OLDER PROJECT · AI IN INTENSIVE CARE · EARLY 1990s ONWARD
From a research system that encoded medical expertise to an automated weaning protocol integrated into Dräger ventilators.
In the 1990s, I worked with Michel Dojat and colleagues on NéoGanesh, a knowledge-based system for adjusting assisted ventilation and supporting the gradual withdrawal of ventilator assistance in intensive care. Michel Dojat later carried this work forward with Dräger, leading to the commercial SmartCare/PS system.
Making medical expertise executable
Weaning means progressively reducing mechanical breathing assistance as a patient recovers. It calls for repeated assessment: enough support to avoid excessive respiratory effort, with opportunities to reduce assistance when the patient can tolerate it.
NéoGanesh addressed this through a closed feedback loop. It interpreted physiological measurements, assessed the patient’s respiratory state, and adjusted the pressure support supplied by a connected ventilator. Its reasoning combined clinical rules, an explicit treatment of time, and a strategy for progressively reducing assistance. The system was developed and evaluated with teams at Hôpital Henri Mondor in Créteil and INSERM.
A simplified view of the research system’s feedback loop. See the 1997 system paper (PDF) for the architecture and evaluation.
On the AI side, the early implementation used Smalltalk and the NéOpus rule engine. An important research question was how to combine objects and production rules so that medical knowledge could be adapted to different clinical contexts and reused within a defined domain. The joint papers below document both the working system and these representation techniques.
From NéoGanesh to Dräger SmartCare/PS

The connection is documented by the people who developed the product. In their 2004 ECAI paper, Stefan Mersmann of Dräger Medical and Michel Dojat describe SmartCare as a system for executing clinical guidelines within medical devices. They explicitly build on the earlier academic ventilator-weaning prototype and cite the 1997 NéoGanesh paper.
The industrial step involved more than transferring a prototype: the paper describes a methodology for modelling clinical guidelines, generating code and integrating it into a target device. Ventilator weaning was its first application.
A direct trace in the manufacturer’s documentation. Dräger’s SmartCare/PS instructions for Evita V800/V600, bibliography, p. 48 (PDF, Norwegian edition), include our 1997 NéoGanesh article alongside Dojat’s clinical studies. The early research remains part of the product’s documented scientific background.
SmartCare/PS is a commercial Dräger system available on its Evita V series. It monitors respiratory parameters, adjusts pressure support and proposes separation from the ventilator when its protocol criteria are met. The clinician evaluates that recommendation and decides whether to extubate. Dräger’s English SmartCare/PS booklet (PDF) explains the product’s operation.
Evidence of clinical usefulness
A major milestone was the 2006 multicentre randomized trial by Lellouche and colleagues, including Dojat and Mersmann. It compared computer-driven weaning with physician-controlled usual care in 144 patients.
| Outcome | Usual care | Computer-driven weaning |
|---|---|---|
| Weaning until successful extubation | 5 days | 3 days |
| Total mechanical ventilation | 12 days | 7.5 days |
| ICU stay | 15.5 days | 12 days |
The trial found no statistically significant difference in reintubation rates.
For me, this is a particularly meaningful outcome of the early collaboration: research on representing expertise became part of a medical product, supported by a substantial programme of clinical evaluation.
Joint papers with Michel Dojat
The NéoGanesh system papers are collected here with the related work on context, reuse and temporal reasoning. Links lead to full text where available, or to a bibliographic record.
- NéoGanesh: a working system for the automated control of assisted ventilation in ICUsMichel Dojat, François Pachet, Zahia Guessoum, Dominique Touchard, Alain Harf & Laurent Brochard · Artificial Intelligence in Medicine 11(2), 97–117, 1997.
The main account of the system: knowledge representation, distributed architecture, closed-loop control and clinical evaluation.
PDF · DOI · Publication page - An extendable knowledge-based system for the control of mechanical ventilationMichel Dojat & François Pachet · IEEE Engineering in Medicine and Biology Society, Paris, 920–921, 1992.Conference paper (PDF) · Extended bilingual manuscript (PDF)
- Representation of a medical expertise using the Smalltalk environment: putting a prototype to workMichel Dojat & François Pachet · TOOLS 7, Dortmund, 379–389, 1992.Bibliographic record
- NéoGanesh, un système extensible à base de connaissances pour la ventilation artificielle des patientsMichel Dojat & François Pachet · Innovation et Technologie en Biologie et Médecine 14, 267–279 · listed as 1992 in the authors’ bibliographies.HAL record
- Effective domain-dependent reuse in medical knowledge basesMichel Dojat & François Pachet · Computers and Biomedical Research 28(6), 403–432, 1995.
How object-oriented and rule-based mechanisms support limited, practical reuse of medical expertise.
PDF · DOI - Representing medical context using rule-based object-oriented programming techniquesMichel Dojat & François Pachet · AIME ’95, Lecture Notes in Artificial Intelligence 934, 423–424, 1995.Publisher’s proceedings and chapter listing
- Three compatible mechanisms for representing medical context implicitlyMichel Dojat & François Pachet · IJCAI ’95 Workshop on Context in Knowledge and Reasoning Modelling, 1995.PDF
- Un framework pour la représentation de connaissances temporelles par objets et règlesFrançois Pachet & Michel Dojat · LAFORIA-IBP research report 1995/19, 38 pp., June 1995.HAL record
- Consistency of temporal constraint networks for medical monitoring: binary or nary constraints?Alain Liret, Nicolas Ramaux, François Pachet, Dominique Fontaine & Michel Dojat · ECAI ’98 Workshop on Non-Binary Constraints, Brighton, 1998.
Related work on recognizing a patient’s evolving situation using temporal constraints, in the DéjàVu system.
PDF
Two key follow-up papers
The industrial implementation and the multicentre clinical trial mark the two main steps from research prototype to medical product.
2004 · From academic prototype to industrial system
SmartCare™ — Automated clinical guidelines in critical careStefan Mersmann & Michel Dojat · ECAI 2004, Valencia, 745–749.The key account of the industrial transition, written jointly by Dräger Medical and INSERM researchers.
Author-posted text · Bibliographic record2006 · Multicentre randomized evaluation
A multicenter randomized trial of computer-driven protocolized weaning from mechanical ventilationFrançois Lellouche, Jordi Mancebo, Philippe Jolliet, Jean Roeseler, Frédérique Schortgen, Michel Dojat, Belén Cabello, Lila Bouadma, Pablo Rodriguez, Salvatore Maggiore, Marc Reynaert, Stefan Mersmann & Laurent Brochard · American Journal of Respiratory and Critical Care Medicine 174(8), 894–900.The 144-patient trial summarized above is the central clinical milestone in this story.
DOI · Open-access full text · PubMed