!

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING. VOL. 38. NO. 2. FEBRUARY 1991

I99

Foetos: An Expert System for Fetal Assessment Amparo Alonso-Betanzos, Member, I E E E , Vicente Moret-Bonillo, Carlos Hernandez-Sande, Member, IEEE

Abstracf-Assessment of the fetus in a high-risk pregnancy uses a variety of tests for screening and continued detection of in utero compromise. This paper describes Foetos, an expert system designed to help clinical personnel to interpret several fetal assessment tests: fetal biophysical profile, contraction stress test, and nonstress tests. Foetos has been built using the knowledge engineering tool Genie, which adopts a mixed frame- and rule-based approach to represent the clinical knowledge in the field. Foetos includes diagnostic, prognostic, and therapeutic structures, based on heuristic interpretation of such tests and contextual structures which relate interpretation to the overall clinical picture. The results of initial retrospective and prospective program validation are included in the report. These results show a substantial level of agreement between Foetos’ recommendations and clinical management. Interpretation of the results indicates that the obstetrical field could be an area of interest for the application of AI techniques.

INTRODUCTION

T

HE evaluation of intrauterine fetal condition using elec-

tronic recording of antepartum fetal heart rate patterns has been developed over the past 20 years. The first important advances in the field were those undertaken by Barcia and collaborators, who pioneered signal acquisition and processing techniques that became progressively more simple and accurate [ l ] . These methods make it possible to obtain “cleaner” signals from which parameters characterizing fetal heart rate (FHR) and uterine pressure (UP) could be established, leading to the identification of their normal and pathological patterns [ 2 ] . The advent of computers enabled an important advance in the quality of health care associated with this clinical modality: in the practical implementation of signal processing algorithms that extracted easily recognizable patterns and in providing significantly improved quality of the acquired signals [3]. Efforts have continued to automate the extraction of information from tracings of fetal heart rate and uterine pressure, and to combine this information with a knowledge of antenatal and perinatal factors so that an evaluation of intrauterine condition of the fetus and

Member, I E E E ,

and

a prediction of the newborn’s state can be made [4], [ 5 ] . Such prognoses can be of vital importance since failure to recognize and respond in early symptoms of fetal stress may have irreversible neurological consequences for the child or cause intrauterine fetal death 161, [7]. Since the first essays in the creation of artificial intelligence and the appearance of expert systems, medicine has been one of the preferred fields for knowledge engineers. Some of the new capabilities that expert systems incorporate are: the ability to manipulate symbolic concepts rather than numerical values, the incorporation of contextual structures for dynamic and accurate interpretation of the “data,” and the ability to explain their conclusions [8], [9]. These system features are desirable for clinicians who are reluctant to use conventional systems which do not use the terminology proper to the domain of application [lo]. Successful applications of artificial intelligence in the medical arena include Mycin [ 1 I], Casnet [ 121, and Internist [13], among others. The field is still very active and new intelligent programs are currently being developed [ 141. However, the field of obstetrics is virtually unexplored from the expert systems point of view. In this paper we describe Foetos, an expert system designed to help the clinician assess fetal status by means of the interpretation of the results obtained from a series of “fetal wellbeing tests, ” normally performed in high-risk pregnancies. The strategy adopted by Foetos is: 1) to quantify the level of risk of the pregnancy by means of the analysis of the risk factors present in the current pregnancy, 2) to perform an adequate fetal assessment by means of the evaluation of a series of tests normally used by obstetricians in high risk pregnancies, and 3) to achieve an adequate interpretation of the results of the latter tests in the light of the patient’s context and the conditions in which such tests have been performed, in order to obtain the diagnosis and prognosis of the fetal state, and to suggest appropriate therapy.

CLINICAL BACKGROUND Manuscript received March 8, 1989; revised May 10, 1990. This work was supported in part by Spanish CICYT under Project PA-86.0230 and the FPI “Plan General en el Extranjero.” A. Alonso-Betanzos was with the Department of Biomedical Engineering, Medical College of Georgia. Augusta. GA 30912. She is now with the Department of Computer Science, Faculty Informatics, University of La Coruna, La Coruna 15071, Spain. V. Moret-Bonillo was with the Applied Biophysics and Artificial Intelligence Laboratory, Department of Applied Physics, Faculty of Physics, University of Santiago, 15706 Santiago de Compostela, Spain. He is now with the Department of Biomedical Engineering, Medical College of Georgia, Augusta. GA 30912. C. Hernandez-Sande is with the Applied Biophysics and Artificial Intelligence Laboratory, Department of Applied Physics, Faculty of Physics, University of Santiago, 15706 Santiago de Compostela, Spain. IEEE Log Number 9041305.

In order to prevent the possibility of fetal compromise, analysis of the pregnancy is needed [15], [16]. The first step of analysis consists of the evaluation of the pregnancy in the light of the risk factors exhibited by the mother. If a situation of low risk is encountered no special supervision for the pregnancy is required. However, if the mother has been classified as high risk or some indications exist which suggest special attention by clinical personnel, the pregnancy is supervised, and a set of “fetal well-being” tests is initiated: nonstress test (NST) [ 171, fetal biophysical profile (BPP) [ 181, and contraction stress test (CST) [19]. Adequate interpretation of such tests and prescription of adequate therapies could avoid the possibility of fetal distress, frequently associated with such high-risk situations

[201.

0018-9294/91/0200-0199$01.00 @ 1991 IEEE

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 38. NO. 2. FEBRUARY 1991

200

TABLE I

A . Nonstress Test (NST) Basically, the NST is a reflection of fetal movement-related events. Its rationale is based on the observation that a fetal heart rate acceleration associated with fetal movement is a good prognostic sign and is predictive of fetal well-being. The test is performed by means of the monitoring of FHR and UP, without any kind of external stimulus. Typically, the fetus is classified as “reactive” (normal situation) if in the course of 10 min the FHR exhibits two or more accelerations of at least 15 beats in amplitude and 15 s duration [21], [22]. The main advantage of the NST is that it is reassuring when it is reactive and is quite easy to perform.

