Prediction of pre-eclampsia at St. Mary's hospital lacor, a low-resource setting in northern Uganda, a prospective cohort study

Pre-eclampsia (PE), a pregnancy syndrome, is characterised by hypertension and proteinuria [1, 2]. Approximately 90 per cent of cases present in the late preterm (≥ 34 weeks) period and have good maternal and fetal outcomes [2, 3]. However, 10 per cent of cases who have an early presentation (< 34 weeks) have more severe disease and carry the additional high risks associated with preterm birth [3, 4]. In addition, mothers with a history of pre-eclampsia are at increased risk for developing the cardiovascular and renal disease [2, 3].

Risk factors of PE include low socio-economic status, nulliparity, multiple pregnancies, obesity, chronic hypertension, being a woman of African descent, previous maternal or family history of pre-eclampsia, and maternal age ≥ 35 years [5‐7]. In addition, high second-trimester artery Doppler resistive index, pulsatility index, and end-diastolic notch are known risk factors for pre-eclampsia [8, 9]. Early diagnosis and delivery of the fetus is the only known treatment, thus necessitating the need for prediction models of this disorder.

De Kat et al. [10] summarised risk factors and models for predicting pre-eclampsia. Black race stood out as a significant risk factor in all the studies where the communities had a mixed race, insinuating that the predominantly black Ugandan communities are at high risk for pre-eclampsia. Studies by Al-Rubaie et al. [11] achieved the highest area under the curve (AUC) for predicting pre-eclampsia at 76% using maternal history. Gallo et al. [12] screened for PE using maternal history and mean arterial pressure (MAP) at a false-positive rate of 10%; their detection rate of total pre-eclampsia was 49.3%. Jhee et al. [13] used laboratory tests (serum urea, aspartate aminotransferase (AST), alanine transaminase (ALT), creatinine, and haemoglobin levels) to predict pre-eclampsia. They got an area under the curve (AUC) above 57%. Delic et al. added uric acid, urea thrombocytes, hematocrit, AST, and leukocytes to the regression model and correctly classified 83.8% of patients with pre-eclampsia [14]. Yucel et al. [15] predicted pre-eclampsia using mean platelet volume (MPV) with an AUC of 64.1% and plateletcrit (PCT) with an AUC of 71.2%.

Antwi et al. [16] reviewed prediction models for pre-eclampsia between 2000 and 2019 and found diverse prediction accuracy ranging from 45 – 95% in the different regions of the world. After observing the wide variation in prediction rates of pre-eclampsia and consistently having the black race as a risk factor for pre-eclampsia, we developed and validated risk prediction models based on maternal characteristics from northern Uganda.

The research was a prospective cohort study at St. Mary's Hospital Lacor. This hospital is a private, not-for-profit hospital founded by the Catholic Church. It is located six kilometres west of Gulu city along Juba Road in Gulu district (Longitude 30 – 32 degrees East and Latitude 02 – 04 degrees North). St. Mary's Hospital Lacor is one of the teaching hospitals of Gulu University with a bed capacity of 482. It is staffed by specialists, medical officers, midwives, nurses, laboratory and radiology staff, and support and administrative staff. The hospital receives about three thousand six hundred antenatal mothers and conducts about six thousand deliveries annually [17]. Some mothers go to the hospital for delivery without prenatal care; others are referred from smaller health units. The mothers pay five thousand Uganda shillings (Ugx 5,000/ =) ($1.5) as the cost per visit. This cost is often waived for most mothers who cannot afford it. Normal labour and delivery cost fifteen thousand (Ugx 15,000/ =) (about $4.50), and Caesarean section costs twenty-five thousand (Ugx 25,000/ =) (about $7.5) Uganda shillings.

Using Yamane's 1967 formula [18] for calculating sample size for cohort studies using finite population size: St. Mary's hospital Lacor receives approximately three thousand six hundred antenatal mothers annually. Since my study duration was 24 months, the limited population we could access was about 7,200 mothers.

