SETH: an expert system for the management on acute drug poisoning

Jean-Michel DROY, MD (1) , Stéfan J. DARMONI, MD, PhD (2) , Philippe MASSARI, MD (1), Thierry BLANC, MD, M. Sc, (3) , Fabienne Moritz, MD, M. Sc, (1), Jacques LEROY, MD (1)

(1) Poison Control Centre and Adult Intensive Care Unit, Rouen University Hospital, 1 Rue de Germont, F76031 Rouen Cedex, France; Tél: (33) Fax : (33) E-mail:
(2) Information System, Computing and Telecommunications Department , Rouen University Hospital, E-mail:
(3) Child Intensive Care Unit, Rouen University Hospital,

Correspondence and reprints to Jean-Michel Droy
This work was supported by a grant from the Conseil de l'Informatique Hospitalière et de Santé (C.I.H.S.), 1993


Computers in clinical decision making can be used to refer to a data base or to carry out a part of the medical reasoning by algorithmic, Bayesian, or knowledge based systems (1). In toxicology, product and case data bases (2, 3) are easier to use and update than books or written forms. Reasoning systems can go further: (a) in a known intoxication to give non-toxicologist physicians better advice according to drugs ingested, clinical manifestations and delay, (b) in an unknown intoxication to identify products according to clinical manifestations. Drug poisoning is a frequent problem, representing in Rouen University Hospital 8% of adult emergencies and 3% of childhood emergencies. The aim of SETH is to give end-users specific advice for the treatment and monitoring of drug poisoning in adults and children.


Decision in toxicology:

In most cases of acute drug poisoning (85% of the overall intoxication), drugs and ingested quantities are hypothetically known from patient or family report. According to the presumed drug ingested, the physician has to evaluate the initial severity and the potential toxicity to define monitoring and treatment. In case of poisoning with a single drug, toxicological reasoning is simple: the intoxication can be confirmed by the presence of one or more clinical findings. It can be eliminated if one very sensitive sign is absent, such as bradycardia in beta-blocker poisoning, after checking that the patient has been examined within a time frame compatible with the drug's toxicokinetics. Poisoning severity depends on the type of drug, the quantity ingested which can be assessed by questioning, the clinical context of the patient, his/her medical history, and the possible ingestion of other toxic, such as alcohol. Monitoring will depend on initial findings and potential toxicity of the drug toxicological class.

In case of poisoning with more than one drug or with a drug containing more than one product (the most frequent drug poisoning in adult), toxicological reasoning is far more complex. Textbooks (4-6), product and case data bases and the advice of a Poison Control Center (PCC) are often necessary to handle such a patient. The physician must confirm or eliminate the ingestion of each drug, or at least each toxicological class, taking into account toxicity delays and modifications of clinical findings related to multiple product intoxication. The physician must assess for each class whether the physical examination is performed during the toxicity interval, whether clinical findings confirm or eliminate the ingestion and whether the hypothetical ingested dose is greater than the toxic dose. Clinical findings are certainly the most important argument in toxicology for or against the ingestion of a poison. The absence of a very sensitive sign can eliminate the ingestion of a drug if all the other signs found can be explained by the other presumably ingested classes, and the examination was done within the expected toxicity interval. On the other hand, a common sign could confirm the ingestion of a given class if it is extremely rarely caused by the other presumably ingested classes. Possible interactions between drugs ingested must be taken into account to define overall advice on treatment and monitoring.

In case of an intoxication with unknown drugs, the only data available to the physician are signs and context. From these data, he/she must deduce the list of toxicological classes which can explain all the signs and the context.

Knowledge and data representation

This expert system in clinical toxicology was designed by a project group consisting of four experts in toxicology and two medical informaticians from the Rouen University Hospital. The expert system shell is KBMS of Trinzic (8) primarily in the MS-DOS environment and secondarily in 1993 in the Windows environment. Access is the database management system used to create and easily maintain tables of drugs and toxicological classes. Hardware is an IBM compatible microcomputer with a 80486 microprocessor and 8 Megabytes of RAM. Hard disk occupation is less than 20 Megabytes.

