As the sophistication of diagnostic and monitoring devices grows, care of acutely ill patients requires rapid assessment and accurate interpretation of a large and ever-growing volume of medical data. Automated decision support systems capable of assimilating these numerous data streams, tracking and anticipating clinically significant changes, and presenting pathophysiologic hypotheses, would be of great value to clinicians facing ``information overload'' in the intensive care unit (ICU). We have begun to explore the design of ``intelligent patient monitors'' to accomplish these goals. To support this effort, we are creating the MIMIC Database.
Reference databases [1, 2, 3] are essential resources for developers and evaluators of algorithms and systems for analysis of physiologic data. The complexity of such data renders hopeless efforts to develop analytic techniques by pure reasoning. It is essential to test algorithms with realistic data, and to be able to perform these tests repeatedly and reproducibly as algorithm refinements are proposed. The requirement of reproducibility stems from the complexity of all but the most trivial analysis algorithms: without a strictly reproducible test, one can never establish with certainty if an observed difference in the results of analysis is attributable to a difference in the analytic method, or to a difference in the input data. It follows from this reasoning that a fair evaluation of an algorithm, whether against a standard or in comparison with another algorithm, also requires a reproducible test. Digital recordings are ideal test data for these purposes.
Experience with our earlier databases has also demonstrated their potential for stimulating basic research. By making well-characterized clinical data available to researchers, these databases make it possible to formulate and answer numerous physiologic questions, without the necessity of developing a new set of reference data at great cost in each case. Even non-specialists, who may lack other means of access to clinical data, can readily participate in such efforts. Since the databases are generally available, research results based on them can be readily replicated by others, fostering both healthy competition and wide-ranging collaborations with minimal barriers to entry.
These databases have also had value in medical education, by providing well-documented case studies of both common and rare but clinically significant pathologies. In clinical practice, however, it is common to have much more information to support diagnoses than has generally been available in existing reference databases.
In the current project, we have taken several steps beyond our previous reference databases. First, the recordings themselves are far longer than in our previous work. We sought to make the recordings long enough to be able to illustrate both the development of critical events and the results of interventions, in many cases several times in a single recording. Second, the MIMIC database includes comprehensive data from each subject's medical record, providing context for each recording. Although existing ICU monitors do not often make use of these data, it is clear that any process intended to provide competent support for medical decisions must be aware of all of what is known about the patient. Thus we regard this component of the database as essential for the task of developing intelligent ICU monitoring in the future. Third, although we have previously published a database of polysomnographic recordings containing blood pressure and other signals, the MIMIC database project is our first extensive and systematic effort to record hemodynamic variables together with multi-lead ECGs and other real-time signals.
We select patients from those considered likely to be hemodynamically unstable during the planned recording period. We chose to focus on this group of patients because their care in the ICU presents a particularly demanding challenge to physicians and nurses. The cardiovascular control system functions primarily as a regulator of blood pressure. Sudden changes in blood pressure hence signify a seriously compromised patient, whose control system is unable to react appropriately (whether by increasing cardiac output or by vasoconstriction) to a challenge such as hemorrhage, sepsis, trauma, medication, or cardiac arrhythmia (among other possibilities). These events require rapid intervention from the ICU staff, but the appropriate response depends on the cause. Thus, a timely and appropriate intervention requires rapid assimilation of a large amount of data from a diverse set of sources, including not only the current signals and recent trends from a bedside monitor, but also (for example) the patient's symptoms, medications, fluid balance, and response to previous therapeutic interventions.