B. Contraction Stress Test (CST) The aim of this test is to determine possible placental insufficiency by observing whether a transitory reduction in the quantity of oxygen supplied to the fetus is associated with a change in its FHR [23]. In clinical practice, one of the ways in which this transitory reduction in oxygen supply is induced is by administering oxytocin to the mother (OCT = oxytocin challenge test). The result of the test is conventionally obtained by counting the number of late decelerations in the FHR, which the cardiotocograph trace shows to have occurred in response to the uterine contractions, as it is this kind of response which indicates the presence of insufficient oxygen reserves in the fetal tissues [24] (see Table I).

C. Fetal Biophysical Projile (BPP) The rationale for the BPP is that since no single fetal assessment parameter predicts adequately a bad fetal outcome, testing a combination of factors should improve test accuracy. It has greater discriminative power over the other tests (NST, CST) alone. Therefore, this test is an important confirmatory one for those fetuses which exhibited abnormal results in the other two tests [25]. The BPP consists of an NST performed simultaneously with ultrasonic observations of fetal respiration, movements, tone, and the volume of amniotic fluid. Each observation is scored from 0 (bad) to 2 (good), and the five individual scores are summed to give a total score which determines the corresponding suitable actions to perform, in accordance with the equivalences listed in Table 11. While the performance and analysis of these tests is possible if a systematic, well-defined procedure is employed, a clinically meaningful interpretation of such tests requires the application of a set of heuristic criteria so that they may be optimally adapted to the contexts of varied clinical problems. This converts the “simple” interpretation of these antenatal tests into a challenging task which is quite appropriate for the development of “intelligent” diagnostic structures.

DETERMINISTIC MONITORING The maternal/fetal signals undergo processing to extract various patterns of interest which are then introduced to the expert system for subsequent interpretation. Two basic types of signals are of interest: U P and FHR. These signals, picked up by an HP-8030A cardiotocograph, are fed to the computer after sampling and digitization. UP and FHR contain both low-frequency components representing, respectively, the contraction of the uterus and the fetal heart rate’s response to this stimulus, and high-frequency components due to straining, retching and fetal movements (in the

CLASSIFICATION CRITERIA FOR A CONVENTIONAL CST

3 or more uterine contractions in 1 0 without late decelerations

Negative

late decelerations present after

Positive

each uterine contraction

Suspicious

late decelerations observed but not with each uterine contraction decelerations present but uterine ntractions are tm frequent or too bng

I

Hyperstlmulated

case of UP), or to the activity of the fetal autonomic nervous system (in the case of FHR). Both frequency ranges thus contain clinically important information. Once the FHR and UP signals have been digitized at 1 Hz, the high and low frequencies are then separated by means of a digital second-order low-pass Butterworth filter. The duration of each contraction is detected using a derivative and threshold detector, but relaxed periods are included at the beginning and end of the segment of signal used for contraction classification, which is thus one and a half times as long as the contraction itself [4]. Following this initial preprocessing, the results are presented to the user on the monitoring screen. The information obtained from this preprocessing of the maternal/fetal signals (baseline fetal heart rate, FHR variability, accelerations, decelerations and types, etc.) provides part of the data that Foetos requires during its reasoning process [ 5 ] , [26]1281.

FOETOS Foetos has been built with the aid of the knowledge engineering tool Genie [29], which adopts a mixed frame- and rulebased approach. Both control knowledge concerning the way in which diagnoses, prognoses, and other values are sought, and factual knowledge comprising descriptions of paradigms and patient data are contained within frames [ 3 0 ] ,while the rule bases handle inferential knowledge [31I. Although several modifications have been done in the original tool (i.e., explanatory facilities, redefinition of some functions, etc.), the structure of Foetos is in part conditioned by that tool. A . Design Considerations The clinicians who have collaborated in the construction of Foetos established three basic design objectives: 1) the system must incorporate a highly interactive “user-program” interface, 2) explanatory facilities must be included, and 3) the system must be able to adapt the results of the evaluation of a given test to the particular situation of the case under consideration, so as to permit different interpretations with the same test results, in different contexts. In order to meet these design goals, it was decided at the outset that a menu-based structure had to be implemented in the system. Each of the possible options of the menu involves the acknowledgement of both the static and dynamic structures necessary to perform the system reasoning. Once a given option

ALONSO-BETANZOS er

U/.:

FOETOS-SYSTEM

20 1

FOR FETAL ASSESSMENT

TABLE I1 BPP MANAGEMENT T A K I NINTO G ACCOUNT T H E RESULTSOF THE TEST A N D T H E MATERNAL CONDITIONS BFTBCORE

CONDITIONS

MANAGEMENT

10 I

No Special Conditions

Repeat weekly

Mother Diabetic or

&pat

estatlonal Age D 42 weeks

weekly

8

I

Ollgohydramnios

Deliver

No Special Conditions

Repeat Weekly

Oligohydramnios or Perslstent Score 6

I

Deliver

6

No Special Condltlons

Repeat test in 24 hours

;eStationalAge < 28 Weeks

Repeat Test in 24 hours

No Special Conditions

Deliver

ktational Age < 28 Weeks

Repeat test Tor 120

No Special Conditions

Deliver

4

I

2

has been selected, the subsequent inferential process runs in parallel with an explanatory module during the protocol of data input. That means that each time that the user is asked for the value of a given parameter, or a given situation must be explicitly stated, the explanatory routine for that parameter is activated. If such a routine is triggered, the system presents information about the parameter and the context in which it is being studied (Fig. 1). This strategy permits the on-line followup of the system’s reasoning during the consulting session. When the consultation process is completed for a given option, and the report of the case has been rendered to the user, the system triggers, on demand, the “inferences justification module.” This module will cause the display o f the bookkeeping information corresponding to the successfully applied rules’

bodies stated in natural language. This permits the off-line followup of the system’s reasoning once the inferential process has been concluded. To adapt the results of a preliminary evaluation of a given test to the real context of the case being studied, thus allowing adequate diagnoses and therapies, the system includes some high-level hierarchical context’s rules. These structures define the environment in which a given parameter, previously analyzed, must be interpreted. These structures act as “semantic filters” and mostly perform interpretations of the results in the light of certain particular conditions (Fig. 2).

