After the Attenuated Pregnancy: Emulating a Decade-Scale Postpartum Trial of Gestational RAAS Phenotypes

The renin–angiotensin–aldosterone system (RAAS) is the body's hormone cascade for controlling blood pressure and blood volume: the kidney releases renin, renin produces angiotensin II (which constricts vessels), and angiotensin II tells the adrenal gland to release aldosterone, which makes the kidneys retain salt and water. A healthy pregnancy deliberately drives this cascade far above non-pregnant levels — plasma volume must grow by 40–50% — while vessels become resistant to angiotensin II so blood pressure stays normal. Two prospective cohorts from Kentucky showed that pregnancies destined for hypertension show the opposite pattern: the cascade stalls. Renin-activity surrogate markers show a median change of 0% across gestation instead of the ~31% rise seen in normotensive pregnancy, and aldosterone rises an insignificant ~57% instead of ~135%, leaving third-trimester aldosterone significantly depressed (Activation of the Renin–Angiotensin–Aldosterone System Is Attenuated in Hypertensive Compared with Normotensive Pregnancy). Worse, the danger sign appears early and paradoxically: first-trimester aldosterone above 704 pmol/L, combined with a declining renin trajectory, identified a subgroup with 68.4% chronic hypertension, 31.6% preterm birth, and 15.8% fetal growth restriction (Association of Elevated Serum Aldosterone Concentrations in Pregnancy with Hypertension).

Every one of those measurements stops at delivery. Whether the attenuated RAAS pregnancy leaves a lasting imprint on the mother's cardiovascular system, or simply resolves after the placenta is gone, is unknown — and the authors of both cohorts say so explicitly. This is a textbook case of the "sustainability gap" that plagues clinical research: the same pattern in which cardiac resynchronization studies showed improved surrogate remodeling (LVEF +3.25%, response RR 1.33) but no detectable mortality benefit (RR 0.69, p = 0.24) because follow-up was truncated at 6 months to 2 years (Comparative efficacy of image-guided techniques in cardiac resynchronization therapy: a meta-analysis), and in which a 6-month dyssynchrony study could predict reverse remodeling (OR 5.67) but never observed survival (A prospective evaluation of cardiovascular magnetic resonance measures of dyssynchrony in the prediction of response to cardiac resynchronization therapy). Waiting a decade for a prospective obstetric cohort to mature is not an answer. Retrospective hospital-record cohorts are the only way to get a ten-year outcome horizon now, and the causal machinery for using them well — propensity scoring over 15-year horizons (Long-Term Survival in Men With High Grade Prostate Cancer) and target trial emulation with inverse-probability weighting on linked English records (Weight Change and the Onset of Cardiovascular Diseases: Emulating Trials Using Electronic Health Records) — already exists, but has never been pointed at gestational RAAS biology.

Core Claim

The claim is that the biphasic gestational RAAS phenotype — high first-trimester aldosterone with a declining renin trajectory, terminating in third-trimester attenuation — predicts incident chronic hypertension and cardiovascular events in the mother at one, five, and ten years postpartum, independently of the hypertensive-disorder-of-pregnancy (HDP) diagnosis code, and that this can be established now by emulating a target trial inside linked multi-site electronic health records. The direction of the question is not obvious: in non-pregnant essential hypertension, aldosterone excess drives diastolic pressure upward (+3.66 ng/dL PAC per 10 mmHg of DBP in a 309-patient retrospective cohort; Relationship between blood pressure level and activity of renin-aldosterone axis in patients with essential hypertension—a retrospective study), while in pregnancy the same hormone first spikes and then collapses in the women who do worst. Whether the postpartum mother resembles the essential-hypertension patient or reverts to a reset volume-regulatory set point is precisely the open mechanistic question, and the retrospective emulation is the fastest instrument that can answer it.

Research Program

The first workstream is cohort construction. We will assemble, from linked obstetric–medical record systems of the kind used in the CALIBER emulation (primary care, hospital admissions, and death registry linkage spanning 1998–2016 in that study; Weight Change and the Onset of Cardiovascular Diseases) and following the multi-site aggregation logic demonstrated across three academic hospitals in the AICare evaluation (Augmenting Clinical Decision-Making with an Interactive and Interpretable AI Copilot), all deliveries over a 15–20 year window with at least one first-trimester and one later-pregnancy clinical RAAS panel (plasma renin activity, plasma aldosterone concentration, aldosterone-to-renin ratio — the same assays extracted retrospectively by Relationship between blood pressure level and activity of renin-aldosterone axis in patients with essential hypertension). This is admittedly a lab-enriched subgroup, since RAAS panels are not routine prenatal care; the selection problem is handled explicitly below rather than wished away.

