Contraceptives and sexual behaviours in predicting pregnancy rates in HIV prevention trials in South Africa: past, present and future implications

Abstract

Objective:

Despite all efforts, high pregnancy rates are often reported in HIV biomedical intervention trials conducted in African countries. We therefore aimed to develop a pregnancy risk scoring algorithm for targeted recruitment and screening strategies among a cohort of women in South Africa.

Methods:

The study population was ~10,000 women who enrolled in one of the six biomedical intervention trials conducted in KwaZulu Natal, South Africa. Cox regression models were used to create a pregnancy risk scoring algorithm which was internally validated using standard statistical measures.

Results:

Five factors were identified as significant predictors of pregnancy incidence:<25 years old, not using injectable contraceptives, parity (<3), being single/not cohabiting and having ≥ 2 sexual partners in the past three months. Women with total scores of 21–24, 25–35 and 36+ were classified as being at “moderate”, “high”, “severe” risk of pregnancy. Sensitivity of the development and validation models were reasonably high (sensitivity 76% and 74% respectively).

Conclusion:

Our risk scoring algorithm can identify and alert researchers to women who need additional non-routine pregnancy assessment and counselling, with statistically acceptable accuracy and robustness.

Introduction

After more than three decades of research, the epidemic in South Africa is still alarmingly high and best described as a hyper-endemic due to its generalised and clustered nature [1,2,3]. HIV prevalence was estimated as high as 40% among antenatal clinic attendees in KwaZulu-Natal, which is the most populated province of South Africa [4]. Over the past decade, the HIV prevention research has been accelerated and primarily focussed on developing biomedical interventions including several vaginal-microbicides, pre-exposure prophylaxis (PrEP) and vaginal-rings [5–9]. While PrEP and the vaginal rings [10,11] have been shown to have some level of efficacy, all the microbicides trials have reported negative results [5–9].

Besides poor adherence to the study products, high pregnancy rates have also been frequently reported as one of the major challenges impeding the success of HIV prevention trials conducted in the region [10,12,13]. Since the safety of the investigational products are usually not well established, study protocols require pregnant women to be taken off the study products and/or withdrawn from the trials. As a result, high pregnancy rates can compromise the trials’ planned sample size and power to show the efficacy of an intervention product under investigation [14,15].

Implementing effective pregnancy prevention programs has been proven to be challenging in the biomedical intervention trials conducted in South Africa [16–18]. Although, there has been extensive research to identify the risk factors for high pregnancy rates in HIV prevention trials, to date, there has not been an attempt to develop and validate a simple scoring tool which can estimate and monitor women’s risk of pregnancy during a trial. We aimed to develop a simple risk prediction algorithm to identify the minimum set of data to estimate women’s risk for pregnancy with statistically acceptable reliability and robustness [19]. Given the significant impact and implications of the high pregnancy rates in biomedical intervention trials, it is therefore crucial to estimate and monitor the participants’ risk of pregnancy during the study follow-up. We hypothesised that women’s risk of pregnancy may change during the trials and needs to be assessed regularly. A combined dataset from approximately 10,000 women enrolled in six biomedical intervention trials conducted in KwaZulu-Natal, South Africa, was analysed to develop our pregnancy risk assessment algorithm (2002–2017) [5–11].

Our prediction algorithm may potentially have significant applications in clinical research setting by identifying and alerting researchers to women who need additional non-routine pregnancy risk assessment and counselling, with statistically acceptable accuracy and robustness. To our knowledge, this is one of the most comprehensive and certainly the largest study to attempt to develop a scoring tool which can be used in a research setting by study personal to assess women’s changing risk of pregnancy during the study follow-up and can alert site personal to women who need additional non-routine monthly pregnancy assessment and counselling.

