Hybrid Expert System in Anesthesiology for Critical ... - IEEE Xplore

5 downloads 0 Views 489KB Size Report
Fernando Passold. Depto. de Engenharia ElCtrica. Universidade de Passo Fundo. Caixa Postal: 567. 99.001-970 Passo Fundo RS - Brazil. Phone: + 55 54 3 11 ...
Hybrid Expert System in Anesthesiology for Critical Patients Fernando Passold

Renato Garcia Ojeda, Jorge Muiiiz Barreto

Depto. de Engenharia ElCtrica Universidade de Passo Fundo Caixa Postal: 567 99.001-970 Passo Fundo RS - Brazil Phone: + 55 54 3 11 1400 - Fax: 4-55 54 3 1 I 1307 E-Mail: [email protected]

Grupo de Pesquisas em Engenharia BiomCdica (GPEB) Universidade Federal de Santa Catarina Caixa Postal: 476 88.040-900 Florianbpolis SC - Brazil Phone: + SS 48 23 1 9872 E-Mail: [email protected]

Abstract - It is showed the results achieved using neural networks to represent anesthesiological knowledge specially to the treatnrcnt of paticnts considered critical in anestlicsiological point of v i c v . l’liis was accomplished in a hybrid expert system appficd to anestliesiology, called PROVANES (Proposal and Evaluating of Anestliesia Plan), expanded to c:ises of critical patients in anesthesiological point of view. Its main knowledgc base is formed by ten artificial neural networlis that perform tlic tasks involved i n an anesthesia plan in the same byay a specialist would do. These nets are embedded in an expert system shell, driven by production rules. This approach bypasses the knowledge elucidation bottleneck by take advantage of the neural net’s knowledge extraction capability.

INTRODUCTION It is showed the results achieved in a hybrid expert system applied to anesthesiology, called PROVANES (Proposal and Evaluating of an Anesthesia Plan), expanded to cases of critical patients in anesthesiological point of view [2!]. This system was emerged as a Dr. thesis [7] exploring tectonics of artificial intelligence suitable to represent the knowledge of the area of anesthesiology. The system aids tlie anestliesiologist in the decision making process of an anesthesia plan, by suggest the technique and anesthetic agents based in many information such body weight, age, sex, routc and time of administration, tolerance, physiological variables, pathological state, genetic factors and drugs interaction. No simple summary of the effect of these variables is possible [SI, so thc main knowledge base of this system use artificial neural networks [9, 161.

Physical Stale Classification

Pre-medication Drugs Anesthetic Technique Selection

Induction Modalily

Ac‘dtt m a l Procedures

Loco-regional Anesthesia

inductionni4aintenance DIUQS

1

E?euroMusculal

4

Net ,o

Blockerts)

i bqaintenance o i Air Ways

General Anesthesia

Fig. 1 . Neural networks used by PROVANES resembling the

aiicsthesia plan steps.

The firsts net (or the first step) classifies the physical state of the patient according to the ASA classification [22]. The second net determines the need of a premedication drug. The third net determines what anesthetic technique should be applied. This net also separates the followins actions of the system: in the case of locoregional anesthesia, it perform more tree nets that determine: the level of blockage (net 4), the sedation drug(s) (net 5 ) and additional procedures (net 6). In the case of general anesthesia, the system key on four nets: the induction modality (net 7), the induction/maintenance drugs (net 8), the neuro-muscular blocker(s) (net 9) and the maintenance of the air ways (net 10). The production rules manage the overall system, ciriticiziiig the neural nets outputs, further controlling ihe user interface, data flow, help texts and help syndromes’ texts. A drug’s interactions database is also used for the criiicizing step. Fig. 2. points out tlie diagram block of

PROVANES.

I . SYSTEMOVERVIEW PROVANES was asseinblcd

us;n& a traditional cxpet-t

system shell driven by production rules with ten artificial neural networks embedded featuring this as a hybrid expert system where more than one agent is used to process information [3, 61. This approach bypasses the knowledge elucidation bottleneck by take advantage of the neural net’s knowledge extraction capability. ‘The ten nets resemble the process of anesthesia planning i n tlie same way as a specialist would do, as demonstrated i n Fig. 1.

0-7803-3 1 -09-5/96/$5.00

0 1996 IEEE

1486

’The expansion introduced consists of doubling the formei- ten neural networks and its criticizing rules to compose a new, distinct knowledge base concerning to critical patients. When the system runs, it identifies the patient’s case and switches to the corresponding knowledge base. Some frequent cases present in the hospitals where the system has been developed as appears in Fig. 3.

of the nets, it was measured the performance [16,19], mean squared errors (MSE) [ I 11, the percentage of “hard errors’’ (HardErr) [13], and the hit taxes reached by each outputs of the net, both for the training and testing set, as illustrated by the Fig. 4. Training Evolution 801

~~

1

during training step

Fig. 3. Distribution of‘ most frequent cases of critical patients.