The SETH's knowledge is comprised of terms, objects, requests, rules and descriptive tems. The consultation model consists of findings, hypotheses and decision rules. Findings are requested from the end-user. Hypotheses are conclusions that may be inferred by the system; they include treatment and monitoring recommendations and intermediate hypotheses, representing relevant aggregations of observations useful for organizing the reasoning. Rules are used to link findings to hypotheses. The typical expression of a rule is an IF ... THEN statement, where the IF clause contains the pattern and the THEN clause contains the action.

The data base contains information on drugs, toxicological classes, potential clinical findings, advice on treatment and monitoring according to severity of poisoning. After each update in the data base, these information are transferred to corresponding objects in the knowledge base. Currently, the data base contains the 1153 most toxic or most frequently ingested French drugs from 78 different toxicological classes.


Our cognitive analysis was transposed in the knowledge base. The SETH expert system simulates the expert reasoning, taking into account for each toxicological class delay, signs and dose. SETH describes a level graph, where each level represents a step of the reasoning. The first level contains initial conclusions on delay, dose and signs. These three initial conclusions generate a final conclusion, which represents the second level of the graph. This final conclusion defines for each class accurate monitoring and treatment advice, taking into account drug interactions. All the conclusions are done at the toxicological class level. SETH checks if the patient had ingested drugs from the same class (Cf. Figure 1).

We are using the set theory to determine the conclusion on sign. If a sign is explained by a unique class, the conclusion on sign is 'yes'. If all the signs explained by one class are also explained by an another class, the conclusion is '?', unless one of these signs are exceptionally present for all the other classes; in that case, the conclusion is also 'yes'. If a class explain no signs, the conclusion is 'no'. If the conclusion on signs is '?' for a class, SETH checks if there are some very frequent signs absent; if it the case, the overall conclusion on signs is 'no'. We defined a toxic dose and massive dose for most of the drugs. Some information about these doses came from the literature and textbooks; some data were estimated by the experts. We also defined delays of toxicity for each toxicological class. These data are mandatory to define the conclusions on dose and delay. SETH calculates the ingested cumulative dose for each toxicological class and compares it with toxic and massive doses defined for this class to determine the conclusion on dose.

Inferencing is used to compute initial conclusions with respect to delay, clinical manifestations and doses, to give global conclusions regarding each ingested class, and to take into account interactions between classes or drugs and treat specific problems. An example of the logic of SETH regarding a case is displayed in Table 1.

A first version of SETH was primarily developed in adult poisoning (9). In April 1993, the extension of the knowledge base to child poisoning began. Although the model previously described for adult poisoning is very much the same for child poisoning, this extension implied some specific adds-on and modifications in the knowledge base and the databases: (a) the calculation of the toxic dose takes into account the weight of the child; (b) some signs are specific for some toxicological classes, especially their frequency, specificity and severity; and (c) accurate monitoring and treatment advice is specific for most of toxicological classes.

SETH also contains a case database : all data inputed by an end-user, (PCC's resident), such as names of drugs, or generated by SETH such as the conclusions about the intoxication, are stored in the case database.

Operating characteristics:

The input of the patient and poisoning data is as simple as possible to minimize the time spent to input data. The first input screen includes patient and physician data, the name and quantities of drugs. Data on drugs can be inputted by proprietary and/or non-proprietary names. Quantities of ingested drugs are expressed in tablets or milliliters for proprietary names according to the type of ingestion and in milligrams for non-proprietary names. The second input screen only clinical manifestations chosen from a list of potential symptoms according to the presumed drug ingested. The report includes : a reminder summary of poisoning, an overall conclusion, and advice on emergency actions needed, monitoring, gastric lavage, specific actions, and biological assays. The advice generated by SETH takes into account the ingestion of other toxics, e.g. alcohol. The recall of the poisoning includes all the input data, calculated data such as hypothetically ingested dose for a class, and data from databases such as the name of a toxicological class for a drug. The overall conclusion about the intoxication displays in a textual way, the initial conclusions and the global conclusion for each class, and some details like the list of clinical symptoms which can be explained only by one class. The potential clinical manifestations conclude the report ; for each of them, the classes which can explained them are displayed. A printout of the main points of the SETH consultation is possible. The report is available for adult and child poisoning, up to 4 drugs or 12 toxicological classes ingested.

