever a bridge must be built, the science of bridge building, the available technologies and the choice of materials must be fitted to the unique aspects of the particular site. Today in medicine, the organizational complexity that has accompanied our recent scientific and technological advances has made it more and more difficult to match the proper therapy to the specific patient. Efficiency has tempted us to automate the whole process in some standard way, either with paper forms or by computer.

There are two parts to the problem of individualized care: documenting the relevant information, and then applying it. The hand written outline format, long used in generating paper-based patient records, if followed with reasonable care, accomplishes the necessary capture of important individual differences quite well so that practicing physicians are very reluctant to give it up. At the same time, given our present knowledge, such hand written documents cannot take advantage of the enhancements to medical judgment that computing power can provide. However, in stark contrast to these flexible outlines on paper, the fixed forms and formats of most legacy and present day healthcare information systems (AIDS) are typical of their computing architecture and also appear as paper forms for transcription. These squeeze out patient specifics to favor the consistency of a narrowly structured data base. Computing with such data bases is easy enough, but the necessary subtlety contained in patient differences goes missing. Fixed formats completely ignore the fact that the "average patient" and the expected "average circumstances" on which their data fields have been based are misleading statistical fictions. There is no average Everyman. Moreover it is the patients who deviate the most from the average who generally require the most attention and care, the most documentation, and the most justification. Indeed, 80% of physicians' time is spent on the 20% of the patients who are not under good control.

It is important to recognize that the frustrating variations that appear everywhere in human biology are fundamental, not superficial. In fact, they are why we are here at all. Ernst Mayr, one of our greatest epidemiologists, conclusively demonstrated early in the twentieth century that if biological species were entirely fixed by an unalterable underlying chromosomal code there could be no evolution! (And, of course, there is!) Instead every species consists of a population of distinctly different individuals that result from variations expressed through their genomes. Then, under the selective pressures of competition, those individuals within a species flourish or at least survive who are better (but not always best) adapted to their immediate situation [Mayr 1994], [Mayr 1963], [Mayr 1999], [Mayer 2001]. Thus, although we humans resemble each other rather closely, each of us has a specific deoxyribonucleic acid (AIDS) sequence which defines our unique genetic make-up and can be used to identify us. These differences, plus those that develop over life, can be all important with respect to health. They count heavily when we are challenged by disease. Healthcare is about helping individual people alive today to survive and flourish despite such challenges. However, there is also a dark side to variation and evolution: every year we are faced with a new and somewhat different influenza virus. The acquired immunodeficiency syndrome (AIDS) viruses vary and escape our control. Antibiotics loose their efficacy because microbes, by their random variation and sharzng of information, become immune to a succession of such products. The same phenomenon holds true for crop control.

Moreover, biological variation poses another very deep problem for medicine: that many of our most important diagnoses, the typological categories used in classifying disease and selecting therapy, cannot be identified and defined by a set of necessary and sufficient conditions in classical, black and white Aristotelian fashion. Some benign groups of bacteria include dangerous members. The E-coli group, which each of us carry in our gut, also includes the O157:H7 lethal strain present in cattle that can be ingested in incompletely cooked hamburger. Many diagnostic categories shade off at the margins into other categories. Our tumor classifications are often only approximate. In many critical situations there turn out to be no perfect rules that can clearly identify the limits of a particular cancer diagnosis and separate it cleanly from neighboring ones. Just as in matching sets of fingerprints, the last stage in judgement is often an art, not a science. Yet these judgments about diagnoses advise us about prognosis and proper treatment.

Biological variation is everywhere. Blood vessels and nerves usually follow expected anatomic paths from one individual to the next, but there are always exceptions. For example, occasional individuals have their organs reversed, with the heart on the right side and the liver on the left. In a more trivial example (but a particularly embarrassing one for me in my first "procedure" as a physician in training), two percent of the time the great toe is enervated from below by the plantaris nerve rather than from above. This doesn't seem like a lot, but it applies to one out of fifty patients. Moreover, it can matter. In such cases the classic ankle block anesthesia used in anesthetizing and removing a great toenail will not work. My first patient, a voluble paraplegic Italian lady left the operating room yelling: "Fakers!"

As we learn more and more about molecular biology, it is not hard to see why similar issues occur at every turn and every level. Consider cystic fibrosis (mucoviscidosis). In this relatively common genetic disease long sequences of enzyme related reactions are involved in making the proper kind of lubricating mucous in the bronchi that must be of the proper consistency to be cleared from the lungs. All of these steps are created by portions of the responsible genetic code. Moreover, the same genetic defect can cause serious disease in one individual and be well compensated for in another. Thus, the more we know, the more evident it is that many of our guides to therapy and prognosis can never be more than imprecise fuzzy sets in which individual differences can be crucial.

