Dimension of Time in Illness: An Objective View

As a professional medical informatician whose main interest is harnessing artificial intelligence and other techniques of information sciences to help both health care providers and patients, I am repeatedly impressed by the tasks that clinicians achieve on a daily basis. Although these tasks require extensive knowledge about both medicine and the world at large, clinicians rarely express any sense of wonder at their accomplishments and usually take them in stride. I can only hope that my computer programs will some day be able to carry out a small fraction of these tasks. I am also repeatedly struck by the fact that clinicians often perform tasks involving large amounts of clinical data without the assistance of computers, although such assistance would be highly beneficial. Indeed, in my view, the science fiction-like, futuristic idea that computers will eventually replace physicians misses the mark. The best term to use when pondering the relation between clinicians and computers may well be synergy. A case in point, and the focus of my work over the past decade, is the role of time in medical care. In particular, I am interested in the implications of the temporal aspects of diseasesfor example, their courses and management. I seek to improve collaboration of care providers and information systems to provide optimal care to patients over time. It is almost impossible to try to represent and analyze clinical data without including a temporal dimension. Clinical interventions must occur at one or more time points (for example, Laparoscopy was performed on 9 June 1998) or over periods (for example, Angiotensin-converting enzyme inhibitors were administered from early February 1992 to late May 1992). Similarly, patient characteristics and measurements (such as results of laboratory tests, findings on physical examination, or a diagnosis) must be noted during time points or periods. Various qualitative and quantitative temporal relations, such as a high fever occurring after immunization for mumps, can exist between measurements or interventions. Periodicity (and, in general, any form of a repetitive temporal pattern) is inherent in our clinical thinking (for example, Over the past 6 months the patient experienced nausea three times after ingesting fatty foods). Time is important for representing information within an electronic medical record system and for querying medical records; it is also important when time-oriented clinical data are considered as part of various decision support applications, such as determining a diagnosis, prescribing therapy, and browsing electronic patient records for management or research purposes. Representing, querying, and analyzing time-oriented clinical data are crucial tasks for care providers and for automated decision support systems; both need to extract certain information from one or more patient records. For example, during treatment of a patient based on an experimental chemotherapy protocol, the care provider or an intelligent (computer-based) therapy support system may need to refer a complex temporal query to the patient's record. Such a query may ask whether the patient had more than two episodes of grade II or higher bone marrow toxicity (as defined by the experimental protocol), each lasting at least 3 weeks, within the past 8 months. Examples of tasks that depend heavily on such access to time-oriented data and their interpretation include patient monitoring, management of patients by using therapeutic guidelines, and interactive visualization and exploration of longitudinal patient data. It is useful to distinguish two research directions that are distinct with respect to their focus (and, of interest, the research communities pursuing them) and are relevant to the temporal dimension of patient illness: temporal reasoning and temporal data maintenance. Temporal reasoning supports various inference tasks that involve time-oriented clinical data and inference, such as monitoring patients, diagnosing disorders, and planning and applying therapy. Traditionally, the temporal reasoning task has been investigated by researchers in the artificial intelligence community. Temporal data maintenance deals with storage and retrieval of clinical data that have heterogeneous temporal dimensions. The temporal maintenance task is usually associated with the temporal database research community, which is often separate from the artificial intelligence and medical communities. The goals of these two disciplines, however, have enough in common to justify their integration in medical information systems. This integration is particularly important for improving the care of patients with chronic diseases. In this paper, I will demonstrate several ways in which these improvements might be achieved. Temporal Reasoning: Interpretation of Time-Oriented Patient Data To determine the course of an illness over time, patient data must be measured and stored, often in electronic media. Ironically, the almost limitless capability of these storage media poses new problems for care providers who must make diagnostic or therapeutic decisions on the basis of voluminous amounts of data. From this point of view, automated assistance in reasoning with the data would be welcome. Experienced clinicians, unlike well-meaning but sometimes misguided computer scientists, know that arriving at a diagnosis is not always the main goal of analyzing the patient's data and therefore is not necessarily the task on which information systems should concentrate. A coherent intermediate-level interpretation of the relations between data and events and among data is often necessary. This is especially true when the overall context (for example, a major diagnosis) is already known, as is the case for most patients seen by clinicians, particularly chronically ill patients. For example, if a system serves only to inform a general practitioner that her patient has type 1 diabetes mellitus, which she and the patient have known for the past 20 years, it will probably not be of substantial assistance to her or increase her confidence in medical information systems. Busy clinicians often need to reduce information overload. Information systems can accomplish this by abstracting clinical data, much of which is acquired or recorded as time-stamped measurements, into higher-level concepts that are relevant to a particular clinical situation. These abstracted concepts, or clinical states, will often remain constant over extended periods. Therefore, the goal of an automated support system in such cases is to create interval-based time-related abstractions (for example, periods of bone marrow toxicity) from time-stamped clinical data (for example, hematologic measurements in the context of a chemotherapy protocol). An oncologist may be treating patients who have graft-versus-host disease, a complication of bone marrow transplantation, with a prednisone-azathioprine protocol (Figure 1). The definition of grades of bone marrow toxicity depends on the context in which these grades are assessed. Therefore, the physician may query her automated assistant for periods of longer than 2 weeks in which the bone marrow toxicity level was grade I or more, as defined within the prednisone-azathioprine protocol. The task of creating temporally extended concepts and patterns from raw time-stamped data is often called the temporal abstraction task. (The term temporal abstraction is somewhat misleading because it is the data that extend over time, not the time itself, that are being abstracted.) Figure 1. Temporal abstraction of platelet and granulocyte values during administration of a prednisone-azathioprine ( PAZ ) clinical protocol in patients with chronic graft-versus-host disease ( CGVHD ). Another example of the need to interpret patient data over time can be seen in monitored patients with type 1 diabetes mellitus. For example, a physician may need to realize that a patient could be getting an insufficient amount of intermediate-acting insulin in the morning. This would result in normal morning blood glucose levels and high blood glucose levels during the afternoons or evenings (Figure 2). This kind of repetitive, periodic pattern is much more significant than a single observation. It should be noted that periodic or episodic patterns may occur on a daily, weekly, or monthly basis (for example, hyperglycemia could recur after lunch on most weekends). Figure 2. Monitoring a patient with type 1 diabetes mellitus ( DM ). A system's ability to automatically create interval-based abstractions of time-stamped clinical data has many implications for the care of patients over extended periods. Data summaries of time-oriented electronic patient records could be valuable to a human user, such as a busy physician scanning a long patient record for meaningful trends (1). Temporal abstractions can be used to support recommendations by intelligent decision support systems, including both diagnostic and therapeutic systems (2), and to monitor therapy plans (for example, clinical guidelines) during their application (3). Use of meaningful time-related states (for example, post-bone marrow transplantation and therapy with the prednisone-azathioprine protocol) enables systems to generate abstractions that are specific to a particular clinical context. In addition, they allow several interpretations of the same data to be maintained within different contexts and enable support of hindsight and foresight (4). Temporal abstractions can often help explain actions recommended by a decision support system, especially those that are determined on the basis of concepts taken from the patient's record. In addition, temporal abstractions are a useful way in which to express the process and outcome intentions of designers of clinical guidelines; they also enable systems to develop real-time and retrospective critiques and quality assessments of care providers' applic