Identification of drugs according to 56 clinical manifestations is also available in a different module in the case of an intoxication with unknown drugs. The end-user inputs the clinical signs and SETH is giving the list of toxicological classes which can explain all the signs. We designed an imputability model to hierarchy this list of toxicological classes. This model takes into account the prevalence of each toxicological class in adult and child poisoning and a predictive score given by the expert of each sign for each class.



We designed three phases to evaluate SETH in our hospital (internal evaluation) (10). There is no objective criterion to evaluate SETH, therefore we did use the expert advice as the gold standard. The aim of the first phase was to test respectively the initial conclusions on delay, signs and dose, and the global conclusion. The purpose of the second phase was to test the accuracy of monitoring and treatment advice generated by the expert system. The aim of the third phase was to evaluate the functionality of SETH, specially the impact of the practical use of SETH (11) using a field trial .

In the first 'internal' test, including 47 poisoning cases representing 72 toxicological classes, no errors were detected in final conclusions but initial conclusions were false in 14% of cases in symptoms conclusion, and in 4% in doses conclusion. Unsatisfying conclusions have been explained by data base errors for 3 classes and lack of precision for 5 classes: symptom conclusions were not taking into account the frequency of signs. The second phase of validation included 120 cases and demonstrated the relative robustness of the report (accuracy of 71%). Errors were analysed and were explained by the lack of SETH to take into account : (a) some interactions between classes, (b) specific reasoning for some classes, and (c) hierarchy between clinical manifestations in terms of emergency actions. For example, the treatment of a cardiac conduction block could improve a bradycardia. We did take into account this deep knowledge in our model.

All the errors of phase one and two of the evaluation were corrected by the beginning of 1992, improving the accuracy of the report. The monthly review of the cases analyzed by SETH has not detected any important dysfunctions (9% of overcautious conclusions). The project group installed SETH in our Poison Center in April 1992. SETH was then daily used by residents primarily for support during consultations and secondarily as an educational tool on drug poisoning. The use of SETH is not mandatory for the PCC's residents. Between April 1992 and October 1994, 2099 drug intoxication cases were inputted, representing 2.1 cases a day, 10% of the Poison Centre's overall activity and 40% of the Poison Control Center's phone calls regarding drug poisoning. There was an increase in the daily use of SETH: between January and October 1994, 850 cases were analysed by SETH (average of 3.15 cases a day) vs. 608 cases (average of 2.25 cases a day) during the same period of 1993. Since the implementation in our Poison Control Center, there has been an increase in telephone calls from inside the Rouen University Hospital, especially by the residents in the Intensive Care Unit and Emergency Department. Some are coming to obtain a printout of the conclusions of SETH.

In November 1994, in order to estimate the impact of SETH on the clinical decision-making of residents, a clinical trial was carried out at the PCC of the Rouen University Hospital. The trial included the nine residents of the PCC who used SETH in the past 30 months. The median of SETH's utilisation at the time of the trial was 24 months (minimum = 10 months and maximum = 30 months). The main objective was to measure whether residents with a computer-aided decision support system changed their management of drug poisoning. The results of the questionnaires are displayed in Table 1.