The second workstream defines the exposure. We anchor to the LC-MS/MS findings of the Kentucky cohorts but recalibrate thresholds to cohort-specific percentiles, because no international reference ranges for gestational RAAS peptides exist — a limitation the Kentucky authors themselves flag. The phenotype is

$$\text{Decline}_i = \mathbb{1}\left[\text{PAC}_{i,1} > Q_{75} \;\wedge\; \Delta\text{PRA}_i < 0\right], \qquad \Delta\text{PRA}_i = \text{PRA}_{i,3} - \text{PRA}_{i,1}$$

where $\text{PAC}_{i,1}$ is first-trimester plasma aldosterone for woman $i$, $Q_{75}$ its cohort 75th percentile, and $\Delta\text{PRA}_i$ the change in plasma renin activity from first ($\text{PRA}_{i,1}$) to third ($\text{PRA}_{i,3}$) trimester. This mirrors the Q4-declining-renin subgroup that carried the adverse outcome burden in Association of Elevated Serum Aldosterone Concentrations in Pregnancy with Hypertension.

The third workstream is the emulation itself, with an explicit protocol: eligibility (delivery at or beyond 20 weeks, mother alive and engaged with the health system at a 6-week postpartum index date), exposure assignment at the first-trimester lab, and outcomes of incident chronic hypertension, major adverse cardiovascular events, and chronic kidney disease at 1, 5, and 10 years. Inverse probability of treatment weighting handles baseline and time-varying confounders — critically including antihypertensive initiation, since labetalol suppresses renal renin release and would otherwise corrupt the trajectory measurement, a confound the Kentucky authors could not remove. Inverse probability of censoring weighting handles loss to follow-up, exactly as the CALIBER emulation did for its 7-year outcome horizon (Weight Change and the Onset of Cardiovascular Diseases). Nonlinear aldosterone–risk relationships are modeled with restricted cubic splines, the technique that exposed a U-shaped BMI–mortality association where linear models saw none (Association of body mass index with severity and mortality of COVID-19 pneumonia), and hospital-level clustering with multilevel models as in the 50,912-patient stroke readmission study (Risk factors associated with 31-day unplanned readmission).

The fourth workstream is mechanistic anchoring: re-contact a subset of the existing 128-woman Kentucky cohort for postpartum LC-MS/MS RAS Fingerprint sampling, to verify that the clinical-assay phenotype built from routine EHR labs corresponds to the mass-spectrometry phenotype the prospective studies measured. This is also where the rare extreme — placental hCG-driven adrenal stimulation in Mirror Image Syndrome (Shuchang Lyu et al.) — serves as a boundary case: the emulation should show whether ectopic adrenal stimulation and ordinary attenuation sit on one dose–response curve or two distinct postpartum trajectories.

Proposed Experiments

The primary analysis is an IPCW-adjusted Cox model estimating the hazard ratio for incident chronic hypertension at 10 years postpartum, comparing the Decline phenotype against the referent (normal or rising renin). The secondary analysis is discrimination: C-statistics at 1, 5, and 10 years for nested models, with conformal prediction intervals for individual-level risk calibration in the spirit of the statin-response model that achieved empirical coverage near its nominal 90% target (A Machine Learning Approach to Predicting Heterogeneous Lipid Responses to Statins in Type 2 Diabetes). The decisive checks are fourfold. First, the phenotype must add discrimination beyond HDP diagnosis codes — the pre-specified ablation. Second, a negative-control outcome plausibly unrelated to RAAS biology (for example, appendicitis or fracture admission) should show no phenotype association; a positive one flags residual confounding. Third, a lab-ordering-bias audit compares effect estimates within strata of ordering indication (hypertension workup versus other), because panels ordered only when clinicians already suspect disease is the single most credible threat to the whole design. Fourth, the IPCW-adjusted estimate is compared against a naive complete-case estimate to quantify how much truncation and informative dropout would have distorted a plain retrospective analysis — the obstetric analogue of the surrogate-versus-hard-endpoint gap the CRT meta-analysis documented (Comparative efficacy of image-guided techniques in cardiac resynchronization therapy).

Baselines

The naive reference point is what postpartum care does today: flag women with an HDP diagnosis code in the record and treat that binary flag as the risk marker. Concretely this is a Cox model with the HDP code, maternal age, and baseline body mass index as inputs; no reported result found on Lacuna for its long-horizon discrimination in postpartum women, so we will establish that number ourselves on the assembled cohort. The strongest existing method is the long-horizon EHR trajectory model of Predicting cardiovascular health trajectories in time-series electronic health records with LSTM models, which achieved AUROCs from 0.84 (blood pressure) to 0.99 (smoking) across 216,445 patients over 13 years of record time — but explicitly contained no obstetric histories or RAAS parameters. We will re-implement and tune it on our identical cohort, same feature window, same compute and tuning budget, and beat it only if the RAAS-informed model's C-statistic at the 5-year horizon exceeds it by at least 0.05 with bootstrap confidence intervals excluding zero, alongside net-reclassification improvement over the code-only Cox model. The key ablation strips the RAAS laboratories out of our own emulation pipeline, leaving diagnoses, demographics, medications, and encounter history: whatever discrimination survives is what the codes were already carrying, and whatever disappears is the specific contribution of the gestational RAAS phenotype — the ingredient this proposal exists to test. A second ablation removes the IPCW machinery, isolating whether the target-trial weighting changes the hazard ratio materially or is methodological decoration.