II. METHODOLOGY A distinct anesthesiological database necessary for training the new neural networks was created, derived from a larger database that nowadays has estimated more than 10,000 cases reportirig real patient’s data submitted to a surgery intervention [4]. This database reflects the resolutions adopted by different specialists to different cases of patients collected from the Celso Ramos Hospital and University Hospital, both of them in Florianopolis city where the system has oeen developed.

The “hard errors” refer to a quantitative measure approach of the network lcarning [13]. I-lard errors occur when the net compute:; a response far from one class of the desired response. For example, the net classifies a physical state of a patient as ASA I when the correct (has stored in the database) is ASA 111. Otherwise the net could compute this patient has ASA 11, in this case, this is considered an “optimistic error” [ 131, and reflects different judgments made by different specialists. To prevent over learning, optimal learning points were detected during the training process, when: a) the net reaches higher performances for the testing set [161, or; b) the performance of the testing set remains constimt but: i) the net has reached higher performance for the training set, mor; ii)the net has reached the less residual mean squared error for the testing set [ I 11. Refer to Fig. 5 to see some results.

63 / 66 91, 7 3 / 74 96 63 1-56Ob

Ill. NEURAL NETWORKS

63 I66 %

This system uses nets with 3 layers, trained by the back. propagation algorithm (the, most popular) expanded with a simulated annealing process and momentum term [ 16, 191. It was prepared two diztinct data sets, one for training and another to the testing phase for each network. These: sets keep in common, similar output patter’s distribution and no repeated sequence:, of equal patterns to avoid introducing bias into the rtmlts. During the training phase

Obs: the numbers Close the blocks refers to tralning and testing performances reached, respectwally

Pig. 5 . A summary results for [he neural network training.

Iv. SYSTEM EVALUATING After the nets were trained, the production rules necessary to complete the knowledge base for critical

1487

be achieved if it was be used a combination of discrete and continuous input data, even if some normalization was necessary, like in the case to inform the level of leukocytes (considered normal between 4,000 and lO,OOD), as adopted in [2]. Also, it could bee used fuzzyfication technique to inform certain input variables, as how much a patient can be considered obese [17]. This last example also clarifies the absence of a criteria to categorize certain variables as how obese is a patient (this particular case justify the high percentage of in critical patient’s distribution cases, Fig. 3.). This aspect suggests a . reevaluating of the anesthesiological database used to train the neural networks of this system, which no contains continuos data. The system also seems a little “ingenuous” for some specialists as it requires a large amount of input data variables to proceed correctly some patient’s cases. But it must be keep in mind that: better a specialist are, he uses less input information to reach his conclusions [ 181; 17) the lack of time dimension information (“timing”) to some input data variables [l]. For example, it is different to anesthetize a patient submitted to an heinodialise yesterday between another, a week ago; c) the system exhibits some partial inconsistency between some suggestions of induction/maintenance drugs (outputs of net 8 + rules), more characteristic of other anesthetic technique (net 3 + rules), although not always it was become an error. It could be happens due it was registered in the anesthesiological database (used to train the neural nets), cases of patients attended in emergency situation when is common changes of inductionhnaintenance drugs during the anesthesiological proceeding, or even so, changes of the anesthetic technique and related drugs. This fact, suggests the use of filters (rules) trying to detect this cases even before or after the data are stocked in the anesthesiologycal database or present to the system

patients was introduced through an interactive build-testrefine cycle. The expanded system was tested in a static and dynamic manner [5]. In the static step, the logical errors were fixed. As the system is in developing stage, a dynamic test was taken in a laboratory stage [ 5 , 121. Some problems were observed in the evaluation phase: a ) t h e absence of formal validation techniques to evaluate decision support system or expert systems in medicine even in nowadays [ I , 14, 20, 231; b) the lack of a “gold-standard’ reference in the area of anesthesiology [14, 5, 12, 20, 231; c) the difficulty to evaluate connectionist (neural networks) systems due its “impenetrable” and inexplicable behavior featuring them as “black-box” systems [ 121; It was observed some items: a) a sinall set of test cases with similar distribution cases that occurs in the hospitals where the system was developed was builder even including some atypical cases [ 121; b) the principle of independence was respected using others specialists to evaluate the system than those who participated in the developing phase [ 5 ] . c) equal access information between the system and its judges was proceed [SI. Table 1. shows a summary result achieved. Some deficiencies (Cap. 4 and 5, [2 I]) could be noted, the most interesting: a) A low sensibility to certain input variables. For example: a little difference between 5.5 and 7.5 units of potassium in patient’s blood exam influence the anesthetic procedure that could be applied, but had not affected the system. It can be justified due the fact that the internal information flow through the system (neural networks) consists only of discrete data variables. For example, information about blood pressure is treat as -1 for low pressure, 0 for normal and +1 for high pressure. Maybe better results could

Correct Suggestions

. . .. . .

.



. ... . ... . .... . . . _.

.. . ... . . 1. .... . .. . ..!;

.

.. ... ...o. .

.

* The system only suggests up to 4 inductionimaintenancc drugs in cases of general ancsthesia. In this cadcolumns is bzcn sumniing the correct and \srong suggestions done for each case. Table 1 . Summary results from the last test in the expanded system over critical cases.

