NéoGanesh: a working system for the automated control of assisted ventilation in ICUs

https://doi.org/10.1016/S0933-3657(97)00025-0Get rights and content

Abstract

Automating the control of therapy administered to a patient requires systems which integrate the knowledge of experienced physicians. This paper describes NéoGanesh, a knowledge-based system which controls, in closed-loop, the mechanical assistance provided to patients hospitalized in intensive care units. We report on how new advances in knowledge representation techniques have been used to model medical expertise. The clinical evaluation shows that such a system relieves the medical staff of routine tasks, improves patient care, and efficiently supports medical decisions regarding weaning. To be able to work in closed-loop and to be tested in real medical situations, NéoGanesh deals with a voluntarily limited problem. However, embedded in a powerful distributed environment, it is intended to support future extensions and refinements and to support reuse of knowledge bases.

Introduction

There is a growing need for computerized systems able to assist clinical staff in decision making, especially in medical environments such as operating rooms or intensive care units (ICUs), where the flow of information is abundant, false positive alarms are common and life-threatening situations should be prevented. These intelligent patient monitoring systems must reason about complex situations under real-time constraints such as resource limitations and must guarantee a timely response. Building such systems is a challenging goal for the emerging research area of real-time artificial intelligence (AI) [31]and more specifically `adaptive intelligent systems' [22]. This paper describes an intelligent patient monitoring system for the automatic control of mechanical ventilation.

For a number of reasons, including the complexity of medical reasoning, interference from noise, considerations of liability, and social and cultural factors, most intelligent patient monitoring systems are open-loop systems with respect to planning and control. Sepia [38]for monitoring patients hospitalized in hemato-oncology departments is a good example of such a system. However, there are specific well defined medical problems, in particular planning drug therapy [11], where closed-loop systems can be proposed. Such closed-loop systems can further improve the management of patient care because they operate continuously on a daily basis.

Computers have been used in clinical practice for traditional tasks such as database management, data acquisition and physiological signal processing. Sophisticated systems which might provide advice in the choice of therapy and assistance in diagnosis are not yet widely used. Several such systems are described in the literature, but only a few of them are routinely used in clinical practice. `Intelligent' decision support systems will help the physician to make better clinical decisions but they have yet to be tested at the patients bedside.

Five years ago in collaboration with the ICU at the Henri Mondor Hospital (Créteil, France), we started a project to explore the feasibility of and the clinical interest in a system providing automated control of mechanical ventilation and the decision for extubation. We decided that such an intelligent system should be able to: (1) handle in real-time a huge mass of information about the patients state; (2) diagnose observed situations; (3) predict the evolution of the patients state; (4) construct action plans with prompt reaction in emergency cases; and (5) execute planned actions.

In this paper we report on how new advances in knowledge representation techniques, in particular the association of the object-oriented paradigm with production rules, temporal reasoning and distribution of knowledge, have been used to design the system. We report on the clinical results obtained with our prototype, called NéoGanesh, which has been used to ventilate a large number of patients in our ICU. To be able to work in closed-loop and to be tested in real medical situations, NéoGanesh deals with a voluntarily limited problem. However, embedded in a powerful distributed environment, it is intended to support future extensions and refinements and to allow reuse of the knowledge bases developed so far.

Our paper is structured as follows. Section 2defines the medical problem, details the different levels of control for the management of mechanical ventilation and briefly outlines several systems that address a similar problem. Section 3describes the main characteristics of the NéoGanesh system. Section 4and Section 5respectively detail the distributed architecture of NéoGanesh and how medical expertise is modeled. Section 6is devoted to the clinical results. Finally, we conclude on the interest of our approach and discuss future extensions of NéoGanesh.

Section snippets

The control of mechanical ventilation

The mechanical assistance provided for a patient with respiratory insufficiency must be well adapted to his or her physiological needs. The clinician must assess the respiratory comfort of the patient and the time-course of his or her ventilation and must set the ventilator parameters accordingly. A second task of the clinician is to reduce mechanical assistance gradually until the patient is able to breathe alone. This task is known as the weaning process.

