Computer Simulation Modeling and Birth Outcome

Lewis Mehl-Madrona, M.D., Ph.D.

 

 

Introduction:  Historically birth outcomes have been relatively unpredictable to physicians and midwives. The development of systems dynamics computer simulation methods provided a tool through which prediction might be realized. The goal of prediction is to better allocate treatment resources and to prevent problems from occuring, both in terms of prevention before the fact and detection of problems in an earlier stage than previously possible.

 

Methods:   (Predicting general outcome.) Subjects were recruited from local obstetrics and midwifery practices. Patients who were willing to be interviewed were contacted by a member of the research team to complete questionnaires, authorize release of medical records and receive an oral interview. The sampling was consecutive. Subjects were usually interviewed once each trimester. Modeling studies have been carried out on almost 500 women, 118 of whom were prospectively studied. The sample was enriched with 33 cases of fetal demise in which data was collected immediately after delivery.

 

Data was collected on heroin and cocaine use, amount of alcohol consumption, marijuana use, caffeine use, cigarette smoking, exercise, number of previous deliveries, perceived marital quality, prenatal bonding, estimated date of confinement, perceived life stress, activity level and conventional obstetrical risk factors. This data was collected as vector data over time. Data was collected for the past 24 to 132 months, depending upon the patient's recall. An "Attitudes toward Pregnancy" questionnaire [1] was given to record feelings of closeness toward the baby.

 

The infant's condition at birth rating was developed as a simple scale where 0 = normal condition, 1 = 1 min Apgar < 7, 5 min Apgar > 8; 2 = 5 min Apgar score > 5 and < 8; 3 = 5 minute Apgar score > 1 and <= 5; and 4 = 5 minute Apgar score <= 1. An additional point was given for transfer to neonatal intensive care, three points for hospitalization longer than 10 days and 4 points for death. Thus, our 0 to 8 score reflected severity of infant difficulties after birth.

 

A theoretical model was developed from the literature to explain the hypothesized effects of alcohol, drugs and stress. Differential equations were written to facilitate computer simulation of our theoretical model. The relationships rendered mathematical in the model were taken from a meta-analysis of the obstetrical risk literature reported elsewhere [2]. A wide range of fetal condition was present in the sample (Table 1).

Predicting Premature Labor with computer modeling:   The purpose of this research was to determine the potential usefulness of DSM for psychosocial and biomedical research and on predicting which women are at risk for premature labor. A dataset was obtained from the Fetal Alcohol Research Center of Wayne State University and consisted of data derived from 650 low income women who delivered in five hospitals in Detroit and Wayne County, Michigan and were interviewed 2-6 days postpartum.

 

Variable domains and measures: Variables from the interview and medical chart had been initially grouped into domains as follows:

 

1. Social support with two social support factors identified as representing Intimacy and Comfort, respectively.

 

2. Medical risk: A composite medical risk score was derived from summing the weighted individual risks as follows: weights of two points each for diabetes, hypertension and previous low birth weight infant, and a weight of one point for age < 16 years or > 35 years and for hematocrit < 28%.

 

3. Habits: Use of tobacco, alcohol and illicit drugs was assessed over the pregnancy.

 

4. Prenatal care: Two components characterized prenatal care: source and amount.

 

5. Other variables: Amount of insurance was coded as none, present during part of pregnancy (1) or over the entire pregnancy (2); feelings about pregnancy was coded as a five point scale from very unhappy to very happy; how hopeful the woman was about the future was a four point scale from not at all hopeful to very hopeful; and the month of gestation pregnancy was first suspected was calculated based upon the calendar month pregnancy was suspected, month of delivery and length of gestation. Birthweight was recorded in grams.

 

Statistical analysis:  Discriminant function analysis was developed on half the sample (randomly selected) and then tested on the other half of the sample. Discriminant Function Analysis. Variables for entry into the discriminant function analysis (DFA) consisted of race, suspect (month the woman first suspected she was pregnant), risk (weighted medical risk not including substance abuse), hopeful (how hopeful the woman was about the future), drinking (total drinking during pregnancy), care (where the woman goes for most of her care, insurance (level of health insurance), firstfeel (how the woman first felt about being pregnant), drugs (total drugs used during pregnancy), comfort (how much social support the woman felt that she had), parity, smoking (during pregnancy), intimacy (how intimate the woman feels with her support providers) and the Kessner variable. DFA was compared with the performance of the systems dynamics computer simulation model (DSM).

 

Model operation:  The heart of the computer model (DSM) is a differential equation which controls the variable called "Time to Start Labor." When the current time within the simulation equals "Time to Start Labor," Gestational Age is fixed along with Birthweight (dependent upon Gestational Age). "Running" the simulation consists of starting at 26 months prior to the estimated date of confinement to put the model into homeostasis prior to conception and activation of the pregnancy module. When the "Conception" variable switches from "0" to "1", the pregnancy module is activated, Time to Start Labor has a value of the current time plus 40 weeks, and the factors listed in Table 1 begin to increase or decrease this 40 week value.

 

Results (General Outcome):   The model predicted correctly 83% of the time, using a normal (Scores of 0 to 2) versus abnormal (Scores of 3 or more) discrimination. None of these women were predicted to be at high risk by their health care providers. All but one of the fetal demise cases were correctly predicted as severe outcomes.