B. Knowledge Base Acquisition

In order to acquire the knowledge embodied by the expert system, a mixed strategy has been used. The basic knowledge

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 38. NO. 2. FEBRUARY 1991

202

@A=-

(IFAIL))

(”

($IF

(TRIPIS (PATIENTNST-INDIVIDUAL-LLEBuLT8 FETAL-m-IS) MILDTACHYCARDIA) (TRIPIS (PATIENT CONDITIONS ACllWIT-UVW =OED)

‘ T N C L U D E (PATENT NSTJNDIVIDUALRESUL~ )SI---..” NORYAL)

(FWT‘(PATIENT REPORT3 &I (CAADR = (FORT ‘(MILD-TACHYCARDIA MESUCEl))))))) Fig. 2. A “context” rule of the system. Structurally and functionally, this rule acts as a semantic filter. The slot FPUT’ . ’ . in the THEN part defines this rule as an “inferential-reporting” rule.

Fig. 3 . Menu structure in Foetos.

structure was, at first, taken from the literature. Once this first approach was completed, a set of structured interviews between knowledge engineers and clinical experts was organized. From these interviews, the basic knowledge based on commonly accepted criteria already present in the literature, was modified and refined with heuristic procedures based on clinical experience. At a later point, unstructured meetings during obstetrical practice took place. In that way we were able to incorporate into the system facts and heuristics that are used clinically, but that are not easily obtained directly from the expert.

C. System Description Foetos is capable of performing five high-level tasks, three tasks relative to the interpretation of the previously mentioned tests, one task devoted to the supervision of labor, and another task which provides a diagnosis concerning the state of the newborn. These tasks are organized mutually independently within the expert system, so that the clinician can choose which task or tasks are appropriate at each moment. Each of these options

include a set of low-level tasks which can be called once a main option has been selected. Several of these low-level tasks are common to certain main options to permit easy interaction among different parts of the expert system even if a concrete selection has been performed (see Fig. 3). As previously noted, the first interaction between the user and Foetos is through a main menu with the following options: 1) NST interpretation, which interprets the results obtained from the analysis of the fetal patterns to obtain a final conclusion about fetal reactivity. The parameters to be studied are the fetal heart rate baseline (BFHR), the long-term fetal heart rate variability (BTBV), the accelerations (frequency, duration, amplitude, percent acceleration time) and the decelerations (frequency), (see Appendix). 2) BPP interpretation, which analyzes the results of a conventional NST together with the results of the ultrasonic observations previously noted using the criteria listed in Table 111. 3) CST interpretation, which performs an analysis of the FHR response to uterine contractions according to the criteria discussed earlier, and listed in Table 11.

ALONSO-BETANZOS er

U/.:

FOETOS-SYSTEM

203

FOR FETAL ASSESSMENT

BPP VARUBLE

NON SIRESS TEST

NORMAL SITUATION

2 acceleratlons of more than 15bpm

ABNORMAL SITUATION

otherwise

Fetal Breathing Movements

1 episode of more than 30” In 30’

otherwise

Fetal Movements

more than 3 body movements in 30’

otherwise

more than one active extension and flexion of limb. trunk or hand

otherwise

pocket greater than 1x1 cm in two perpendicular planes

otherwise

Amniotic Fluid Volume

The data needed to perform these interpretations are obtained from the database records created by conventional monitoring systems [26]-[28], and keyed into the expert system by the operator. The final diagnoses and therapies obtained from such interpretations are performed taking into account the maternal/fetal contextual information, which includes: a) information about the maternal clinical history (both from previous and current pregnancies), b) information about the maternal conditions at the moment of the test (gestational age, medication, fasting, etc.), and c) environmental circumstances which could affect the interpretation of the test (stress, etc.). 4) Labor supervision. This option comprises two main suboptions: a) analysis of the characteristics, etiology, diagnosis and therapy of the diverse fetal patterns, and b) analysis of the characteristics, etiology, diagnosis and therapy of maternal labor patterns [3]. To perform the fetal supervision track, Foetos analyses the FHR baseline patterns, the FHR variability patterns and the contraction-related periodic patterns. Each of the patterns is studied individually and then the system evaluates the whole set of parameters in order to obtain the final diagnosis (i.e., uncompromised fetus, prolonged fetal stress, fetal asphyxia, fetal distress, etc.). Once the final diagnosis has been reached, and taking into account the available information, Foetos suggests an appropriate therapy, if needed (change maternal position, fetal blood sampling, maternal oxygen administration, etc.). To perform the maternal supervision, Foetos evaluates parameters such as cervical dilatation, fetal stage and time in hours since labor was established. Based on such analysis the system will be able to recognize the following patterns: “normality,” “prolonged latent phase,” “slow slope active phase,” “active phase arrest,” “prolonged second stage,” and “failure of descent.” Again, based on these results Foetos will suggest appropriate therapy. Also in this case, the data are obtained from database records created by conventional monitoring systems [321. 5) Neonatal outcome diagnosis. In this case, the system performs an evaluation of a series of indexes associated with the

degree of neonatal well-being. The corresponding diagnoses are obtained based on the variables that are usually analyzed after labor: Apgar score, umbilical cord blood pH, blood gases, relation between birthweight and gestational age at delivery, etc. 131, 1331. The system also performs analysis of other obstetrical indexes as, for example, favorability and morbidity. Some of them are low-level options of the main menu options and are referred in the literature [34]-[37].

