Clinical database variables can be either categorical (nominal) or continuous. Categorical variables may be limited to only two possible values and thus be dichotomous; examples include yes or no, 0 or 1, male or female, and presence or absence of chronic heart failure. Alternatively, categorical variables may be ordinal, that is, have a limited list of possible values that bear an ordered relation to one another; examples include New York Heart Association functional class and severity of mitral valve regurgitation (6). Finally, categorical variables may be polytomous, that is, have an unordered but finite list of possible values, such as etiology of subacute bacterial endocarditis or names of bones in the body. Continuous variables can hold an infinite number of values; examples are age and systolic blood pressure.

If data are to be understood, valid, and reliable, it is essential that all data elements be defined accurately. A common means of accomplishing this is to create a separate computer file that contains the list of all files in the database, a list of all variables, definitions and data type of each variable, and lists and definitions of all possible values with each variable. Such a file is known as a data dictionary (7).

With the advent of worldwide exchange of information, attention has focused on data definitions and formatting that are independent of specific machine types and operating systems using internationally agreed on standards. Data definitions are now more fully expressed in terms of metadata. Metadata are “data about data” that contain all structural attributes of the data and database. Metadata may be separated from the data to form part of the structure of the database, in which case data elements are simple, consisting of just the value, or metadata may be directly associated with each data element, in which case data elements may be complex (semistructured data). Metadata may be further organized with interrelationships to become a knowledge base or ontology.

Determination of the size, structure, and content of clinical databases has been a major source of frustration, in part because no standards for these existed before the modern computer age. Ellwood described certain rules that should be followed if a large clinical database is to remain robust over time (8). The goals of assembling the database have to be prospectively determined to define appropriate data elements. The size of the database must be carefully considered; large databases contain more information, but excessively large numbers of data elements can lead to significant deterioration of data quality. Although data may be collected in many geographic locales, a mechanism for centralizing data is essential. Finally, the ideal clinical database is not confined to hospital events. It is the frustration with the predefinition of databases, which presupposes the relations and types of questions likely to be asked of it in the future, that has stimulated development of infinitely extensible databases, such as exemplified by the World Wide Web and semistructured databases.

For clinical data obtained outside the confines of a clinical trial or registry to be considered reliable for analyses, it is essential for standards to be developed to allow data fields to be consistent between different databases and to allow data to be transferred between systems (9). Numerous coding systems (such as ICD-9, SNOMED, Unified Medical Language System [UMLS], CPT, and others) for clinical databases have been developed to represent clinical classifications and nomenclatures (10,11).

A database “design” refers to the model according to which data fields are logically linked to one another (12); the most commonly cited database designs are flatfile, spreadsheet, hierarchical, relational, and object oriented. The simplest design is a single table of values, called a flatfile. A flatfile is non–self-documenting, requiring a separate data dictionary to accompany the data file. The next-simplest design is the spreadsheet. It is superior to the flatfile not only in containing self-documentation, such as column headings, but also in allowing in-place associations to be established among the data cells. A major limitation of flatfiles and spreadsheets is that they have primarily been proprietary products, and they allow entry of data values in an unstructured fashion that may permit human understanding of the data but not computer understanding necessary for analysis. Another early design was the hierarchical model, which consisted in linking multiple collections of similar information into a tree structure, with root and child nodes as branches and ending with leaf nodes. A major disadvantage of this database design is that hierarchies are arbitrary and limiting.

A database model that ties together multiple tables of information and provides more flexibility is the relational model. Relational databases allow for efficient, powerful storage of large quantities of information (13,14). Tables within a relational database are linked by user-defined cross-references or “linking variables” (Fig. 44.1). Relational databases are, however, static repositories. Imagine a database that could note that an entered value for a glucose tolerance test was well above the normal range and therefore set into motion a predetermined sequence of activities, such as paging the physician, printing a diabetic diet based on patient characteristics, discovering a clinical trial of pancreatic islet transplantation, or a host of other functions. This is possible in object-oriented databases. Objects contain not only data, but also instructions on how to react to certain data values.

Although the relational model is the most pervasive in clinical databases today, two other challengers deserve mention. One is an old technology, time-oriented databases, and the other is new, semistructured databases. Semistructured databases permit the self-assembly of an ad hoc structured database at the time of an analysis using information embedded in the data elements themselves. This is the exact opposite of the traditional approach to structured databases, in which the database is fixed and, after some lapse of time, queried and an analysis performed.

As described by Safran (19), there are four fundamental uses of clinical databases: (a) results reporting, (b) case finding, (c) cohort description, and (d) predictive modeling. An additional use that has gained much recent attention is for assessment of medical care quality (20). Results reporting refers to the ability to retrieve information about a specific patient or procedure; this is the type of activity by which practicing clinicians would most often interact with, or at least “seek rewards” from, a clinical data repository. Case finding involves looking for patients with similar problems or outcomes as a patient of interest, often to direct management based on prior individual experiences. Cohort description seeks to simply describe a group of patients or procedures based on a common set of attributes. Prediction modeling involves applying sophisticated analytical methods to describe changes or associations that can lend insight into the natural history of disease processes, the effects of different clinical interventions, and/or quality of care within given institutions or by specific providers. It is this use of databases that presents the most powerful, and perhaps also the most controversial, benefit of computerized clinical information technology.

In the 1988 Shattuck Lecture to the Annual Meeting of the Massachusetts Medical Society, Paul Ellwood advocated the nationwide adoption of outcomes management, which he described as a new “technology of patient experience” (8). Work by Wennberg and others (21,22,23) that showed tremendous variation of practice patterns without clear differences in patient outcome has caused the public and third-party payers to demand better-quality clinical effectiveness information. Physicians, too, are frustrated by their inability to accurately and quantitatively predict the effect of their care on patient outcomes because good-quality data from randomized, controlled trials simply do not exist for many common clinical scenarios (24). To help to solve these problems, the process of outcomes management would seek to use high-quality data based on real patient experiences to help physicians, payers, and patients make the most rational treatment choices.

A major component of outcomes management is the determination, based on analyses of large observational clinical databases, of how outcomes are associated with physician plan and hospital execution factors after “adjusting for” patient factors. Some have argued, however, that using observational databases to assess the effect of therapies or behaviors under physician or patient control is inherently flawed and that only randomized trials can appropriately answer these questions (25). The main strengths of randomized, controlled trials include (a) assembly of a single, well-defined inception cohort; (b) elimination of potential biases and confounding, particularly those caused by unknown or unmeasurable factors; and (c) adherence to well-defined clinical care and data collection protocols with vigorous quality assurance and safety mechanisms in place (26). However, as reviewed by Califf and colleagues (27), clinical trials suffer from several inherent limitations: (a) they are expensive and sometimes very time consuming; (b) therapies change such that by the time trial results are reported, new technologies may have become available that may replace the methods tested, (c) many patients, particularly those with comorbidities or the elderly, are excluded, (d) trials are performed more often in patients at tertiary referral centers, raising generalizability questions, and (e) the benefits of a therapy as noted in the ideal conditions of a clinical trial may not necessarily apply to the “real world” of clinical care. Furthermore, some clinical treatment questions simply cannot be answered with randomized trials. For example, a recent study found that a prolonged door-to-balloon time was associated with higher mortality among patients undergoing primary angioplasty for acute myocardial infarction (28). It is hard to imagine that the validity of this finding could be tested by a controlled trial in which patients would be randomized to an immediate versus delayed angioplasty or even to a series of prescribed intervals (24).