Toxicological computer-aided decision support systems, such as SETH, could and should be developed because they give better advice than drug databases by themselves. Search in Medline, Toxline and Toxlit, shows that few computer-aided decision support systems have been developed in clinical toxicology (12-18) compared to several dozen in environmental and experimental toxicology (21.). Monov et al. (15.) developed an expert system (Medicotox Consilium) for diagnostics and treatment of exogenous poisonings. It is the most similar system to SETH. We shared some common approaches such as the use of toxicological classes. Nonetheless, we believe that the set theory used in SETH is much more appropriate in clinical toxicology than scoring used in Medicotox Consilium. The domain of the latter is larger that SETH's domain but the management of drug poisoning is far more precise in SETH. Medicotox Consilium is operational in several Bulgarian hospitals. Soto et al. (16.) developed an expert system to model product safety, toxicology and regulatory decision processes. Spiehler et al. (17., 18.) developed an expert system analysis to aid interpretation in forensic toxicology. Amitriptyline (17.) and morphine (18.) were selected in the pilot trial. Interpretation was defined as advising on dose, time since ingestion and effect. The authors searched for patterns in the data therefore to interpret poisoning with unknown data. Although our knowledge representation is very similar, our first goal was to develop an expert system to give advice in poisoning with known drugs.

Compared with other medical fields, clinical toxicology is probably easier to formalise because few heuristics are used and lots of data can be managed in a database, such as storing information about drugs and toxicological classes. From the beginning of the analysis, we have intentionally separated data and knowledge. In SETH, data on drugs, toxicological classes and advice can be updated within the data base application. The maintenance of data on drugs and toxicological classes is performed with an electronic dictionary of French drugs available in our University Hospital and is updated every three months (20). Only reasoning and toxicological classes interaction updates have to be done in the knowledge base. Therefore the overall maintainance of SETH is very easy. The maintenance of the knowledge base includes a review of the data stored in the case database challenging it with an alternative analysis by an expert. It is also possible to translate SETH in other languages quite easily because of the structure of our model. SETH was definitely not developed for the experts; but they are using part of it, especially the information stored in the different databases (e.g. toxic doses) as they could do with an electronic textbook. We have minimized the time spent to input data (2 screens, less than one minute) because it is one of the main reasons to explain the lack of use of expert systems in daily practice (21). We have also minimized the overall time of the SETH consultation (less than three minutes) because drug poisoning is an emergency situation. The telephone support must be as quick as possible. Our effort was successful : the PCC's residents have judged that SETH is faster than sources previously used by PCC's residents. The effect of SETH in the daily practice of our PCC is positive : the performance of the residents increased and they would agree to use it outside our University Hospital. SETH does not reduce the time spent to converse with colleagues about drug poisoning which is a very important point by which to judge the acceptance of the system. The results showing the SETH's acceptance as a useful knowledge source must be underscored because the PCC's residents had a low previous experience with computers (mean score = 2.56). Furthermore, before the use of SETH, the residents thought that CDSS were not really useful in clinical decision making (mean score = 2.71) and after its use, they were willing to use SETH outside the Rouen University Hospital (mean score = 4.44).. Nonetheless, the residents used the expert system in only 40% of the PCC's drug poisoning cases, the most difficult ones, especially multiple drug intoxication. They do not use SETH for a single benzodiazepine intoxication which is very frequent because they already know how to handle it by themselves.

The SETH's domain is very precise: exclusively drug poisoning. We defined in the knowledge base some limits of expertise (21), e.g. in case of a drug intoxication with a short delay, without any clinical sign and any information about the ingested dose, the SETH advice is 'Ask a human expert' but it gives the 'maximal' management which is the correct answer in case of doubt in clinical toxicology.

Next steps of SETH's developments will be:
(1) The evaluation in child poisoning will begin in May 1995 and will follow the same three phases described for adult poisoning.
(2) The evaluation of SETH in adult and child poisoning must be completed by an external evaluation, including other French hospitals (PCCs and Emergencies Departments), before it can be transferred to other hospitals. The transferability of an expert system has been defined as 'the degree of which a system retains its reliability when applied in an another organisational environment (21).

In conclusion, we believe that an expert system in clinical toxicology is a valuable tool in the daily practice of a Poison Control Center.