D. Structure One of the most important characteristics of Foetos’ architecture is the modularity of its knowledge representation. Apart from the explanatory and justificatory modules, three main structural modules can be identified: a) the static knowledge module, b) the dynamic knowledge module, and c) the control module. The static knowledge module comprises the whole set of frames which are used to represent the so-called static knowledge. Although structurally speaking there is no difference among the different frames included in this module, conceptually speaking we can distinguish several categories of frames [38]. We shall refer as static-static knowledge that representing information which will not be altered during the course of the consulting session. These frames are used to represent wellestablished procedures or to represent the ‘‘literature’’ associated with a given event (i.e., all the possible messages attached to “mild tachycardia,” Fig. 4). Another type of frame which is included in this module incorporates the so-called static-dynamic knowledge. The knowledge here represented, static by nature (no procedural), will be modified in the course of the consulting session by the rules’ conclusions or via user-system interaction. These frames will serve to elaborate a specific model of the case being studied. Also, the final report of the consulting session will be incrementally constructed from some of the frames included in the static-static submodule by selecting the appropriate slots of the corresponding static-static frame and putting it as new slots in the static-dynamic frame “Patient” (Fig. 5). Special effort has been done to represent in adequate

204

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING. VOL. 38. NO. 2. FEBRUARY 1991

(MILD-TACHYCARDIA (MESSAGE1 C'MILD TACHYCARDIA, however .")) (MESSAGE2 PMild Tachycardia appeprr to be related with')) (MESSAGE3 ("theadministration of BETAWb¶ETICSto the mothe?')) (MESSAGE4 C'highm a t e m Activity Level")) (MESSAGES ("the administration of PARASYMPATHETIC BLQCEERS to the mother")) (MESSAGE6 C'gertational age l e u than 30 wcckd')) (MESSAGE7 ("mPterno/fetalcomplex I"1Olr')) (MESSAGES C'fetd Tachyarrhyth")) (MESSAGE9 ~oToxIcosIs")) (MESSAGE10

..

. . .))I

Fig 4 The static-stdtic frame mild-tdchycardid Note the semantic elements appearing in some slot\ which dllow the incrementdl construction of statements

CONDITIONS OEBTATIONAL AGE

AcTIvlTpIEvEL FASTING MEDICATION IYBT-PARAMETRRS

BFHR

BTBY AXEISRATIONS FREQUENCY AMPLITUDE

ALONSO-BETANZOS

cf

U/.:

FOETOS-SYSTEM

205

FOR FETAL ASSESSMENT

(ACCEL&RATIOIYll ($TYPE(IFALL))

(8P

(TRIPIX (PATENT NST-P-8 ACCEIERATIONS) pR&o (TRIPIX (PATIENT N8T-PACCEIERATIONS PRE8ENT DURATION)CE 15) (TRIPIE (PATIENT NST-PARAIYETERB ACCEIERATIONS PRESENT FRwu“CE34

(8oR

(8-

(TRlPIE (PATIENT NST-P-8 ACCEIERATIONS P”TAr”) GE 16) (TRXPIB (PATIENT CONDITIONS GESTATIONAI-ME) GE 32))

(PATIENT NST-PARAIYIETER ACCEIERATIONS PRESENT AMPIEIUDE)GE 10) (TRIPIE (PATIENT CONDITIONS GESTATIONAL-AGE) LT 32))) (TRIPIE(PATIENT N8T-PACCEIERATIONS PRESENT P“9 ‘%DIE

(8-

( C O N C m E (PATIENT NSTJNDIVIDUAL-mULm ACCE-TION-STATUS-IS) NORMAL)))

Fig 6 A “pure inferential” rule in the \ysteni This rule cla\\ihe\ sym bolicdlly the Flatu\ of the accelerations, tdking into account the appropriate

parameter\ pseudo-natural language the conclusions obtained from the use of Foetos. In this respect, certain slots of some of the staticstatic frames, which represent partial conclusions, incorporate semantic elements which allow a “smoothly incremental construction of statements” (see, for example, Fig. 4). Beside these two different types of frames, we can also identify the static control frames which include information to be used by the inference engine in order to accomplish its tasks. This third type of frame incorporates the low-level control structures used by Genie (i.e., param-specs frames, etc.) which among other things govern the way in which parameter values, defined as the terminal elements of a given frame, are sought when required by the rules (i.e., specifying the type of the parameter, its multiplicity, its message format, etc.). The dynamic knowledge module comprises several rulebases [ 341, which represent inferential knowledge relative to particular aspects of the problem. This distribution of the inferential procedures leads to a clear representation of the knowledge in particular areas. Since there are some common aspects in each of the different options of the main menu, the system will only activate the dynamic structures which could be used to perform its reasoning. In this way, only relevant rulebases are loaded from disk, depending on the initial user selection. Also, we can distinguish different types of rules in Foetos: “pure inferential” rules and “inferential-reporting” rules. Pure inferential rules are used at very unspecific levels in Foetos’ reasoning. For example, these rules are employed to classify symbolically a given numerical value and to identify a given situation (Fig. 6). Conversely, the inferential-reporting rules incorporate the contextual information needed in order to interpret previous results and contribute to elaborate the final report by selecting particular slots of the appropriate static-static frame (Fig. 2 ) . The rulebases of the system can be invoked either in backward or forward chaining modes. If all the necessary data are available for the system to reach a specific goal. Foetos employs forward chaining. In this case, the system must have all necessary data in order to carry out its inf.rential process. On the other hand, in backward chaining Foetos does not require predetermined initial data to perform its reasoning. Conversely, the necessary data are obtained by the system through a prede-

predetermined initial data to perform its reasoning. Conversely. the necessary data are obtained by the system through a predetermined strategy that uses header or bookkeeping information (static-control frames), physically contained in the corresponding rulebase. Judicious implementation of backward and forward chaining optimizes the inferential process, while at the same time enhancing user interaction. The third module in the sysem is the control knowledge module in which high level control structures reside. Genie provides control facilities in the form of agenda frames, which list sequential tasks to be performed by the system. These agenda frames could also be considered as part of the submodule of static-control frames; however. we prefer to include in the latter only the low-level control structures which operate in the system. Other structures belonging to this module are the Foetos’ metarules, which control the activation of the appropriate rulebases and also the selective application of certain rules within a given rulebase. In order to control the main menu-based structure of the system, Genie provides the menu-input control frames which, operating on a given static-static frame, define the proper pruning of the latter (Fig. 7 ) .

