Canadian Researchers Design Population Tool to Predict Adverse Pregnancy Outcome Risks

Researchers in Canada have established a formal methodological protocol for the Adverse Pregnancy Outcomes Population Risk Tool, designated as PregPoRT, a predictive algorithm designed to quantify population-level risks of major maternal complications using linked survey, administrative, and environmental records. Adverse pregnancy outcomes, including gestational diabetes mellitus, preeclampsia, and placental abruption, represent substantial contributors to maternal and fetal morbidity and mortality with enduring consequences for individual health trajectories and healthcare expenditures. In high-income nations, up to one-third of pregnancies are affected by at least one adverse pregnancy outcome, and in Canada preeclampsia alone contributes over three million Canadian dollars annually to direct medical costs. While existing clinical risk calculators rely almost exclusively on individual biomarker screenings and hospital clinical metrics, they routinely omit distal social, behavioral, and residential environmental conditions. The PregPoRT study protocol outlines the design of a population-based predictive instrument that directly integrates these broader structural influences to provide healthcare planners and public health units with actionable epidemiological risk forecasts.

The conceptual framework underpinning the identification and structuring of predictive variables is derived from the health equity-oriented maternal health model formulated by Kramer and colleagues in 2019. This framework conceptualizes perinatal outcomes as the product of intersecting forces arranged across two distinct dimensions: a horizontal life course trajectory and a vertical hierarchy of multilevel determinants. The life course axis incorporates pre-pregnancy exposures, cumulative physiological burdens, and psychosocial stressors that shape baseline maternal susceptibility prior to conception. These individual-level factors encompass sociodemographic markers such as chronological age, educational attainment, household income, immigration status, visible minority classification, marital status, and household food insecurity. The framework also encompasses modifiable health behaviors, including cigarette smoking, alcohol consumption, nutritional patterns, and physical activity levels, alongside documented pre-existing chronic conditions like pre-pregnancy diabetes, chronic hypertension, asthma, arthritis, and multimorbidity. Psychosocial factors such as self-rated general health, self-rated mental health, life satisfaction, life stress, and community belonging further complete the preconception exposure profile.

To capture population-wide patterns, the protocol establishes a comprehensive data linkage architecture executed through Statistics Canada’s Social Data Linkage Environment housed at the University of Toronto Research Data Centre. The primary survey foundation is the Canadian Community Health Survey, a nationally representative cross-sectional survey covering all ten provinces and three territories that samples approximately sixty-five thousand respondents annually using a multistage stratified cluster design. Individual survey records are deterministically linked to hospitalization separations cataloged in the Discharge Abstract Database, which is curated by the Canadian Institute for Health Information. Because approximately ninety-eight percent of Canadian deliveries occur within acute care hospital settings, the Discharge Abstract Database provides near-complete national capture of obstetrical hospitalizations across all jurisdictions outside the province of Quebec. Individual pre-pregnancy exposure records from survey responses are tied to delivery events occurring within a subsequent two-year observation window, preserving the temporal sequence between baseline determinants and perinatal diagnostic endpoints.

The vertical axis of the predictive architecture incorporates macro-level environmental and area-based determinants by geocoding participants’ residential postal codes reported during their survey interviews to standardized geographic databases. Area-level deprivation is quantified using the Canadian Marginalization Index, which derives standardized scores from census indicators across four specific operational domains: material deprivation, residential instability, dependency, and ethnic concentration at the dissemination area level. Neighborhood built environment characteristics are incorporated through the Canadian Active Living Environments index, which evaluates land-use patterns, dwelling density, street intersection connectivity, and transit accessibility to measure community supportiveness for active transit. Ambient environmental metrics are drawn from the Canadian Urban Environmental Health Research Consortium, which compiles high-resolution exposure estimates at postal code or one-square-kilometer resolutions. These environmental inputs supply objective measurements of ambient atmospheric pollutants, including fine particulate matter with aerodynamic diameters below two and a half micrometers, nitrogen dioxide concentrations, ambient ozone levels, and satellite-derived vegetation greenness indices.

The protocol defines the target study population as female-identifying individuals aged 15 to 49 who participated in the Canadian Community Health Survey between 2000 and 2017 and experienced a singleton delivery hospitalization within two years. Deliveries are identified through validated diagnostic and procedural codes, while multifetal gestations are systematically excluded to maintain clinical and biological homogeneity across outcome pathways. The primary endpoint is defined as a composite binary outcome encompassing the clinical diagnosis of gestational diabetes, preeclampsia or eclampsia, or placental abruption during the delivery admission. Gestational diabetes is characterized by elevated risks of macrosomia, cesarean birth, neonatal hypoglycemia, and a sevenfold increased maternal risk of progressing to type 2 diabetes. Preeclampsia is strongly linked to maternal organ injury and fetal growth restriction, whereas placental abruption presents critical risks of severe maternal hemorrhage and perinatal mortality. Preliminary cohort assembly identified 13,360 eligible singleton deliveries, demonstrating an overall composite outcome prevalence of 7.6 percent across the linked population.