Methods

We used a combined data from 9,948 sexually active women at reproductive age (16–49 years) who enrolled in one of the six phase-II/III HIV prevention trials during the period of 2002–2017 in Kwa-Zulu Natal, South Africa [5–11]. The combined population from the Durban area was analysed. All six trials used broadly the similar testing and diagnostic testing procedures for HIV and other sexually transmitted infections. Briefly, participants’ HIV status were determined using two rapid tests on blood sourced from either finger-prick/venepuncture. All trials used similar eligibility criteria: sexually active, not pregnant and not planning to get pregnant during the study follow-up, residing around the site area past 12 months, HIV negative at screening/enrolment. All participants had consented to enrol in the trials and received counselling on risk reduction and HIV prevention as well as free/unlimited access to male condoms. Urine pregnancy tests were conducted at the screening and women enrolled in the trials if they were not pregnant and were not intending to become pregnant during the study. All the protocols and informed consent forms were approved by each study’s ethical committees separately.

Measurements

The primary outcome of this study was pregnancy incidence. Variety of demographic, socioeconomic characteristics and sexual behaviours were also included in the analysis: age (<20, 20–24, 25–29, 30–34, 35+years); marital-status (single/not-cohabiting vs. married/cohabiting); total number of sex partners in past three months (<2 vs 2+ partners); age at sexual debut (<15, 15–19 vs.20+ years), condom used at last sex (yes/no); method of contraceptive: none, condoms, oral/pill vs. injectables; parity (nulliparity, primiparity, 2 births vs. 3+ births) and average number of sexual acts in the past two weeks were analysed. Study participants were also tested for other sexually transmitted infections (STIs).

Deriving pregnancy risk assessment:

We used “split-sample” method in order to develop the risk scoring algorithm to predict subsequent pregnancies among the study population: 67% of the study population was randomly assigned to the development dataset using the random number generation function in Stata 14.0; the rest (i.e. 37%) were allocated to the validation dataset. All available risk factors were considered in Cox proportional models. Backward-selection technique was used to finalize the “development model”. Only statistically significant factors with p<0.05 were included in the final model. The goodness of fit of the model was assessed using the Hosmer–Lemeshow goodness-of-fit test. We have created our weighted scores for each risk factor by using the coefficients (i.e. logarithms of the hazard function) in the final multivariable model. After rounding up each coefficient to the nearest integer, we multiplied them by 10. We calculated pregnancy risk score for each woman by adding up final rounded integers. These scores were used to categorise women into: “low”, “mild”, “moderate”, “high” and “severe” risk of pregnancy. Statistical robustness and accuracy of these classification were evaluated in development and validation cohorts.

In internal validation cohort, we calculated the area under the receiver-operating curve (AUC)] in order to assess the model accuracy of the scoring algorithm. We added weighted scores for each-item listed in Tables 1 and 2. We split subject-specific scores into quintiles [1st to 5th]. Cox regression models were fitted to measure the increasing trend in hazard ratios; crude incidence rates were calculated across the deciles of the score.

Table 1:

Developing the risk scoring algorithm for Pregnancy:

	Development Dataset (n= 67%)					Validation Dataset (N= 33%)
Characteristics	%0	Adjusted HR	p-value	βX10	Score	%00	Adjusted HR	p-value	βX10	Score
Age groups
<20 years	10%	4.00 (2.97, 5.30)	<0.001	13.8	14	10%	5.40 (2.90, 10.1)	<0.001	16.9	17
20–24 years	35%	3.16 (2.46, 4.07)	<0.001	11.5	12	36%	5.00 (2.88, 8.73)	<0.001	16.0	16
25–29 years	22%	2.27 (1.72, 2.98)	<0.001	8.2	8	20%	4.34 (2.42, 7.76)	<0.001	14.6	15
30–34 years	13%	1.61 (1.17, 2.21)	0.003	4.8	5	14%	2.70 (1.41, 5.14)	<0.001	10.0	10
35+ years	20%	1		-	0	20%	1	-	-	0
Age of sexual debut
20+ years1	17%	1				17%	1		-	-
15–19 years	78%	1.06 (0.80, 1.40)	0.697	-	-	78%	1.77 (1.00, 3.15)	0.053	0	-
<15 years	5%	1.28 (0.81, 2.03)	0.294	-	-	5%	1.37 (0.52, 3.61)	0.537	0	-
Married/cohabitating										-
Yes	77%	1		-	0	77%	1			0
No	23%	1.28 (1.06, 1.55)	0.011	2.5	3	23%	1.57 (1.10, 2.26)	0.015	4.5	5
Sex partners (past 3 months)
<2	87%	1				86%	1
2 +	13%	1.37 (1.03, 1.83)	0.033	3.2	3	14%	2.02 (1.28, 3.16)	0.002	7.0	7
Parity
0	11%	5.66 (4.30, 7.44)	<0.001	17.3	17	12%	4.31 (2.53, 7.34)	<0.001	14.6	15
1	44%	2.10 (1.63, 2.71)	<0.001	7.4	7	46%	2.26 (1.40, 3.63)	0.001	8.1	8
2	24%	1.64 (1.24, 2.18)	<0.001	5.0	5	22%	1.74 (1.02, 2.98)	0.042	5.6	6
3+	22%	1				21%	1	-	0	0
Number of sex acts in past 2 weeks
0–1	38%	1		-	-	40%	1
2–3	39%	1.04 (0.66, 1.62)	0.866	-	-	42%	1.29 (0.60, 2.76)	0.516	-	-
4+	23%	1.18 (0.72, 1.92)	0.505	-	-	18%	0.71 (0.23, 2.21)	0.556	-	-
Contraceptive use
Injectables	53%	1		-	0	53%	1		-
No contraceptives	14%	4.58 (3.74, 5.60)	<0.001	15.2	15	13%	6.15 (4.20, 9.01)	<0.001	18.1	18
Oral contraceptives	10%	6.04 (4.94, 7.40)	<0.001	18.0	18	15%	9.32 (6.47, 13.41)	<0.001	22.3	22
Condom	15%	6.16 (5.10, 7.44)	<0.001	18.2	18	11%	7.70 (5.37, 11.01)	<0.001	20.3	20
IUD	8%	0.25 (0.21, 0.31)	<0.001	<0	0	8%	-	-		0
Condom used last sex
No	33%	1				33%	1
Yes	67%	0.99 (0.85, 1.15)	0.878	0	0	67%	0.86 (0.65, 1.14)	0.298	0	0
Diagnosed with STI1
No	82%	1				80%	1
Yes	18%	1.13 (0.94, 1.37)	0.185	0	0	20%	1.10 (0.80, 1.48)	0.607	0	0

⁰in development dataset;

⁰⁰in validation dataset;

¹diagnosed with at least one STI (chlamydia, gonorrhoea or syphilis)

Table 2:

Hazard ratios for pregnancy outcome during the study follow-up across the risk score categories in the development dataset:

	Development dataset			Validation dataset
	Pregnancy Incidence rate (95% CI)	Hazard Ratio (95% CI)	p-value	Pregnancy Incidence rate (95% CI)	Hazard ratio (95% CI)	p-value
Risk score categories†
Low risk	1.8 (1.2, 2.6)	1		1.8 (1.2, 2.6)	1
Mild risk	3.2 (2.3, 4.3)	1.77 (1.10, 2.85)	0.019	4.0 (3.0, 5.0)	2.19 (1.39, 3.46)	0.001
Moderate risk	3.7 (2.8, 4.9)	2.08 (1.32, 3.28)	0.002	3.7 (2.7, 5.0)	2.10 (1.29, 3.37)	0.003
High risk	10.1 (8.4, 12.2)	5.70 (3.78, 8.58)	<0.001	10.1 (8.4, 12.2)	5.73 (3.76, 8.73)	<0.001
Severe risk	25.0 (22, 28)	14.00 (9.54, 20.56)	<0.001	22.3 (20.0, 25.2)	12.56 (8.44, 18.68)	<0.001

^†1^st quintile (score 0–15); 2^nd quintile (16–20); 3^rd quintile (21–24); 4^th quintile: (25–35); 5^th quintile: (36–55)

Results

A total of 1093(11%) pregnancies occurred among 9,948 women; with an overall incidence rate of 9 per100 person-year (95% CI:8.5,9.6); median age of the study population was 25 years old (Interquartile-range (IQR):22–32). At baseline, 55% of them had either no children or one child; 18% of them were diagnosed with at least one STI(s).