SYSTEMEVALUATION Validating an expert system of Foetos’ characteristics requires its utilization in the environment in which the system is applied. The validation process we have chosen consists of the comparison of the man-machine performance [39]. This has been done taking into account not only the final results after a consulting session, but also the initial and intermediate results, the reasoning schemes as well as the logic of the inferential processes. The validation of Foetos was accomplished through three steps:

A. First Validation-ReformCycle The knowledge base of the original version of the system was tested on 87 cases, randomly chosen from files of the Ob-Gyn Department of the Hospital General de Galicia, Spain. The cases formed a representative sample. Modifications were made until our collaborating experts were satisfied with the results.

LEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 38. NO. 2, FEBRUARY 1991

206

(PROMPT-SPECS

man

‘(PATENT IDENTIFIERS

8%n

n

(E

sgn 8%n

W

O

n 8%n c€H 8%n (E

sgn

81Y

Fig. 7. An example of a menu-input control frame. The slots of “past” and “present” (risk factors) will be pruned by the user and the results will be stored (in the form of new slots) in the slot “risk factors” of the frame “patient.” The second part of this frame incorporates a procedure which stores the frame “patient” in a file which has the same “name” that the hospital number of the patient.

I

I‘

I

Fig. 8. All the possible interrelations among the modules and submodules of Foetos.

B. Retrospective Validation After this first evaluation, a retrospective sample of 20 patients was chosen for whom NST, BPP, and CST tests have been done, and whose labor data was available. The patients were chosen from the files of the Maternal-Fetal Section of the Ob-Gyn Department of the Medical College of Georgia,Augusta, GA. All these 20 cases represent a population in which all data were useful because of their good quality. In all cases the percentage of signal loss oscillated within 0-12%. Table IV shows the quantitative criteria associated with the level of agreement between human experts and the expert sys-

tem for the different possible diagnoses in each of the system’s tasks. In this table, the diagnoses have been organized according to a scale gradually ranging from the most favorable to the least favorable for each diagnostic possibility. For each couple of clinician/Foetos diagnosis, we have made a numerical assignment in which the value of 0 represented diagnostic agreement. The lack of agreement is characterized by a number (indicating the absolute disagreement between the two diagnosis), and a positive or negative sign (indicating the sense of direction of the disagreement: positive if Foetos’ diagnosis was more favorable, negative if Foetos’ diagnosis was less favorable) [40]. The application of these criteria to the 20 patients

ALONSO-BETANZOS er

U/.:

FOETOS-SYSTEM

207

FOR FETAL ASSESSMENT

TABLE IV Q U A N T I T A T I V E C R I T E RASSOCIATED IA W I T H T H E LEVEL OF AGREEMENT BETWEEN HUMANEXPERTS A N D FOETOS,FOR T H E DIFFERFNT POSSIBLE DIAGNOSES I N EACHOF T H E SYSTEM’S TASKS

NON STRESS TEST

k 1 REACTIVE LNDECIDED IUNREACTIVE PClvrtFI.

UNDECIDED UNREACIIVE

CONTXACTION STREW TEST

MOIIBIDITI NORMAL UNDECIDED

ABNORMAL

Negatfve False Negattve I

False PwtUve

2

1

0

-1

hsltlve

3

2

1

0

I I

UNDECIDED

I

1

I

ABNORMAL

1

2

1

0

1

I

-1

1

0

I 1

F’EMLBIOPHYSEALPROFIIX REACTIVE bNDEClDED IUNREACllVE I

I t

UNDECIDED

I

I 1

I

-1

0

REACTNE

1

I

I

I

I

O

-2

I

1

I

-

l

Ruaonable

I

1

Sunplclous

Non Rcp.onablc 0”rU

selected for this study, for each type of diagnosis, produced the results shown in Table V (percentage of total agreement). To verify the degree of man-machine agreement regarding possible therapeutic decisions, the criteria employed was in terms of absolute agreement-disagreement between the recommendations made by Foetos and the clinician. The corresponding results are shown in Table VI. The results indicate a substantial agreement between the clinician and the expert system described herein. Man-machine disagreements appearing in Table V are based exclusively on whether there was any disagreement at all, irrespective of the category in which the disagreement occurred. With respect to the results shown in Table VI dealing with therapeutic recommendations, agreement was defined exclusively when the entire set of therapeutic recommendations offered by the system agreed with that offered by the physician. For example, in the case of patient 15 who exhibited a labor pattern consisting of persistent late decelerations, the physician recommended the following: change maternal position, administer oxygen to the mother, and perform an analysis of fetal blood. For this case, and taking into consideration the correspondent contextual information of that patient, Foetos recommended: change maternal position, administer oxygen to the mother, and administer tokolytic

3

0

agents to the mother. This result was considered as an overall disagreement.

C. Prospective Validation The results of the retrospective evaluation helped us to point out the errors of the system, indicating where revisions were needed. Taking these into account, the knowledge base of the system was refined. After that, a prospective blind study was designed in which the NST, BPP, CST, and labor data of the patients were analyzed by both the physician and the expert system. From this analysis, a prognosis of neonatal outcome (if the same trends were continued) was established and compared with the early neonatal outcome. The sets of patients included in the study were the following: 228 patients for whom NST and labor data were available, 33 patients whose BPP and labor data were available, and 85 patients who underwent both NST and BPP and whose labor data were available. Since CST has become less frequent after the establishment of BPP as a routine clinical screening test, the quantity of available cases was not sufficient to evaluate this modality. Table VI1 shows the results obtained for man-machine performance and the outcomes of the newborns. We considered as