The authors thank Benoît Thirion for his librarian assistance and Karen Benattasse for her linguistic assistance

Table 1. Example of the logic of SETH regarding a case

Intoxication with :
- VERAPAMIL 5400 mg
- CLONAZEPAM quantity unknown
Delay after ingestion : 4 hours
Signs : obduration, bradycardia, junctionnal rythm
Initial and global conclusions :
	       Conclusion on delay	Conclusion on signs Conclusion on dose  Globalconclusion
VERAPAMIL 	Y	                    M	                 O	                    O
ACETAMINOPHEN 	N	                   ?(*)	                 N	                    O
CLONAZEPAM 	Y	                   ?(*)	                 ?	                   O
(*) obduration can be observed in intoxication by the 3 drugs, so the expert system can't concluded on the signs for ACETAMINOPHEN and CLONAZEPAM

Summary of the situation :
For ACETAMINOPHEN, delay after ingestion is short, others signs of toxicity can appeared. For the others drugs, time issued after ingestion corresponds of the toxicity interval.
Intoxication by VERAPAMIL is responsible of bradycardia and junctionnal rythm, the ingested dose is very important, the main toxicity is cardiac. The dose ingested for ACETAMINOPHEN is in the range of toxicity, no conclusion can be done on signs because of the too short delay after ingestion. The main toxicity is hepatic. Intoxication by CLONAZEPAM can be responsible of obduration, but this symptom can also be observed after ingestion of the others drugs.

Emergency advice
If bradycardia is important, junctional rythm can indicate temporary electrosystolic-stimulation. Treatement by N-acetyl-cystein must be immediatly started.

Survey advice
- In an intensive care unit - EKG and hemodynamic must be monitored - Consiousness and respiratoy functiun must be surveyed - Survey of more than 72 hours is necessary

Gastric lavage
- Gastric lavage, if the is no "contre indication"

Biological advice
- ACETAMINOPHEN dosage is very important to perform - Survey of hepatic tests, glycemia and prothrombin time - Survey of renal funtiun and kaliemia

Specific advice
- N-Acetyl-Cystein treatment is depending of ACETAMINOPHEN dosage : IV infusion initially of 150 mg/kg in 20 mn, then of 50 mg/kg every 4 hours. - Flumazenil test can be performed.

Table 2. Acceptance and attitudes towards SETH as a useful knowledge source

Question                            Frequency of Scores                        Mean Score
                                    -1 1 2 3 4 5
1.Previous experience with computers 0 3 3 0 1 2     2.56
2. CDSS* might be useful 
in clinical decision making          2 1 2 2 2 0     2.71
3. Easy to use                       0 1 1 1 5 1     3.44
4. Clear and convenient presentation 0 1 1 1 4 2     3.56
5. Faster than previous sources      0 0 0 3 2 4     4.11
6. Less conversation with colleagues 0 6 1 0 2 0     1.78
7. SETH gives appropriate answer 
for most of patients                 0 1 0 2 4 2     3.67
8. Performance increases             0 0 1 1 4 3     4.00
9. SETH serves as a useful 
training source                      0 0 2 2 3 2     3.56
10. Outside Rouen, would use 
in daily practice                    0 0 0 0 5 4     4.44

Scores: 1=strongly disagree 2=disagree 3=in between 4=agree 5=strongly agree -1=no opinion
* CDSS stands for Computer Decision Support Systems