Development and validation of risk assessment tool

Pregnancy incidence rate were 8.9 per 100 person-year and 9.3 per 100 person-year in the development and validation datasets respectively. In multivariable analysis, we identified five risk factors as the significant predictors of increased risk of pregnancy: (1) younger age (<35 years); (2) single and/or not cohabitating; (3)≥2 sexual partners in past 3 months;(4) parity<3; (5)not using an long acting contraceptives such as injectables. The same factors were also significantly associated with increased risk of pregnancy when we conducted the analysis using the validation dataset. The adjusted HR’s in the validation model were also broadly similar to the ones in the development model. The p-values for the Hosmer-Lemeshow test were 0.561/0.432 for the development/validation models; indicating a reasonable fit for our primary outcome of interest and the risk factors considered in the multivariable models.

We calculated subject-specific risk scores by summing up the scores assigned for each risk factor. There was a significant increasing trend in the HRs for pregnancy incidence across the quintiles of the subject-specific scores in both data sets (P_trend<0.001). Overall, discriminative power of the risk score was 76% (95%CI:73%,78%) and 74% (95%CI:71%,76%) for the development and validation models respectively (Table 3). The cut-point of ≥21 (moderate-to-severe) had a sensitivity of 95% with a specificity of 21%; while the score of ≥25 (high-to-severe) had 89% sensitivity and an increased specificity of 36%.

Table 3:

Performance of risk scoring algorithm for different cut points:

	Development Data (AUC‡=76 %, 95%CI: 73%, 78%)		Validation Data (AUC‡= 74%, 71%, 76%)
	Sensitivity	Specificity	Sensitivity	Specificity
Risk score†
≥16	98%	10%	99%	5%
≥21: moderate to severe risk	95%	21%	98%	9%
≥25: high to severe risk	89%	36%	96%	15%
≥36: severe risk	86%	41%	89%	32%
≥40	74%	45%	86%	37%

^‡Area under the curve

Figure 1a presents the age-specific median pregnancy risk score for the risk factors included in our algorithm in the overall study population. The median score was the highest among younger women and ranged from 30 to 50 regardless of the other factors included in the algorithm; while older women had the highest median score if they had less than three children. We also presented probabilities of pregnancies at month 6,12,18 and 14 by the categories of subject-specific risk scores (Figure 1b). Women with moderate-to-severe risk scores (i.e. score≥25) were three-to-four times likely to become pregnant within 24 months compared to those who had low or mild risk scores. Final risk scoring tool is presented in Supplementary table S2.

Figure 1:

Total median score by risk factors and age groups

Age-specific probabilities of pregnancy by risk factors

Age-specific probabilities of pregnancy were also tabulated for each risk factor on the Supplementary Table S1. Overall, the highest probabilities of pregnancy were consistently observed among the younger women regardless of the presence of the other risk factors. For example, the probability of pregnancy for a woman who was not using injectable contraceptive and younger than 20 years of age ranged between 44%−48% within 24 months; while this probability declined to 15% for those who were using injectables in the same age group; while probability of pregnancy for a woman with no children and younger than 20 years of age was as high as 58% within 24 months and declined to 7% for those 35 years or older.

Discussion

High pregnancy rates have been reported to be one of the major barriers for the effectiveness and efficaciousness of the biomedical interventions being tested in HIV prevention trials [16–18]. We developed and validated a 5-item risk assessment tool using combined data from approximately 10,000 women who enrolled in one of the six HIV biomedical intervention trials between 2002–2017. Younger age, not using injectable contraceptives and being nulliparity/primiparity collectively had the highest impact on pregnancy incidence. In addition, single women and having a higher number of sexual partners in the past three months were also identified as significant predictors of pregnancy and was therefore included in our risk scoring algorithm. Although, these results are generally consistent with the previous research, a direct comparison was not possible because of the differences in methodologies and populations across the studies [16–18].