208

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 38, NO. 2. FEBRUARY 1991

TABLE VI RESULTSOF THE MAN-MACHINE PERFORMANCE IN THE EVAI-UATION OF FOETOS'THERAPEUTIC SUGGESTIONS

a bad outcome any of the following: Apgar less than 7 at 5 min, arterial pH and blood gases reflecting metabolic problems, and abnormal relation between birthweight and gestational age at birth; conversely, we codified as good outcomes those babies whose Apgar score was more than or equal to 7 at 5 min, whose arterial blood gases reflected a normal situation or an acute respiratory problem, whose birthweight was normal, and who did not exhibit any major fetal abnormality. We also included as a bad outcome one neonatal death which took place in the BPP patients' group, and which was accurately predicted by both the physician and Foetos. In examining the results, it must be pointed out that those corresponding to the "bad prognosislgood outcome" group reflect good clinical management rather than a failure of the ob-

stetrician. Once this is acknowledged, the results obtained indicated a good level of agreement between physician and expert system, which was already expected after the retrospective validation. It also can be seen from Table VI1 that the percentage of success in predicting the early neonatal outcome is satisfactory.

DISCUSSION AND FUTURE WORK This prototype has been implemented on a 386 machine and has been developed using the IQLisp version of Genie. The knowledge contained in Foetos has been organized into five independent, principal tasks. The static knowledge, in which we distinguish several categories, has been organized in the form of frames. Dynamic knowledge has been implemented in the

ALONSO-BETANZOS er a / . : FOETOS-SYSTEM

209

FOR FETAL ASSESSMENT

TABLE VI1 RESULTS OF MAN-MACHINE PERFORMANCE FOR T H E PROSPECTIVE EVALUATION OF T H E SYSTEM PERCENTAGES OF

CA”

GREEMEI’TT/DISAGREMEI’TT

-1

Bd5

160

85% Agreemat

Bad (man)/ Bad (machine)

physician: 95%

-2

Bed:=

34 Good [man) / Bad ( m a w

Goat 14

Bed:3

17

1

Bad

(man) /

Good ~machhc’

Expatsystan: 84%

Good.2 Bed: 15

hod (man) / Good (Innchine)

Good: 15

7996 Agrement

126-1

phyeldan: 91% Bad (man) /

Bad (mnchlne)

21% Msagnement (7 ==4

hod

(IMn)/

Good (Innchine)

Phyaiclan: 94% Bad (man)/ Bad (muchlnc)

Good: 1

M:6 Good (man) / Bad (nuchinc)

29% l n q p ” t (25 -1

I

m

25 Bad (man)/ Good (machine)

0

form of production rules, each of which being applicable to a specific knowledge domain. This defines a hybrid of a rule-based system and a frame-based system. The rules of Foetos, which contain procedural and inferential knowledge, mostly use the factual information contained in static-static frames to elaborate conclusions and reports, which will be subsequently stored in the static-dynamic frames. Dynamic control of the system is optimized through metarules. The knowledge structure of Foetos involves establishing different levels of abstraction that are reached as the inferential process proceeds, resulting in a transition from numeric, through diagnosis to therapy. The structure of the explanatory and justification facilities, perhaps the most successful part of the program, provides an excellent means to follow the system’s reasoning, as well as facilitates further modifications of both, the inferential process and the knowledge base. The performance results obtained to date do not substantially differ from those obtained by other expert systems in other clinical fields [41], [42]. However, evaluation of expert systems is always a complicated problem and sometimes the performance results that can be obtained may not reliably evaluate the effec-

Good:=

I

G0ad:O

M : O

I

2

Jhpcrt systan:68%

I

tiveness of such systems [43]. As far as Foetos is concerned only its actual use can conclusively demonstrate its validity or reveal possible design flaws. Foetos, which is actually being used to assist obstetricians in clinical practice as well as part of the training of residents, is the intelligent element of a larger computerized system, in which several conventional monitoring units are used, depending on the task under consideration. At present, the system is not able to acquire directly from the databases the information needed to perform its reasoning. This presents a problem which sometimes limits its use. Despite the fact that some information has to be acquired directly from the user, the quantity of numerical data required by the expert system (variable depending on the first selection from the menu) is not too large. Moreover, this task can be accomplished by personnel not specially trained in obstetrics. However, to obtain a final product which could link with the hospital databases is an attractive goal. We are now working on the translation of Foetos’ knowledge base into the NEXPERT [44] environment, which allows this latter linkage, in order to free clinical personnel from the sometimes tedious task of data input.

210

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 38, NO. 2. FEBRUARY 1991

REFERENCES

APPENDIX This section includes definitions of some of the clinical parameters more commonly used by the system. A . Baseline Heart Rate The baseline heart rate is that occurring between contractionassociated changes (periodic changes). In the absence of contractions the baseline is calculated as an average of the FHR each minute, and integrated for the 30 min segments used for the antenatal tests.

B. Fetal Heart Rate Variability The term refers to the irregular fluctuations noted on a tracing from a cardiotochometer. Foetos uses long-term variability and admits three different scales: 1) statistical variance V of FHR in beatslminute, calculated from N

v2 = ( l / N )

c IF,

I =

-

I

F12

(1)

where F is the basal FHR, F, is the value of each FHR point between contractions, N is the number of data points, and Vis the statistical variance. 2) The standard deviation of the FHR divided by the mean basal heart rate, which is the parameter currently used by the maternal-fetal section of the Ob-Gyn department in the Medical College of Georgia, and 3) variability V‘ coded as shown below, which is the parameter currently in use in the High-Risk section of the Ob-Gyn department in the Hospital General de Galicia, Spain.

V 0-2 3-6 7-1 1 12-15 > 15

V’

0 +1

$2 +3 +4

C. Accelerations They are defined as increments in the fetal heart rate, lasting more than 15 s and with an amplitude greater than 10 or 15 bpm, depending on the gestational age.