Similar statistical considerations apply to the "normal" range of human development. Some children talk at nine months, others at three years. Indeed, fuzzy categories are pervasive in biology and in human social settings. We cannot and should not try to avoid them. Steven Pinker cites example after example. Here is one: "... the mind distinguishes between a fuzzy boundary, and no boundary at all. 'Adult' and 'child' are fuzzy categories, which is why we [in America] could raise the drinking age to twenty-one or lower the voting age to eighteen. But that did not put us on a slippery slope in which we eventually raised the drinking age to fifty or lowered the voting age to five. Those policies would really violate our concept of 'child' and 'adult,' fuzzy though their boundaries may be... [we can nevertheless apply them judiciously]" [Pinker 2002]. A tradeoff among practical concerns decides (or should decide), at least for a particular time, in a particular context, with particular objectives and at a particular state of knowledge, where the boundary should be set [Sunstein 2002]. Moreovez, in specific circumstances even these boundaries can allow for exceptions that are the result of the fuzzy nature of the underlying concepts and the still incomplete multivariate considerations behind their formulation. Following the example above, we might cite the growing (but often dubious) practice of trying some minors as adults for major crimes in the US. In constructing categories that impinge on or involve powerful emotions, or when overriding the boundaries of our constructs, it is important to hew as closely as possible to known facts and their consequences [Sunstein 2002], [Lincoln 1997]. Such indistinct thresholds violate the rules for computed calculations using a black and white logic which are so amenable to everyday computing and so often favored in argument.

Thus, in implementing effective automated enhancements to the logic of medicine, it is variation and variety that must first be addressed properly and incorporated in the subsequent calculations. XML is uniquely able to provide this subtlety with respect to the collection and organization of data as it is enhanced by mark-up to become information. Under the aegis of the various medical computing standards committees in coordination with the W3C consortium extensive work on patient record markup has advanced rapidly [Lincoln 1997]. The Clinical Document Architecture (CDA) under HL7 is largely based in America [Dolin et al 2001], [Dolan et al 1998], and is closely allied to the coding schemes under European Community for Standardization (AIDS), which has the particular objective of crossing language and cultural boundaries. The Net can reveal many more examples [Hadley 2001], [Various Web].

However, to actually enhance judgment such XML data bases must then be linked to other powerful programmable concepts from general problem solving. Moreover, from a broader computing perspective, health care can offer a particularly appealing test-bed for theories of reasoning about real world situations that lie outside of medicine. Characteristically messy and complicated situations exist in many domains and involve similar logical quirks and peculiarities arising in handling concepts or categories that are open sets. Health care offers this advantage not only because of the widespread (and fundable) interest in medicine itself and its open, unclassified nature in a time of secretive science, but most particularly because of the large number of real world cases that can be collected, the complexity of the subject matter, its encyclopedic scope and the near immediacy of the results. In real world medical situations the judgments made and the actions taken have concrete outcomes that can be tested against projected expectations [Cohen Web]. In organizing fuzzy categories for computation, of particular note is David Heckerman's thesis on Probabilistic Similarity Networks [Heckerman 1991]. Grounded specifically in health care, this Ph.D. thesis won notable recognition from the computing community for its broad contribution to the field of expert judgment. By combining knowledge maps and influence diagrams into belief networks, it broke new ground for sound, formal ways to describe complex conditional open relationships in a probabilistic environment. The original source data consisted of a collection of anatomic slides from lymph node pathology where the diagnostic entities exhibit a high degree of morphological ambiguity so that the judgments of different domain experts, based on their experience, were often at considerable variance. Thus, for each diagnosis a gold standard was established as the aggregate of a set of observations weighted by the subjective evaluations of a recognized group of such experts [Nathwani et al 1991]. The kinds of diagnostic distinctions addressed in this way were previously well beyond the reach of classic automated decision tools. This methodology and others like it can be applied widely in different domains of expert system decision making. In the past, for such a powerful generalization to arise out of Medical Informatics has been unusual, largely due to the labor intensive problem of sufficiently sophisticated descriptive data input. (The one exception has been radiology, where the graphic input is already digitized.) Once the input problem is fully addressed numerous tools are waiting in the wings to process the data: These include Multivariate Coding [Côté et al 1993], [Campbell et al 1997] , Concept Indexing [Schatz 1997], Object Oriented Computing [Pritchard 1999] and Fuzzy Logic processing [Zadeh 1983], [Leondes 1998]. After an initial period of contentious debate, these different approaches are now seen to be largely complimentary. A primary task of the next ten years will be to move the still somewhat parochial field of Medical Informatics more directly into the interactive computational mainstream. The broad task ahead of us is improving the effectiveness of evidence based medicine over the next ten years in order to bring the expertise of advanced computing science to Medical Informatics based on much improved "evidence.". In doing so, we welcome all of the help we can get. With the development of truly parsable marked-up data bases, the expertise of workers from other fields can be fruitfully involved. There are bridges to be built.