Why not just…?

Why not just extend the prospective Kentucky cohorts with postpartum visits? This would give protocolized LC-MS/MS measurement, which routine EHR assays cannot match — the retrospective phenotype will not contain angiotensin peptide panels, only the cruder PRA/PAC/ARR. That is a real weakness, and it is why the re-contact substudy is part of the design rather than an afterthought. But as the primary vehicle it fails on three counts the prospective papers themselves acknowledge: the cohorts are small (74 and 128 women), recruited exclusively from high-risk referral clinics, and would need another decade before the 10-year horizon closes. The emulation delivers that horizon now, at population scale, and the prospective re-contact anchors rather than replaces it.

Why not just use HDP diagnosis codes as the exposure? Codes are downstream of the physiology and demonstrably coarse: elevated first-trimester aldosterone appeared in women across all HDP subcategories as well as in women who stayed normotensive (Association of Elevated Serum Aldosterone Concentrations in Pregnancy with Hypertension), so the at-risk phenotype is not contained in the diagnosis. But the challenge cuts both ways, which is why the ablation is decisive rather than rhetorical: if the code-only model matches the RAAS-augmented model's discrimination, that is a genuine finding — it would mean cheap codes suffice for postpartum risk stratification and the labs add mechanism but not prediction — and we will report it as such.

Why not just train a large sequence model on the full EHR and let it find the signal? The LSTM trajectory work proves black-box models can exploit long record histories, but it modeled five general cardiovascular metrics with no pregnancy strata and no hormone data, and its output is a prediction score rather than a hazard ratio attached to a named exposure (Predicting cardiovascular health trajectories in time-series electronic health records with LSTM models). The question here is causal and mechanistic — does a specific dysregulated trajectory carry forward? — which needs the target-trial estimand, not only an AUROC. We nonetheless run the LSTM as a fair, equally tuned baseline; if it wins on discrimination while the emulation still delivers an interpretable and confounding-robust hazard ratio, both results stand and we will say so.

Decisive Checks

The proposal succeeds if the Decline phenotype shows a 10-year hazard ratio for incident chronic hypertension whose confidence interval excludes 1 after IPCW and passes the negative-control and ordering-bias audits, and if it improves 5-year discrimination over both the code-only Cox model and the tuned LSTM baseline. It fails informatively if the codes alone suffice — in which case the field learns that postpartum screening needs no hormone panel — or if the ordering-bias audit shows the phenotype is an artifact of who gets tested. Either terminal state resolves a question that, right now, every gestational RAAS study leaves open at the moment the baby is born.

References

  1. Activation of the Renin–Angiotensin–Aldosterone System Is Attenuated in Hypertensive Compared with Normotensive Pregnancy
  2. Association of Elevated Serum Aldosterone Concentrations in Pregnancy with Hypertension
  3. Comparative efficacy of image-guided techniques in cardiac resynchronization therapy: a meta-analysis
  4. A prospective evaluation of cardiovascular magnetic resonance measures of dyssynchrony in the prediction of response to cardiac resynchronization therapy
  5. Long-Term Survival in Men With High Grade Prostate Cancer: A Comparison Between Conservative Treatment, Radiation Therapy and Radical Prostatectomy—A Propensity Scoring Approach
  6. Weight Change and the Onset of Cardiovascular Diseases: Emulating Trials Using Electronic Health Records
  7. Relationship between blood pressure level and activity of renin-aldosterone axis in patients with essential hypertension—a retrospective study
  8. Augmenting Clinical Decision-Making with an Interactive and Interpretable AI Copilot: A Real-World User Study with Clinicians in Nephrology and Obstetrics
  9. Association of body mass index with severity and mortality of COVID-19 pneumonia: a two-center, retrospective cohort study from Wuhan, China
  10. Risk factors associated with 31-day unplanned readmission in 50,912 discharged patients after stroke in China
  11. Mirror Image Syndrome: 1 Case and Literature Review
  12. A Machine Learning Approach to Predicting Heterogeneous Lipid Responses to Statins in Type 2 Diabetes
  13. Predicting cardiovascular health trajectories in time-series electronic health records with LSTM models