R. Caldeyro-Barcia, C. MCndez-Bauer, J . J . Poseiro, L. A. Escarcena, S. V . Pose, J . Bieniarz, I. Arnt, L. Gulin, and 0. Althabe, “Control of human heart rate during labor,” in The Heart and Circulation in the Newborn and Infant, De Cassels, Ed. New York: Grune & Stratton, 1966, pp. 7-36. E. H. Hon, An Atlas of Fetal Heart Rate Patterns. New Haven, CT: Hardy, 1968. J. T. Parer, Handbook of Fetal Heart Rate Monitoring. Philadelphia: W. B. Saunders, 1983. C . Hernandez Sande and J . E. Arias Rodriguez, “Syntatic pattern recognition of foetal stress,” J . Biomed. Eng., vol. 6, pp. 97101, 1984. C. Hernandez Sande, J . E. Arias Rodriguez, and L. Gdmez Gonzalez. “A perinatal monitoring display based on the fetal topogram,’’ IEEE Trans. Biomed. E n g . , vol. 33. pp. 6-12, Aug. 1986.

M . L. Gimovsky and S. L. Bruce, “Aspects of FHR tracings as warning signals,” Clin. Obstet. Gynecol., vol. 29, no. I , pp. 5163, 1986.

A. Parnes LaSala and H. T. Strassner. “Fetal death,” Clin. Obstet. Gynecol., vol. 29, no. I , pp. 95-103, 1986. A. Barr and E. A. Feigenbaum, The Handbook of Artificial Intelligence. Kaufman, 1981. R. Davis and J . King, “An overview of production systems,” in Machine Intelligence, E. W. Elcock and D. Michie, Eds. New York: Wiley, 1977, pp. 300-332. F. Hayes-Roth, D. A . Waterman, and D. B. Lenat, Building Expert Systems. Reading. MA: Addison-Wesley, 1983. E. H. Shortliffe, S. G. Axline, B. G. Buchanan, T. C. Merigan, and S. N . Cohen, “An artificial intelligence program to advice physicians regarding antimicrobial therapy,” Comput. Biomed. Res., vol. 6 , pp. 544-560, 1973. S. M. Weiss, C. A. Kulikowski, and A. Saphir, “Glaucoma consultation by computer,” Comput. Biol. Med., vol. 8, pp. 25-40, 1978.

R. A. Miller, H. E. Pople, and J. D. Myers, “INTERNIST-I: An experimental computer-based diagnostic consultant for general internal medicine,” New Eng. J. Med., vol. 19, pp. 468476, 1982.

S . M. Finkelstein, P. L. M. Kerkhof. and M. Okada, Eds., Special Issue on Medical Applications of Artificial Intelligence and Information Systems, IEEE Trans. Biomed. Eng., vol. BME-36, May 1989. A. D. Haeri, J. South, and J . Naldrett, “A scoring system for identifying pregnant patients with high risk of perinatal mortality,” J . Obsret. Gynecol. Brit. Commun., vol. 81, pp. 535-538, 1974. S . B. Effer, “Management of high-risk pregnancy,” Canadian Med. Assoc. J . , vol. 101, pp. 55-63, 1969. L. R. Evertson. R. J . Gautier. B. S . Schifrin. and R. H. Paul,

“Antepartum fetal heart rate testing. I. Evolution of the nonstress test,” Amer. J . Obstet. Gynecol., vol. 133, no. 1, pp. 29-33, 1979.

D . Decelerations

They are defined as decrements in the fetal heart rate, lasting more than 15 s and with an amplitude greater than 15 bpm.

ACKNOWLEDGMENT The authors wish to thank the Department of Electrical and Biomedical Engineering of Vanderbilt University, Nashville, T N , for generously making Genie available to them. In the construction of the knowledge base, we have received the assistance of Dr. Iglesias and Dr. Ucieda from the High-Risk section of the Ob-Gyn Department of the Hospital General de Galicia, Spain. In further modifications of the knowledge base and in the validation of the system, Dr. Devoe and Dr. Castillo, from the Maternal-Fetal section of the Ob-Gyn Department of the Medical College of Georgia, Augusta, have collaborated.

F. A. Manning, L. D. Platt, and L. Sipos, “Antepartum fetal evaluation: Development of a fetal biophysical profile,” Amer. J . Obstet. Gynecol., vol. 136, no. 6, pp. 787-795, 1980. J . V . Collea and W. M. Holls, “The contraction stress test,” Clin. Obstet. Gynecol., vol. 25, no. 4 , pp. 707-717, 1982. S . B. Thacker and R. L. Berkelman, “Assessing the diagnostic accuracy and efficacy of selected antepartum fetal surveillance techniques,” Obstet. Gynecol. Survey, vol. 41, pp. 121-134, 1986. J. P. Lavery, “Nonstress fetal heart rate testing,” Clin. Obstet. Gynecol., vol. 125, no. 4 , pp. 689-705, 1982.

L. D. Devoe, J. McKenzie, N. Searle, and D. Sherline, “Nonstress test: Dimensions of normal reactivity,” Obstet. Gynecol., vol. 66, no. 5 , pp. 617-620, 1985. M. Ray, R. K. Freeman, and S . Pine, “Clinical experience with the oxytocin challenge test,” Amer. J . Obstet. Gynecol., vol. 114, no. I , pp. 1-9, 1972. B. S. Schifrin, “The rationale of antepartum fetal heart rate monitoring,” J . Reprod. Med., vol. 2 3 , no. 5 , pp. 213-221, 1979. F. A. Manning, I . Morrison, and I . R. Lange, “Fetal assessment based on fetal biophysical profile scoring: Experience in 12,620