Sample size determination for the multivariable predictive modeling protocol was formally computed using the pmsampsize package in the R statistical environment, implementing the precision-based criteria developed by Riley and colleagues to minimize parameter estimation optimism and prevent model overfitting. Calculations incorporated the observed 7.6 percent outcome prevalence, a conservative anticipated Nagelkerke pseudo-R-squared value of 0.20 to approximate explained variance, a targeted heuristic shrinkage factor of 0.90, and 118 planned candidate regression coefficients reflecting continuous splines and categorized predictors. Under these stringent statistical constraints, the minimum sample size required for stable prognostic model estimation was calculated at 12,196 individuals, corresponding to at least 927 primary outcome events and an effective events-per-parameter ratio of 7.9. The established retrospective cohort of 13,360 linked deliveries comfortably exceeds this theoretical minimum threshold, ensuring adequate statistical power to support robust parameter estimation without generating unstable coefficient estimates across extensive multi-variable parameterizations.

The mathematical formulation of PregPoRT relies on an accelerated failure time survival model parameterized using a Weibull distribution to evaluate the survival-free time from the survey interview date until delivery outcome. A Weibull distribution was chosen because adverse pregnancy complications manifest across a bounded gestational duration exhibiting a hazard rate that monotonically increases as gestation advances. Continuous variables will be modeled flexibly using restricted cubic splines with piecewise cubic functions smoothed across standard knot placements recommended by Harrell to accommodate non-linear biological associations. Predictor penalization and regularization will be conducted through the least absolute shrinkage and selection operator, using ten-fold cross-validation to select the optimal tuning parameter lambda while controlling multicollinearity among highly correlated social, environmental, and behavioral indices. Maternal mortality will be handled as a competing event through administrative censoring at the time of death, and complex survey weights will be applied across all stages of model fitting to adjust for unequal sampling probabilities and maintain national representativeness.

Missing data will be addressed through multiple imputation by chained equations using the mice library in R, grounded in an empirical evaluation of missing at random assumptions. To account for structural variations in survey design over time, imputation procedures will follow a sequential three-stage architecture. The first stage executes within-cycle imputations to preserve unique survey module properties, the second stage conducts between-cycle imputations for items omitted across specific survey iterations, and the third stage imputes missing area-level environmental and marginalization variables independently. Five distinct imputed datasets will be generated and pooled using Rubin’s rules, while auxiliary predictors, including the primary composite outcome indicator, will inform imputation models to preserve covariance structures without imputing missing outcome labels directly. Sensitivity checks will evaluate model stability against complete-case analyses and expanded twenty-imputation specifications to confirm the robustness of pooled regression parameters and verify that missing data algorithms introduce negligible distortion into baseline risk calculations.

The validation architecture follows a multi-tiered framework designed in accordance with the Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis guidelines. Internal validation will first be executed on a random seventy percent development split of survey cohorts spanning 2000 to 2014, applying five hundred bootstrap resampling iterations to quantify optimism in discrimination and calibration. Split-sample verification will be conducted on the remaining thirty percent split, after which the full historical dataset will be combined to derive final model coefficients. Temporal validation will subsequently evaluate model transportability using independent data from the redesigned 2015/16 and 2017/18 survey cycles to assess performance stability across major survey revisions. Predictive discrimination will be assessed using Harrell’s concordance statistic, overall accuracy will be quantified through Nagelkerke’s R-squared and the Integrated Brier Score, and calibration fidelity will be examined across calibration-in-the-large, calibration slopes, loess calibration curves, and the Integrated Calibration Index across sociodemographic subgroups.

The study protocol explicitly details several methodological boundaries and structural data limitations inherent to the population linkage design. First, cohort inclusion requires explicit participant consent to administrative linkage, introducing potential consent selection bias that cannot be fully corrected through standard survey weighting. Second, the hospital separation registry excludes deliveries taking place outside hospital institutions, such as planned home or birthing center births, although these comprise approximately two percent of Canadian births. Third, because the registry tracks deliveries resulting in live births or stillbirths, spontaneous early pregnancy losses and miscarriages remain unobserved. Fourth, survey historical boundaries span 2000 through 2017, meaning contemporary changes in obstetric management and pandemic-era disruptions are not reflected. Finally, the survey frame excludes institutionalized residents, individuals living on First Nations reserves, and residents of Quebec, limiting generalizability among populations with distinct risk profiles. The investigators intend to publish open-source R code and model coefficients upon completion to facilitate maternal health surveillance and health system equity evaluations.

Subject of Research: Protocol for a population-based predictive risk model for adverse pregnancy outcomes

Article Title: A study protocol for a predictive model to assess population‑based risk of adverse pregnancy outcomes: The Adverse Pregnancy Outcomes Population Risk Tool (PregPoRT)

Article References: Chiodo, S., Grandi, S. M., Gronsbell, J., & Rosella, L. C. (2026). A study protocol for a predictive model to assess population‑based risk of adverse pregnancy outcomes: The Adverse Pregnancy Outcomes Population Risk Tool (PregPoRT). Diagnostic and Prognostic Research, 10(1), Article 5. https://doi.org/10.1186/s41512-026-00220-3

Image Credits: AI Generated

DOI: 10.1186/s41512-026-00220-3

Keywords: adverse pregnancy outcomes, predictive modeling, study protocol, gestational diabetes, preeclampsia, placental abruption, maternal health, social determinants of health, public health surveillance, protocol, predictive, model