Adam Seiver

Uncertain reasoning and forecasting

Abstract We develop a probability forecasting model through a synthesis of Bayesian belief-network models and classical time-series analysis. By casting Bayesian time-series analyses as temporal belief-network problems, we introduce dependency models that capture richer and more realistic models of dynamic dependencies. With richer models and associated computational methods, we can move beyond the rigid classical assumptions of linearity in the relationships among variables and of normality of their probability distributions. We apply the methodology to the difficult problem of predicting outcome in critically ill patients. The nonlinear, dynamic behavior of the critical-care domain highlights the need for a synthesis of probability forecasting and uncertain reasoning.

Distributing Intelligence within an Individual

Distributed artificial intelligence (DAI) refers to systems in which decentralized, cooperative agents work synergistically to perform a task. Alternative DAI models resemble particular biological or social systems, such as teams, contract nets, or societies. Our DAI model resembles a single individual, characterized by adaptability, versatility, and coherence. The proposed DAI architecture comprises a hierarchy of loosely coupled agents for specific perception, action, and reasoning functions, all operating under the supervision of a top-level control agent. We demonstrate the proposed architecture in the Guardian system for intensive-care monitoring.

Guaranteeing real-time response with limited resources

Unanticipated problems detected by patient-monitoring systems may sometimes require real-time response in order to provide high-quality care and avoid catastrophic outcomes. In this paper, we present an approach for guaranteeing a response to such events by a monitoring agent even in situations where we have limited problem-solving resources. We show that an action-based hierarchy can accomplish this goal. We also analyze the performance of this hierarchy under varying resource availability and discuss decision-theoretic approaches to enable us to best structure such a hierarchy. We also describe an implementation of these ideas, called ReAct, in the BB1 architecture. All the ideas are illustrated with examples from the surgical intensive care unit (SICU).

Decision Analysis: A Framework for Critical Care Decision Assistance

The ultimate goal of medical computer systems is to help clinicians make good decisions. Such systems must be based on sound principles. Decision analysis is a 25-year-old discipline that provides the needed rigorous foundation for decision assistance. Decision analysis comprises the philosophy, procedures, and tools that can correct the flaws in existing critical care decision-making practice. Intelligent decision systems--computer-based systems that automate decision analysis--make it practical to apply decision analysis to critical care. Orchestra is a pilot intelligent decision system (now under development) that coordinates the efforts of the critical care specialist, the bedside physician, and the bedside nurse in building decision models that can provide recommendations and insight for ventilator management decisions. Decision analysis delivered by intelligent decision systems has great potential for improving critical care decision-making.

Regular low frequency cardiac output oscillations observed in critically ill surgical patients

A serendipitous finding during development of an automated “electronic flow chart” system to gather data on ICU patients [1] was the observation of low frequency oscillations in blood pressure that were not explained by systematic variability in the environment. [2] This finding has since been confirmed by others. [3,4] In the present report, hemodynamic data for critically ill surgical patients was continuously collected and visualized on a computer workstation to search for patterns not noted by standard monitoring. With this system, we observed low-frequency periodic oscillations in the cardiac output of ten patients, with regular periodicities of 4 to 280 minutes (average = 34 minutes). The mortality rate in these patients was 40%, while the mortality was only 10.8% in 83 similarly monitored intensive care unit (ICU) patients who did not develop regular oscillations in cardiac output. Interestingly, these oscillatory patterns appear to be associated with inadequate resuscitation of increased metabolic rates. The mathematical definition of “chaos” refers to irregular behavior that appears to be random but is actually deterministic. [5] A surprising finding concerning transitions between states of apparent randomness and order in nonlinear systems is that many systems become more organized after being disturbed. Chaotic behavior in biological systems may represent a normal physiologic state, while the loss of chaotic behavior may herald a pathophysiologic state. [6] The mechanism of the regular low frequency oscillations in cardiac output remains to be determined, but the high mortality rate suggests that it is a pathophysiologic marker, perhaps due to inadequate oxygen delivery in under-resuscitated shock.

A decision class analysis of critical care life-support decision-making

Decision analysis is a powerful methodology that can help clinicians make good decisions. Because it is not practical to place a decision analyst at the bedside in critical care units, the application of this methodology will require leveraging the analyst through computer-based systems. A decision class analysis is a collective analysis of a group of decisions that provides the high-level specification for such a computer system. This paper presents a decision class analysis of critical care life-support decisions. Key elements of this analysis are: the simplification of an otherwise extremely complex multistage sequential decision problem by using a sequence of two-stage models, and the use of six generic knowledge maps that capture the extremely complex relevant medical knowledge.

Capnography: Capnography in non-invasive positive pressure ventilation

The usually more controlled circumstances of airway management in the operating room (OR) often provide better conditions, better monitoring, and more experienced personnel, particularly when a problem occurs, than is available in other critical care environments or the emergency department. While the detection of CO2 by capnography after completion of a difficult intubation procedure may suggest success, it may more precisely indicate only that the tube tip is somewhere in the respiratory path, although perhaps not exactly where the intubationist desires. A capnography pattern indicating declining CO2 in each subsequent breath over several breaths will help identify esophageal intubation. Unilateral pathophysiologic conditions that cause unilateral hypoventilation or high airway resistances would result in a biphasic waveform. Many techniques to facilitate blind nasal tracheal intubation use the detection of significant exhaled gas flow from a spontaneously breathing patient to indicate the proximity of the tube tip to the glottic opening.