Weighted Cumulative Exposure Modelling to Assess the Association Between Reproductive Factors and Future Cardiovascular Disease in Women.

N.D. and M.A. drafted the brief report. M.‐E.B., M.A. and G.D.S. critically revised and reviewed the report.

We propose using flexible WCE modelling to quantify risks of pregnancy events such as SMM associated with future CVD. This strategy permits the assessment of cumulative effects of SMMs as a time‐varying exposure while accounting for their timing/recency and severity across multiple pregnancies [ 11 ]. The WCE methodology has been validated in extensive simulations [ 10 , 11 ] and has largely focused on medication safety or effectiveness [ 16 ]. Indeed, an independent review recognised WCE among the most useful methods for modelling time‐varying drug exposures [ 17 ], where it usually improves the fit to data [ 16 ]. More recent studies employed WCE models to assess cumulative effects of non‐medication exposures, such as radiation, physical activity or air pollution [ 11 ]. Studies have shown that SMM is associated with a reduction in the probability of subsequent birth, likely due to a combination of subfertility and family planning decisions [ 18 , 19 ]. However, our proposed analyses would not be confounded by these patterns because the WCE models include both women who do and do not become pregnant following SMM and non‐SMM pregnancy. SMM complicates ~1.5%–2.0% of pregnancies in Canada, and its incidence is rising [ 20 ]. In offering a detailed picture of the relationship between SMM as a time‐varying exposure and women's future CVD, the use of the flexible modelling approach proposed herein represents an advance over traditional statistical approaches. The WCE approach can be used to not only refine estimates of etiologic associations but to inform novel clinical prediction models with the potential to predict postpartum CVD for a given woman based on her unique pregnancy history. Translated into a risk calculator based on a reproductive risk score and geared for clinical use at routine postpartum visits, such prediction models can contribute to an integrated personalised strategy to improve postpartum cardiovascular health.

The WCE model relies on a weight function that assigns differential weights to past exposures, depending on how long ago they occurred. The exposures of interest in the current study are reproductive events present at a given point in time and aggregated through a reproductive ‘risk score’. The shape of the weight function is estimated using flexible cubic spline modelling to optimise model fit to the empirical data [ 10 ]. Cumulative effects of past exposures are then quantified by the weighted sum of their values. Formally, this approach involves modelling, at any time u during follow‐up, a time‐varying exposure metric WCE ( u ) = ∑[ w ( u  −  t )· X ( t ), for t  ≤  u ] where X ( t ) is the exposure observed at an earlier time t , ( u  −  t ) is the time elapsed since the exposure, and w ( u  −  t ) is the corresponding value of the estimated weight function [ 10 ]. In WCE analyses of longitudinal cohorts, the time‐varying exposure metric WCE ( u ) is continuously updated with increasing follow‐up time u to account for both increasing time(s) since past exposure(s) and potential new exposure(s). The resulting WCE ( u ) exposure metric can then be included in a multivariable regression model for time‐to‐event analysis, such as a Cox proportional hazards model [ 10 ]. To illustrate the insights that may be offered by WCE analyses applied to reproductive events over a woman's life course, we consider an earlier study of the association of low‐dose ionising radiation and cancer incidence in a cohort of adults [ 12 ]. We selected this example because radiation and SMM/reproductive exposures both (i) affect only a relatively small subsample of the respective study population, and (ii) for those ever exposed, exposures are sparse in time. The WCE‐based weight function estimate indicated that radiation exposures that occurred 2–5 years previously had the strongest association (as reflected by the highest estimated weights) with the current hazard of incident cancer, whereas radiation received more than 6 years ago had no material impact (with weights close to 0). Based on the estimated weight function, calculated adjusted hazard ratios (HRs) assessed cumulative effects of different specific patterns of past radiation exposure, accounting for both frequency and recency [ 12 ]. For example, a man who had received low‐dose ionising radiation at 2 times, 4 and 5 years previously at a 15.0 mSv dose, was estimated to have 62% higher hazard of having an incident cancer than another man who was never exposed but had the same values for all covariates. We propose to apply WCE modelling in a longitudinal cohort of ~2 million pregnant women followed for up to 13 years [ 13 ] to estimate associations between an aggregate reproductive risk score across multiple pregnancies and future premature CVD. The WCE implementation would follow three steps. Step 1 will focus on estimating a multivariable reproductive risk score to quantify the overall severity of SMM subtypes identified in an affected woman during a given pregnancy while accounting for other relevant reproductive (e.g., parity, stillbirth, endometriosis, infertility treatment) and non‐reproductive covariates (e.g., diabetes mellitus, hypertension, age, socioeconomic status). To this end, a multivariable Cox proportional hazards model will first be developed to estimate the adjusted associations of individual SMM subtypes observed during a given pregnancy with incident CVD. The Cox model will estimate adjusted β coefficients (log HRs) for each SMM subtype, as well as for other reproductive and non‐reproductive predictors. Then, for each pregnancy, the aggregate reproductive risk score will be calculated as the weighted sum of the SMM subtypes observed during this pregnancy, with weights defined at the birth level by the β coefficients. The resulting scores will then be used to define the time‐varying exposure X ( t ) in the WCE analyses in Step 2. Time zero will be set to 42 days after delivery to allow for the inclusion of pregnancy‐related predictors that occur in the early postpartum period while avoiding immortal time bias. Women will be followed until the CVD event or censored at the earliest of death, end of the observation period or reaching 60 years of age (i.e., marking usual onset of menopause). In Step 2 (WCE modelling) [ 10 , 11 ], a longitudinal dataset will be constructed in which SMM and other reproductive and non‐reproductive predictors will be recorded for all women in regular 3‐month intervals, from 42 days after the woman's first delivery until the outcome or a censoring event. We will begin follow‐up at 42 days after delivery as the majority of postpartum delivery‐related complications occur within this timeframe [ 14 , 15 ]. For each pregnancy, data on SMM subtypes will be aggregated into a corresponding value of the multivariable ‘SMM risk score’ developed in Step 1, which will be assigned to the end of the corresponding 3‐month period. Accordingly, risk scores will be represented by a time‐varying exposure, with different values assigned for each pregnancy for an individual with > 1 pregnancy, and fixed to 0 when not pregnant. Finally, a new multivariable Cox model will estimate the WCE‐based weight function for the reproductive risk scores, adjusted for other reproductive risk factors and additional patient characteristics. Step 3 will estimate adjusted WCE‐based HRs associated with different SMM exposure histories, using validated methods [ 10 ] and illustrated in the radiation example outlined above [ 12 ].