FOETOS--SYSTEM F O R FETAL ASSESSMENT

21 I

referred high-risk pregnancies.” Anier. .I. ODstet. Gynecol.. vol. 151, pp. 343-357, 1985. (261 J. R. Searle, L. D. Devoc, M. C . Phillips. and N . S. Searle. “Computerized analysis of resting fetal heart rate tracings,” Obstet. Gynecol., vol. 71. no. 3 . pp. 407-41 I , 1988. 1271 L. D. Devoe and J. R. Searle, “Thc fetal biophysical profile: Antepartum assessment using a programmed microcomputer,” J . Clin. Eng., vol. 1 I , no. 4, pp. 285-289, 1986. [28) C . Hernindez Sande, A . Alonso Betanzos. and J . E. Arias Rodriguez. “Automatic unit for monitoring and diagnosis with the contraction stress test,” Med. Biol. Eng. Coinput., no. 26. pp. 410-415, 1988. 1291 H. S . H. Sandell, “GENIE user’s guide and referencc manual.” Tech. Rep. 84-003, Dep. Elect. Biomed. Eng.. Vanderbilt Univ., Nashville, TN. July 1984. 1301 M. Minsky. “A framework for representing knowledge,’’ in The Psycliology of CornputPr Vision, P. Winston. Ed. New York: McGraw-Hill. 1975. [31] B. G. Buchanan and E. H. Shortliffe. Ride-Based Expert Systein.s. Reading. MA: Addison-Wesley, 1984. 1321 C. Hernindez Sande, A. Alonso Betanzos. and J. E. Arias Rodriguez, “ A n aid to obstetrical decision-making,” 1nno\~. Tech. B i d . Med., vol. 7. no. 5 . pp. 545-552. 1986. 1331 J. S. Drage and H. Berendes. “Apgar scores and outcome of the newborn,” P r d . Clin. North Anier.. no. 13. pp. 635-643. 1966. 1341 S. Hulkko and M. Kataja. “Forecast model for the outcome of a pregnancy,” Lect. Notes Mcd. Iirfornirit., pp. 176-179, 1982. 1351 H. R. Rey, L. S . James, H. E. Fox, J. M. Driscoll. and H. Samshi. “A computer weighted scoring system for the prediction of fetal and neonatal outcome,” IEEE Troris. Biotned. Gig.. vol. 31, no. 3, pp. 14-21, 1984. [36] J. G . Nijhuis, H. F. R. Prechtl. C. B. Martin, Jr.. and R. S. G. M. Bots, “Are there behavioural states in the human fetus?,” Early Hum. De\,., vol. 6. pp. 177-195. 1982. [37] E. H. Bishop, “Pelvic scoring for electivc induction.” Obstet. Gynecol., vol. 24, no. 2, pp. 266-268, 1964. 138) C. Hernandez-Sande, V . Moret-Bonillo. and A. Alonso-Betanzos. “ESTER: An expert system for management of respiratory weaning therapy,” IEEE Trcins. Biorned. E i i g . , vol. 36. no. 5 , pp. 559-565. 1989. 1391 R. M. O’Keefe, 0. Balci. and E. P. Smith. “Validating expert system performance.” IEEE Expert, vol. 2. no. 4. pp. 81-89, 1987. [40] A. Alonso-Betanzos. L. D. Devoe. R. A. Castillo. V. MoretBonillo, C. Hernandez-Sande. and N. S. Searle. “FOETOS in clinical practice: A retrospective analysis of its performance,” Art$ Intell. M e d . , vol. I , no. 2, pp. 93-99, 1989. 141) J. S. Aikins, J . C. Kunz, E. H. Shortliffc, and R. J . Fallat. “PUFF: An expert system for interpretation of pulmonary function data,” Coinput. Biorned. Res.. vol. 16. no. 3 . pp. 199-208. 1983. 1421 J. R. Slagle, S. M. Finkclstein, L. A. Leung, and W. J. Warwick, “Monitor: An expert system that validates and interprets time-dependent partial data based on a cystic fibrosis home monitoring program,” IEEE Truns. Binnied. Eng.. vol. 36. no. 5 . pp. 552-558, 1989. 143) J. R. Boston, “Automated interpretation of brainstein auditory evoked potentials: A prototype system,” IEEE Trans. Biorned. Eng.. vol. 36. no. 5 . pp. 528-532, 1989. [44] NEXPERT Object. Palo Alto, CA: Neuron Data. 1988.

Amparo Alonso-Betanzos (M’87) was born in Vigo, Spain. in 1961. She received the degree in chemical engineering from the University of Santiago de Compostela, Spain, in 1984 and a post-graduate degree for research in perinatal monitoring in 1985. Since joining the University’s Department of Applied Physics, she has received the doctorate “cum laude” and “premio extraordinario” in 1988 for work on obstetrical expert system’s development. She was a postdoctoral fellow in the Department of Biomedical-Engineering Research, Medical College of Georgia, Augusta. Her main current research interest area is in medical expert systems. Dr. Alonso-Bctanzos is a member of the IEEE-EMBS Society. the IEEE Computer Society, and the ACM.

ALONSO-BETANZOS

(>I < I /



Vicente Moret-Bonillo (M’87) was born in

Valencia, Spain, in 1962. He was graduated with the degree in physical chemistry from the University of Santiago de Compostela, Spain, in 1984, and the first post-graduate degree for research on control and monitoring of hemodynamic variables in 1985. Since joining the University’s Department of Applied Physics, he earned his doctorate “cum laude” in 1988 for work on the application of artificial intelligence techniques to intensive respiratory treatment of patients depending on mechanical ventilation. He is currently a postdoctoral fellow in the Department of Biomedical Engineering Research, Medical College of Georgia, Augusta. His current research area is in knowledge representation and in the application of knowledge engineering techniques to dynamic systems. Dr. Moret-Bonillo is a member of the IEEE-EMBS Society, the IEEE Computer Society. the ACM, and the Mathematical Association of America.

Carlos Hernandez-Sande (M’87) was born in

Santiago de Compostela, Spain, in 1945, where he received the doctor’s degree in science in 1971. He is currently Professor of Applied Biophysics and director of the Laboratory of Applied Biophysics and Artificial Intelligence. His fields of interest are digital signal processing and intelligent monitoring systems for intensive care units and perinatology. Dr. Hernhez-Sande is a member of various IEEE Societies.

Foetos: an expert system for fetal assessment.

Assessment of the fetus in a high-risk pregnancy uses a variety of tests for screening and continued detection of in utero compromise. This paper desc...
1MB Sizes 0 Downloads 0 Views