 

We found no apparent effect of alcohol until a woman reached the ten drinks per week category. At this point, there was an abrupt decline. By the 16 drinks per week level, there can be as much as an 18-fold effect on fetal condition at birth. The dramatic nature of this effect may also exist because of our inclusion in this sample of 33 women with medically unexplained fetal demise, for whom stress and substance utilization was found to play an important role in contributing to the demise [2].

 

Among women with low stress with no other substance utilization, the effect on infant condition at birth of even 20 drinks per week was minimal in this group. The shape of the curve, however, was the same as the curve for women who showed maximal effect and again that the level at which effect occurred was at the 10-12 drink per week level.

The data for the "low stress - one other substance" case was intermediate between the two previously discussed extremes. Again the shapes of the curves were very similar with the same 18-fold difference in effect for the "high stress - one other substance" case, and a three-fold effect for the other cases.

 

Thirty-four cases of medically unexplained fetal demise were included in the sample, and these cases contributed to the high stress - the two substance cases in which higher levels of alcohol contributed to an eighteen fold decrease in fetal condition at birth. The eighteen-fold effect cases were associated with fetal demise.

For evaluation of this model, the proper statistical approach is the use of a multinomial distribution to study the correctness of each combination of predictions. Each case is treated as an independent trial. The number of variables per case is irrelevent, since a deterministic result is produced for several outcome measures. The correctness of these outcome measure is what is studied to show accuracy of predictions. Each prediction should be treated similar to a coin toss experiment with several coins (equal to the number of outcome variables). To evaluate our results we would ask for the probability of predicting correctly within this sample through a random stochastic process. A "normal" fetal condition at birth (outcome ratings of 0 to 2) has about a 63% chance of occuring in the sample. The probability of predicting correct assignments randomly compared to the results obtained results in statistical significance with p < 0.001 for the results of the present model.

 

Predicting Prematurity/Low birth weight: Table 2 shows the results of comparing DSM and DFA. Of the 78 women in the sample who did not have low birthweight infants, DSM correctly identified 74 of them. It missed 4 of these women, predicting them as being in the low birth weight group. DFA correctly identified 69 of these same women, misclassifying nine of them into the low birth weight group. The prematurity/low birth weight rate of this group of inner city Detroit women was 22%, a common rate for this and similar other populations. Of the 22 women who had low birth weight infants, DSM correctly identified 16 and missed 6. DFA correctly identified 11 and missed 11. The adjusted phi-square method was used to compare the two procedures. DSM performed statistically significantly better than DFA at the p=0.01 level.

 

Table 2:   Results of comparing DFA and DSM predictions

  ...........Outcome DFA Outcome DSM

  .................LBW..Not LBW ..LBW ..Not LBW

Actual LBW............11 11 ......16 6

Actual Not LBW .......9 69 ........4 .......74

................Total Test Data Set (n = 100)

 

  ........Comparison DFA Comparison DSM

Efficiency (%) .......82 .................90

Sensitivity (%) ......65 .................80

Specificity (%) ......86 .................92

PPV (%) ......64 .................72

RR (%) .......5.9 ..............14.2

Adjusted Phi .......0.54 ..............0.74

Adjusted Phi2 (%)..29 .................55

 

Discussion:  Models such as this one can be used by clinicians to help pregnant women assess their individual behavior and make changes when necessary to reduce their risk. Actual interaction with the model could help influence patient compliance through showing the woman the effects of her behavior and allowing her to interact with the model to assess different changes she might consider making.

 

Both DFA and DSM showed high levels of prediction, higher than other reports available in the literature. The reason for this can be hypothesized to be due to the inclusion of variables related to psychosocial status and lifestyle. Both techniques showed that increased quality and quantity of prenatal care can improve pregnanccy outcome. Both techniques showed that substances (drugs,alcohol and tobacco) can be major pregnancy risks. DSM, while investigator intensive, can produce comparable results to conventional, sophisticated multivarate analytic techniques. It may outperform these techniques in modeling complex inter-relationships among variables and in allowing the use of complex systems theory to be applied to the prenatal period.

 

The theory of low birthweight generated by the DSM model states that medical risk is reduced by positive psychosocial factors and increased by negative psychosocial factors (intimacy and comfort). The woman's feelings about the baby at the time she learns she is pregnant and her levels of hope and/or depression influence birthweight. Drugs and alcohol decrease birthweight. Tobacco has the most powerful effect of the substances considered. Good prenatal care increases birthweight, possibly through providing adequate attention or through improving psychosocial risk.

 

The theory developed from DSM suggests that medical risk is non-linear, reduced in the young (less than 18 years old), in women who have had prior children, and in women with positive psychosocial circumstances who do not smoke, and among women who are hopeful about the future and very much want the baby. A poor, anxious, 12 year old who initially felt miserable about being pregnant and denied it until the sixth month would experience an eight-fold greater effect on "Time to Start Labor" than a similar 21 year old. A similar 16 year old would experience a three-fold effect.

 

References:

 

1. Newton R.W., Webster P., Binu P.S., Maskrey N., Phillips A.B., Psychological stress in pregnancy and its relation to the onset of premature labor. British Medical Journal 1979; 2: 411-413.

 

2. Mehl L.E., Systems dynamics computer modelling to predict birth risk among medically low risk women. Int'l Journal Prenatal and Perinatal Studies 1990; 1: 47-68.

Published in the Journal of the American Board of Family Practice in the March issue for 1998.

Back to Midwifery Linking page