In a cohort of 1,992,972 births, mean (SD) follow‐up time was 7.3 ± 3.8 years, with 16,846 cardiovascular hospitalisations observed (1.19 per 1000 person‐years). The rate of SMM was 16.4 per 1000 births overall and was 7.8, 2.8 and 0.7 per 1000 births for haemorrhagic SMM, hypertensive SMM and non‐hypertensive cardiac SMM respectively. Risk of CVD was 2.64, 1.95, 4.55 and 9.11 per 1000 person‐years among those with any, haemorrhagic, hypertensive and non‐hypertensive cardiac SMM respectively.

During female reproduction, several sex‐specific risk factors for chronic disease may arise. Specifically, pregnancy is a physiologically demanding state that serves as a window into a woman's future health [ 1 ]. Severe maternal morbidity (SMM)—sudden and life‐threatening events occurring during pregnancy, childbirth or shortly thereafter—often resolves with prompt delivery and good obstetrical care, yet may reveal a predisposition to chronic disease and premature mortality [ 2 ]. Specific SMM subtypes including preeclampsia, obstetric haemorrhage, pregnancy‐associated stroke and renal impairment have been linked with premature cardiovascular disease (CVD; occurring before age 60 years or onset of menopause) as early as 1 year postpartum and up to 15 years after delivery [ 2 , 3 , 4 , 5 , 6 ]. Preeclampsia and other SMM subtypes like placental abruption and obstetric haemorrhage are transient multi‐system disorders affecting the vascular endothelium [ 7 ], resulting in long‐term subclinical vascular impairment [ 8 ], and later, CVD. Other reproductive conditions such as infertility, endometriosis and polycystic ovarian syndrome are also associated with poor metabolic health and premature CVD [ 9 ]. Despite these advances in our understanding of CVD in women, existing research is limited by methodological challenges. Previous studies assessed the presence of reproductive events as a dichotomous time‐invariant condition, failing to assess either (i) their cumulative effects observed across multiple pregnancies or (ii) the time elapsed since each affected pregnancy. We propose to use a weighted cumulative exposure (WCE) modelling approach for future studies assessing or predicting CVD as a function of reproductive events [ 10 , 11 ]. The proposed approach outlined below will capture a woman's total reproductive trajectory over time in addition to time‐varying non‐reproductive risk factors in a model to predict CVD.

The authors declare no conflicts of interest.