1488

(neural nets).

v. COI\JC:I-USIONS Although the “black-box” nature of this system compromises good explanations and the transparency as expected by the specialists, it reaches a good performance, showing tha: neural networks could be use t o represent knowledge in the area of anesthesiology. The lack of good explanation could be overcome through the introduction of some tc,chnique of neural networks sensibility analyses [15] in conjunction with explain by confabulation technique [I)]. Other neural networks methodologies could be explored: use of genetic algorithms or a change dirccted to a fuzzy neural structure [ 181. Even so, some benefits of neural networks must be remember: a) ability to process a massive amount of input data; b) simulation of diffuse medical reasoning [ 121; c) higher performances when compared writ11 statistical approaches [ 2 ] ; d) self-organizing ability - learning capability; e) easy knowledge base updating.

ACKNOLVEDGMENT W e want to acknowledge the specialists who take part in the developing and tes’:ing phase of this version: Dr. D a d o Freire Duarte, DE;. Maria Anita Cristina Costa Spindola Bez Balli, Dr. Getulio R. d e Oliveira Filho, Drs. Marcia Regina Ghellar and Dr. Cesar AntGnio Alisaka.

REFERENCES BERNER, E. S., Performance of Four Computer-Based Diagnostic Systems, The N e w Eiiglad Journal of Medicine. v. 330, n. 25, 1792-1796, June, 23th, 1994. 121 BIJSCIIMAN, T. G.. KUBOS, K. I.,., SETDLER, A. J., SIEGFORI‘II, iM. J. A., A Comparison os Statistical and Connectionist Models f w i.he I’redition of Chronicity I n A Surgical Intensive Care Unit, Crifictrl Core Medicine Journal. v. 22, n. 5, 750..762, May. 1994. , 16, 11. 10, 108131 CAlJDIL, M., Expert networks, B y t ~ v. 116, October, 1991. 141 DUARTE, D. F., ZANCHIN, C. Y., LIMA. W. C., FALQUETO, J., Bases de Dados ern Ancsthcsiologia. In [I1

BRAZILIAN CONGKE!jS OF INFORMATIC IN HEAI,T, 2, 1988. Annals... October. 1988.

151 FIESCIlI, M , Toivard Viilidation or Expert Systems as Medical Decision Aids, fnt. J. Bionzed. C,‘onzpu/. (Ireland): Elsevier Scientific Publishers l,td., n . 26, 93-108. 1990. 161 GALLANT, S. L.. Connectionist Expert Systems, Conzrnunicntions of thz ifCi\f, v. 31, n. 2, 152-169, February, 1988.

I71

GARCIA, R. 0..Timicas de Inteligtncia Artificial Aplicadas ao Apoio r a Ilecisgo Medica na Area dc

Anestesiologia. Dr. Thesis of Federal IJniversity of Santa Catarina. Dept. of Electrical Engineering, 1992. [SI GOODMAN, L. S., GILMAN, A., The Pharmacological Basis of Tlzerczpeutics, 4tf de., New York: The MaeMillian Company, 19’70. 191 GUERRIERE, M. R. J., DETSKY, A. S., Neural Networks: What Are ‘They ?, Annals of Internal Medicine, v. 115, n. 11, December, 1991. [lo] HAMMERSTROM, D., Neural Networks at Work, IEEE Spectrum Mug., 26-32, June, 1993. [ 1I] I-IAMMERSTROM, D., Working with Neural N’etworks, JEEE S@ctruni Mag,, 46-53, July, 1993. [I21 HART, A., WYATT, .I.,Evaluating black-boxes as medical decision aids: issues arising from a study of neiii-a1networks. Med. lj;fimm, v. 15, 11. 3. 229-236. 1990. [13] tIOL,DA\i’AY, R. M., WI-IITE, M.W., MARMAR.OU, A., Classification of. Soniatosensor~-Evokcd 1’83tentials Iiccordcd from Patients with Severe Head Injuries, IEEE Engrnccriirg in AJedicine m d B i o / o & yA4ag v 9, 11.3, 4349, September, 1990. [ 141 KASSIEI1EK, J . P., ’4Report Card On Computer-Assisted Diagnosis, ?%e ATew Englad .Journal of illcciicirze, v. 330, n. 25, 1824-1825, June, 23th, 1994. 1151 KLIMASAlJSKAS, C. C. “CASEY’, Neural Nets Tell Why: A technique for explaining a neural network’s decision-making process, Dr. Dohb ’s ,Journa/, 16-24, 1991. 1161 KNIGHI‘. K . Connectionist, Ideas and Algorithms, Conunuriications of the ACA4, v. 33, n. 11, 59-74, November, 1990. 11’71 LOPES, H., NASSAR, S.. HERNANDEZ, A., BAIIRETO, S . M.. Classificacion Difusa de la Obesidiid para Anestesia, In: CONGRESS0 CHILENO DE INGINIERIA ELECTRICA, 10, 1993, Valdivia (Chile). itnais .....,1993. [I81 MACHADO, I