The first attempts to improve care with electronic medical records began more than 20 years ago with the computerizing of guidelines for simple preventive care and for identifying abnormal test results and potential drug interactions [2, 3]. Over the ensuing two decades, computers have become much faster (by orders of magnitude) and much less expensive. Meanwhile, partly in response to increasing health care costs and research showing that medical practice varied greatly among geographic locations and practices [4], professional organizations and federal agencies began developing more sophisticated clinical practice guidelines [5]. The automation of early guidelines through computers improved health care delivery [6-8] and, occasionally, patient outcomes [9, 10]. Electronic medical records thus offer a way to efficiently improve and monitor the processes and outcomes of care.

The ability to implement practice guidelines using electronic medical record systems depends on having sufficient data. Comprehensive electronic medical record systems that can store long-term data and implement guidelines are still uncommon. However, as more processes in health care become computerized (for example, laboratories, pharmacies, and billing offices), more clinical data are being stored electronically. Emerging standards for data transmission [11] and coding [12, 13] will augment the building of comprehensive data repositories from disparate data sources.

Dictated clinical notes of patient encounters and textual reports of procedures and imaging studies are large, mostly untapped sources of important data. One way to electronically capture these data is by paying technicians to read, hand-code, and enter summary codes from such reports. One data-entry technician who is paid $30 000 can code and enter data for 100 000 reports a year. This cost ($0.30 per report processed) is less than 1% of the charge for these tests and procedures and compares favorably with the typical 6% to 8% overhead charged by most billing organizations. However, data-entry technicians require management (hiring, firing, training, and oversight) and introduce another source of error into patients' electronic records.

Diagnostic impressions can also be directly captured from physicians as they dictate reports for imaging tests, pathology specimens, and procedures. Although we currently use this method for cytology and pathology reports, it requires first that physicians use computers as their primary reporting medium and second that they limit their diagnoses and impressions to standard codes. Neither of these requirements is prevalent today. Yet, a wealth of useful information remains locked in free-text reports.

In this issue, Hripcsak and coworkers [14] from Columbia-Presbyterian Medical Center describe an important advance in our ability to obtain useful data from narrative reports. Their natural language processing software did as well as radiologists and internists in coding diagnoses from chest radiographs. Importantly, they did not erroneously assume that any single coding of the report was the "criterion standard" but rather relied on standard measures of interobserver agreement.

The authors stress that the information gleaned from clinical reports by their language processor is not meant to replace the physician. Rather, their software extracts data for their clinical alert system [15]. Limiting the role these data play in providing care is important because, for now and into the future, neither fast computers nor sophisticated programs [16] can distill all of the nuances of language contained in dictated reports. Although difficult to measure and perhaps impossible to reproduce, the physician's "gestalt" will remain a critical component of clinical decision making. Hence, the electronic medical record system at Columbia-Presbyterian also stores and displays the full-text reports of radiographs [17].

We do disagree with one stance taken by Hripcsak and colleagues. To avoid false-positive reminders and thus prevent physicians from losing confidence in the clinical alert system, they have programmed their language processor to err on the side of specificity. That is, they have used stricter definitions to define their diagnosis codes to reduce the number of false-positive alerts. We believe that the proper role of reminder or alert system is to prevent clinicians from overlooking details and clinical relationships. Physicians at our institution accept our more sensitive (and less specific) system as long as they know from the start that some of the reminders will be wrong and should be ignored [7, 18]. Moreover, the strength of the recommendation can be varied with the reliability of the data. Because malpractice suits and, more importantly, clinical mistakes are triggered more often by errors of omission than errors of commission [19], a stance favoring sensitivity would be more appropriate.

Hripcsak and colleagues studied only one type of radiograph (chest) in one environment (the hospital). To be more broadly useful, natural language processors will have to be specific to both the type of report (for example, radiographs, scintigrams, procedure reports, histories, and physical examinations) and setting (for example, inpatient, emergency department, outpatient). Obviously, much work needs to be done. But the task is finite, and "beginning at the beginning" necessitates taking the technology that Hripcsak and colleagues have developed (and we hope will continue to refine) and applying it first to the most common textual reports and their most common conditions. Some crossover of terminology between reports will occur (for example, some of the codes for describing chest radiographs will apply to the physical examination of the chest), and testing the reliability and validity of such systems will become more standardized (and easier). The authors' methods have added greatly to this end. We applaud their pioneering efforts and the Editors of Annals for publishing a paper that would normally be buried in a highly technical journal with a limited circulation. The success of this experiment in natural language processing is a small but distinct step toward realizing the lofty goals of electronic medical records [1].