One of the most striking results from our study was the increased risk associated with condom use. Women who reported using condoms as their preferred contraceptive method were 6 times more likely to get pregnant during the study compared to those who were using injectables. In fact, the predicted probability of pregnancy incidence within 24 months was approximately 50% among younger women (<25 years) who reported condom use at screening. Since pregnancy is generally associated with condomless sex, these findings provide strong evidence for high rates of unprotected sex in this subgroup of women. These results are also consistent with the other sexual behaviours that we considered in this study; 67% of the women reported that they had used a condom at the last sexual act. However, there was an association with pregnancy incidence. Therefore, it is likely that self-reported condom use might be subject to misreporting. Similar results were reported previously in the studies conducted among African women [14]. Collecting sexually sensitive information has been proven to be challenging in HIV prevention trials conducted in the African population [20]. To avoid additional counselling and/or for social acceptability, study participants are likely to overreport their condom use behaviours [12,20]. Results from our study also confirmed that women who were on oral contraceptives were also six times more likely to get pregnant compared to those who were using injectables. This can partially be explained by the relatively high proportion of women (24%) who switched from oral contraceptives to non-effective methods (i.e. no contraceptives or condom) (data not shown). Although, it cannot be ascertained using the data at hand, these results can also be attributed to a high level of discontinuation and/or incorrect contraceptive use which were also reported previously in African populations [12,17,20,21] These findings are counterintuitive and conflicted with the long-term effectiveness of the oral contraceptives [12]; while consistent with the results reported in Centre for the AIDS Programme of Research in South Africa (CAPRISA)-004 trial [17]. Briefly, CAPRISA-004 was the first trial to implement a comprehensive contraceptive curriculum to reduce pregnancy rates and enhance contraceptive uptake. This study reported a 10-fold increase in pregnancy incidence rates among oral contraceptives users compared to those who were on injectables.

In our algorithm, parity was the second most influential characteristic to predict pregnancy incidence. Particularly, nulliparous women were more than five times likely to get pregnant during the study compared to the women who already had three or more children. Average predicted probability of pregnancy for a nulliparous woman younger than 20 years of age was estimated as 58% within 24 months. Given the high fertility rates as well as cultural and social pressure to have more children in the region, these results are not surprising [17,20].

Identifying the most influential factors associated with pregnancy incidence is crucial to develop more effective strategies for future trials. Our risk scoring algorithm may play a significant role in the clinical research setting by identifying those who are at varying risks of pregnancy. This information may be used to guide recruitment and screening strategies of future biomedical intervention trials. Women at moderate to severe risk of pregnancy can be prioritised for additional close monitoring, more intensive and tailored counselling sessions, frequent pregnancy tests and multicomponent pregnancy prevention packages. Although, these extra efforts may potentially increase risk of bias in the study population, their overall benefit would be higher for the success and integrity of the future biomedical intervention trials.

Four out of five risk factors in our pregnancy scoring algorithm, namely, younger age, single/not cohabiting, low number of children and higher number of sexual partners, were also identified as significant predictors of HIV infections among women who enrolled in HIV prevention trials [19]. Collectively, these overlapping risk factors indicate the seriousness of the problem and the potential future burden in HIV prevention trials particularly for pregnant HIV infected women and their unborn babies.

Our study has several limitations. Firstly, we used data from the women who consented to enrol in HIV prevention trials with certain eligibility criteria, including being at reproductive age and sexually active. Study population may be at higher risk of pregnancy compared to the general population. Contraceptive methods were self-reported; therefore, they would be subject recall bias. There were no data available from male partners of the women who participated in these trials.

Conclusion

The pregnancy risk algorithm developed in this study will potentially provide a simple guide, with acceptable accuracy, to identify and alert clinical research staff to study participants who need additional contraceptive counselling for pregnancy.

Figure 2:

Cumulative predicted probabilities of pregnancy incidence by the quintiles of score and study follow-up

Highlights:

• Identifying the most influential factors associated with pregnancy incidence is crucial in HIV prevention trials.

• We developed an algorithm to predict women at highest risk of pregnancy during the study follow-ups.

• Four out of five risk factors in our pregnancy scoring algorithm, namely, younger age, single/not cohabiting, low number of children and higher number of sexual partners, were also identified as significant predictors of HIV infections among women who enrolled in HIV prevention trials.

• Collectively, these overlapping risk factors indicate the seriousness of the problem and the potential future burden in HIV prevention trials particularly for pregnant HIV infected women and their unborn babies.

• This is the first and the largest study to report age-specific probabilities of becoming pregnant during a biomedical intervention trial among South African women.