We doubled the sample size to take care of loss to follow-up. We targeted all pregnant women attending antenatal care at St. Mary's Hospital Lacor. In Uganda, the clinical guideline advocates for the first antenatal care to be sought by a pregnant mother up to 20 weeks of gestation [19]. While all expectant mothers attending antenatal care at St. Mary's hospital Lacor were eligible, we included gestational ages of 16 to 24 weeks and those who gave written informed consent to participate in the study. Those whose pregnancies were less than 16 weeks were given a return date for the recruitment, while those above 24 weeks or had molar pregnancies, intrauterine fetal death and anencephaly were excluded.

We used consecutive sampling. We informed the mothers about the study during their morning health education meeting given to all mothers on arrival for prenatal care at the hospital. All the women who satisfied the inclusion criteria were approached and requested to provide informed consent. A research assistant administered questionnaire to capture their history and performed a physical examination. Some mothers were asked to give blood samples for full haemogram and liver and renal function tests. A few mothers (after the 1000^th mother) did not undergo laboratory tests for logistical reasons. An obstetrician performed the uterine artery Doppler sonography.

We recruited 1,285 pregnant mothers at 16–24 weeks from April 2019 to March 2020. All the mothers were of African ancestry at the end of the recruitment period. We followed up with the participants until September 2020. One thousand four (1,004) complete delivery records were obtained at the end of the study period. Seven hundred eighty-two (782) participants had laboratory blood tests (full haemogram, liver and renal function tests) done in addition to blood pressure readings, body mass index calculation and maternal history. Details are in Fig. 1.

The outcome was a combination of a blood pressure ≥ 140/90 mmHg and urine protein ≥  + 1 (pre-eclampsia) by delivery time. The data was pre-processed using Stata® version 15 and built models using RStudio version 4.1.3. Model 1 was built from second-trimester maternal history and physical examination findings, model 2 from the ultrasound and uterine artery Doppler indices, model 3 from a combination of maternal history, physical examination and ultrasound findings, model 4 from maternal laboratory tests, model 5 from a combination of laboratory tests with maternal history, and model 6 from the combination of all models.

We included all variables, did a univariate analysis and got unadjusted p-values for every variable collected. Afterwards, we had all variables with p-values ≤ 0.20 or known risk factors for pre-eclampsia in a logistic regression model and removed the non-statistically significant predictors step-wise. Finally, we retained the independent risk factors for pre-eclampsia and used them to build the models of choice, one with the least number of predictors with a higher AUC.

Because the incidence of pre-eclampsia was low (4.3%) [20], we generated synthetic balanced data using the ROSE package [21, 22] in RStudio by over-sampling pre-eclampsia and under-sampling non-preeclampsia. We got 383 (48.8%) for those diagnosed with pre-eclampsia and 399 (51.2%) for non-pre-eclampsia. Then, we evaluated the actual model performance against the ROSE-derived synthetic dataset using K-fold cross-validation in RStudio to obtain the AUC's accuracy, sensitivity, specificity and McFadden's pseudo R². The variables were said to be independent risk factors for pre-eclampsia if their p-value < 0.05 in the model. The models, too, had a good fit if McFadden's value was between 0.2 – 0.4.

One thousand four (1,004) participants were used in a regression model to obtain the unadjusted relationship between the maternal history, clinical characteristics and pre-eclampsia. Seven hundred eighty-two (782) participants had additional laboratory blood tests (full haemogram, liver and renal function tests) and complete delivery records. So they were used to build the prediction models.

In this research, maternal history predictors of pre-eclampsia were maternal age, parity, personal history of pre-eclampsia, body mass index, diastolic pressure, and multiple pregnancies. They predicted pre-eclampsia with 66.6% accuracy, 82.7% sensitivity, and 78.4% AUC with a McFadden's pseudo R² of 0.21. It can identify four out of five participants destined to develop pre-eclampsia. The low specificity of the model of close to 50% reduces the model's accuracy to 66%. This model is of good fit and can be used independently in prenatal clinics to screen for pre-eclampsia. In Ghana [23], predictors of pre-eclampsia were diastolic blood pressure, family history of hypertension in parents, history of pre-eclampsia in a previous pregnancy, nulliparity and obesity, with an AUC of the original model being 70% and 68% in the validation cohort. Gallo et al. [12] screened by maternal characteristics and mean arterial pressure (MAP) at a false-positive rate of 10%; their detection rate of total pre-eclampsia was 49.3%. In a systematic review by Al-Rubaie et al. [11], their detection rate was 76%.