  1. Shortliffe EH. Clinical decision-support systems. In: Medical Informatics, Computer applications in health care. Addison-Wasley: Shortliffe EH & Perreault, ed., 1990:466-502.
  2. Quinton A, Latry P, Biour M.: Hepatox : base de données sur les médicaments hépatotoxiques.. Gastroenterol. Clin. Biol. 1993;17:H116-120.
  3. Scialli AR: A computerized consultation service in reproductive toxicology: summary of the first five years. Obstet. Gynecol. 1988;72:195-199.
  4. Bismuth C, Baud F, Conso F, Frejaville JP, Garnier R. Toxicologie clinique. Médecine & Science, Flammarion Ed., Paris, 1987.
  5. Descotes J, Testud F, Frantz P. Les urgences en toxicologie. Maloine Ed., Paris, 1992.
  6. Hellenhorn MJ, Barceloux DG. Medical toxicology. Diagnosis and treatment of human poisoning. Elsevier Ed., New York, 1988.
  7. Yurman D. Success factors for expert systems. In: Expert systems in environmental applications, American Chemical Society, Hushon JM, ed., Washington DC, 1988:121-147.
  8. KBMS application development guide. Trinzic, Palo Alto, Ca, USA, 1993.
  9. Darmoni SJ, Massari P, Droy JM, Mahé N, Blanc T, Moirot E, Leroy J. SETH: an expert system for the management on acute drug poisoning in adults. Comput. Methods Programs Biomed. 1993; 43: 171-176.
  10. Stunder M: Verification and validation of environmental expert systems. In: Expert systems in environmental applications, American Chemical Society, Washington DC: Hushon JM Ed., 1990:39-51.
  11. Visser, MC. Hasman, A. Van der Linden, CJ. 1994, "Protocol processing system (ProtoVIEW) to support residents at emergency ward". Proceedings of MIE 94, Twelfth International Congress of the European Federation for Medical Informatics, P. Barahona, M. Veloso & J. Bryant, Eds., pp. 138-143.
  12. Cartasegna S; Pancolini O; Spampinato L; Costantino D. I sistemi esperti nell'emergenza medica. Minerva Med. 1988;79:407-412
  13. Durrieu G, Saint-Martin M, Shavant L. Informatique et détermination des champignons. Bull. Soc. Myc. Fr. 1986;102:183-187.
  14. Emerman CL, Patrick EA. Computer-aided diagnosis of drug poisoning: a Consult-I subsystem. In: Proceedings of the Seventh Annual Symposium on Computer Applications in Medical Care, Dayhoff RE Ed., IEEE Computer Society, 1983:126-128.
  15. Monov A, Iordonova I, Zagorchev P, Vassilev V, Nissimov M, Kojuharov R, Tconev R, Damianov V. MEDIXOTOX CONSILIUM: An expert system in clinical toxicology. In: MEDINFO 92, 7th World Congress on Medical Informatics, Lun KC, Degoulet P, Piemme T & Rienhoff O, (Eds), North Holland, 1992:610-614.
  16. Soto RJ, Osimitz TG, Oleson A: Toxpert: an expert system for risk assessment. In: Proceedings of the Twelfth Annual Symposium on Computer Applications in Medical Care, Greenes Ed., Computer Society Press, 1988:48-51.
  17. Spiehler V, Spiehler E, Osselton MD: Application of expert systems analysis to interpretation of cases involving amitriptyline. J. Anal. Toxicol. 1988;12:216-224.
  18. Spiehler VR. Computer-assisted interpretation in forensic toxicology: morphine-involved deaths. J. Forensic. Sci. 1989;34:1104-1115.
  19. Hushon M: Overview of environmental expert systems. In: Expert systems in environmental applications, American Chemical Society, Washington DC: Hushon JM Ed., 1990:1-24.
  20. SJ. Darmoni, F. Dufour, P. Massari, S. Arnoudts, B. Dieu, B. Alizon, F. Hantute, M. Baldenweck. Consultation of the Electronic Vidal dictionary in the Rouen University Hospital: analysis of the first year of utilisation. In: Proceedings of MIE 94, Twelfth International Congress of the European Federation for Medical Informatics, P. Barahona, M. Veloso & J. Bryant, Eds., 1994, pp 384-388.
  21. Clarke K, O'Moore R, Smeets R, et al. A methodology fo evaluation in knowledge-bsed systems in medicine. Artificial Intelligence in Medicine 1994; 6: 107-121.