During weaning from mechanical

The NéoGanesh system

NéoGanesh is based on the knowledge of ventilation management acquired by the clinical staff of the ICU at the Henri Mondor Hospital (Créteil, France). In contrast to other systems 23, 37, NéoGanesh deals with a voluntarily limited problem: only one mode of ventilation is managed by the system, i.e. pressure support ventilation (PSV) used to ventilate patients who can have spontaneous respiratory activity. These limitations allowed us to design a closed-loop system that controls the ventilator

A distributed architecture

Distributed artificial intelligence and especially multi-agent architectures provide a powerful paradigm for the modeling and the development of complex systems. It is based on the decomposition of systems into several interacting and autonomous entities. Recent applications show the growing interest in this paradigm in the medical domain 21, 27. According to the definition of Shoham [40]“an agent refers to an entity that functions continuously and autonomously in an environment in which other

Knowledge representation in NéoGanesh

Interpreting data over time is an essential task of diagnostic and control processes. The time course of a process, determined from the evolution of a set of representative parameters, is central to predicting its future behavior and to choosing actions over time to influence it. Since the early work in AI, many formal studies have been conducted about change and time representation 10, 46. However, the complexity of temporal constraint propagation algorithms have limited the integration of a

Evaluation

In NéoGanesh, both synchronous communication and asynchronous communication between agents are possible, although the latter has not been tested yet in a clinical environment. Preliminary results [21]showed that the response time (the elapsed time between the perception of an event and an action on the ventilator) and the reaction time (the elapsed time between the perception of an event by the DataProcessor agent and its processing by the Classifier agent) are shorter with the introduction of

Conclusion

NéoGanesh is based on current AI technologies: sophisticated knowledge representation and temporal reasoning in a distributed architecture. We have chosen an environment which combines actors, objects, and production rules. We fully exploit the well known mechanisms of object-oriented programming and by using the inheritance mechanism our system can be easily extended.

Preliminary clinical studies performed at the Henri Mondor Hospital have demonstrated that patients show less signs of

Acknowledgements

This work has been partially financially supported by Hamilton AG. We wish to thank Professor Jean-François Perrot for his stimulating advice and constant encouragement that significantly influenced this work.

References (47)

  • Y. Shoham

    Temporal logic in AI: semantical and onthological considerations

    Artif. Intell.

    (1987)
  • J.H. Strickland et al.

    A computer-controlled ventilator weaning system

    Chest

    (1993)
  • D.A. Tong

    Weaning patients from mechanical ventilation. A knowledge-based system approach

    Comput. Methods Prog. Biomed.

    (1991)
  • P. Van Beek

    Reasoning about qualitative temporal information

    Artif. Intell.

    (1992)
  • Temporal reasoning in medicine, Artificial Intell. Med. 8 (Special Issue)...
  • J.P. Briot, Actalk: a testbed for classifying and designing actor languages in the Smalltalk-80 environment, Proc. Eur....
  • L. Brochard et al.

    Inspiratory pressure support prevents diaphragmatic fatigue during weaning from mechanical ventilation

    Am. Rev. Respir. Dis.

    (1989)
  • L. Brochard et al.

    Improved efficacy of spontaneous breathing with inspiratory pressure support

    Am. Rev. Respir. Dis.

    (1987)
  • L. Brochard et al.

    Comparison of three methods of gradual withdrawal from ventilatory support during weaning from mechanical ventilation

    Am. J. Respir. Crit. Care Med.

    (1994)
  • J. Carrive, B. Louis, A. Harf, M. Dojat, BioPhony: An open system to measure the airway area by acoustic reflexion,...
  • M.C. Chambrin, C. Chopin, K.H. Mangalaboyi, Autoregulated inspiratory support system, Proc. 14th Conf. IEEE-EMBS, 1992,...
  • L. Chittaro and M. Dojat, Using a general theory of time and change in patient monitoring: experiment and evaluation,...
  • T.L. Dean, M.P. Wellman, Planning and Control, Morgan Kaufmann, San Mateo, CA,...
  • Cited by (0)

    View full text