With uterine artery Doppler indices, we predicted pre-eclampsia with over 68% accuracy and 71.4% AUC. Unfortunately, the model had a McFadden's pseudo R² of 0.09 and was not a good fit. Therefore, this model cannot be used independently in prenatal clinics to screen for pre-eclampsia. That was way below Trudinger et al. [9], who predicted up to 90% of pre-eclampsia in Australia using an end-diastolic notch. Using a combination of maternal history, physical examination and uterine artery Doppler indices, we got an AUC of 80.4% with 76.0% accuracy. That is comparable to Pedroso et al. [24], who found a combination of uterine artery Doppler indices and maternal history predicted 75% of PE.

With laboratory blood tests, we predicted pre-eclampsia with 67.1% accuracy and 75.6% AUC with McFadden's pseudo R² of 0.11. However, this model cannot be used independently in prenatal clinics to screen for pre-eclampsia. Jhee et al. [13] used a combination of serum urea, aspartate aminotransferase (AST), ALT, creatinine, and haemoglobin levels to predict pre-eclampsia with AUC above 57%. Yucel et al. [15] predicted pre-eclampsia using mean platelet volume (MPV) and plateletcrit (PCT) with AUC of 64.1% and 71.2%, respectively.

When we combined laboratory blood tests with maternal history, the accuracy improved to 72.7% with an AUC of 82.2% with a McFadden's pseudo R² of 0.26. The combination of maternal history and uterine artery Doppler indices improved the accuracy to 76% with 80.4% AUC with a McFadden's pseudo R² of 0.25. Combining maternal history, blood tests, and uterine artery Doppler indices (model 6) slightly improved the prediction accuracy to 77.0% and 80.2% sensitivity with an AUC of 84.9% with a McFadden's pseudo R² of 0.30. All the combined models were of good fit and could be used independently in prenatal clinics to screen for pre-eclampsia.

Delic et al. [14] added uric acid, urea thrombocytes, hematocrit, AST and leukocytes into the logistic regression model and correctly classified 83.8% of patients with pre-eclampsia. That had a better detection rate than 57% in the UK [25]. A low level of serum ALP may signify a reduced viable mass of placental tissue in pregnancy [26, 27], which means a decreased surface area for the transfer of nutrients from mother to baby. This reduced surface area of the functional placenta may increase the number of placental infarcts and, eventually, placental debris released into the maternal circulation. In addition, increased levels of placental tissue in maternal circulation lead to maternal systemic inflammation [28]. That may result in endothelial injury, vasoconstriction and hypertension [29].

Duckit et al. [6], in a systematic review, found controlled cohort studies showing the risk of pre-eclampsia increased in women with a previous history of pre-eclampsia, multiple (twin) pregnancy, nulliparity, family history, raised blood pressure (diastolic ≥ 80 mm Hg) at booking, increased body mass index before pregnancy at booking, or maternal age ≥ 40. Antwi et al. [16] reviewed prediction models for pre-eclampsia between 2000 and 2019 and found diverse prediction accuracy ranging from 45 – 95% in the different regions of the world. The other prediction rates could explain the differences in the populations studied or the test techniques and the ultrasound machines used. Our models seem within acceptable accuracy, although the whole study population was at high risk. These models will ease the identification of high-risk mothers and referral to specialists' healthcare providers. That may, in turn, contribute to reducing maternal mortality and morbidity in the community.

Predictors of pre-eclampsia in the low resource setting of northern Uganda are maternal age ≥ 35 years, nulliparity, maternal history of pre-eclampsia, body mass index, diastolic pressure, white blood cell count, lymphocyte count, serum alkaline phosphatase and end-diastolic notch of the uterine arteries. Prenatal clinics without any ultrasound scans or laboratory can adequately predict pre-eclampsia with up to 66.6% accuracy and 78.4% AUC. However, clinics with Doppler ultrasound and laboratory tests can use maternal history either with ultrasound or laboratory blood tests or both. That will improve the prediction AUC to over 80%.