Deep Dive Podcast:

Introduction: The High Stakes of Maternal Care

Every seven minutes in the United States, a woman is diagnosed with preeclampsia—a pregnancy complication that remains one of the leading causes of maternal and fetal morbidity worldwide. Despite advances in obstetric care, the ability to accurately predict which patients will develop severe complications has remained stubbornly elusive. For those of us who straddle the worlds of clinical medicine and software development—the “Doctors Who Code”—this represents both a challenge and an opportunity.

The stakes couldn’t be higher. A woman with preeclampsia faces risks ranging from seizures and stroke to multi-organ failure. Yet overly aggressive management brings its own complications: iatrogenic prematurity, unnecessary interventions, and increased healthcare costs. This delicate balance makes accurate risk prediction not just helpful, but potentially lifesaving.

Today, we’re at a pivotal moment in maternal risk assessment. A recent multicountry study published in PLOS Medicine (Yang et al., 2025) has shed critical light on how our current preeclampsia risk models perform when used repeatedly over time—findings that have profound implications for the development of our new PreeclampsiaWatch application.

Understanding Preeclampsia and Clinical Prediction Models

What Is Preeclampsia?

Preeclampsia is a multisystem disorder characterized by hypertension and often proteinuria after 20 weeks of gestation. What makes it particularly challenging is its unpredictable course—some women experience a relatively benign progression, while others deteriorate rapidly toward severe complications including HELLP syndrome, placental abruption, and eclampsia.

Risk factors include nulliparity, advanced maternal age, chronic hypertension, diabetes, obesity, and personal or family history of preeclampsia. However, these factors alone provide limited precision in predicting which women will develop severe complications.

Static vs. Consecutive Prediction Models

Most clinical prediction models in medicine, including those for preeclampsia, were designed as “static” tools—they take a single snapshot of patient data to generate risk scores. The PIERS (Preeclampsia Integrated Estimate of RiSk) models represent the most validated approach, with two key variants:

• fullPIERS model: A logistic regression-based model developed from data from well-resourced settings
• PIERS-ML model: A more recent machine learning approach trained on globally diverse data

Both were developed to predict adverse maternal outcomes within 48 hours of hospital admission with preeclampsia. But here’s where clinical reality introduces complexity: in practice, clinicians don’t assess patients just once. They continually reassess as new labs and vitals become available—especially during expectant management of preterm preeclampsia.

This is where “consecutive” (or serial) prediction becomes critical. A consecutive prediction uses the same model repeatedly as new data becomes available. While this approach intuitively makes sense, the crucial question is whether these static models maintain their accuracy when deployed serially.

Key Findings from the “Consecutive Prediction” Study

The recently published PLOS Medicine study by Yang et al. represents the largest evaluation to date of how the PIERS models perform when used for ongoing risk assessment. The researchers analyzed data from 8,843 women with preeclampsia across multiple continents, including diverse populations (32% white, 30% black, and 26% Asian).

Study Background

The researchers tracked the performance of both the PIERS-ML and fullPIERS models during a 2-week window following hospital admission. This multicountry design is particularly valuable as it captures diverse healthcare systems and patient populations, enhancing the generalizability of findings.

Model Performance Highlights

The results reveal critical insights that any clinician-developer should consider:

1. Initial Strength: Both models showed good performance during the first 48 hours after admission, with the PIERS-ML model demonstrating particularly strong discrimination (AUC-PRC of 0.65 on day 0).
2. Performance Drift: The accuracy of both models decreased noticeably beyond 48 hours. This deterioration was especially pronounced for moderate risk classifications, with AUC-PRC values dropping significantly and remaining between 0.1 and 0.5 after day 2.
3. High-Risk & Very High-Risk Alerts: The models maintained excellent “rule-in” capacity for the highest risk groups, with positive likelihood ratios ranging from 70.99 to infinity for the “very high risk” group. However, these groups represented only a small fraction of patients.
4. Clinical Utility: Decision curve analysis revealed a diminishing advantage for treatment guided by these models over time, suggesting reduced clinical impact beyond the first 48 hours.

Clinical Implication

The key takeaway is stark but crucial: static models must be used with increasing caution for ongoing, day-by-day care. As time progresses, the reliability of risk estimates declines unless the models are recalibrated or updated with fresh clinical data.

This finding has significant implications for software designers developing clinical decision support tools for preeclampsia management. The study doesn’t suggest abandoning these models—rather, it highlights the need to evolve our approach to risk prediction over time.

Practical Insights for “Doctors Who Code”

As physicians who develop medical software, we must translate these research findings into actionable development strategies. Here are key considerations:

Handling Performance Drift

The observed performance drift isn’t unique to preeclampsia prediction—it’s a known challenge in clinical prediction modeling. When static models are re-run daily without accounting for changing patient conditions, their accuracy naturally deteriorates.

To combat this, consider:

• Recalibration Strategies: Implementing automated recalibration of model parameters based on recent patient outcomes
• Landmark Modeling: Using different models or parameters based on how much time has elapsed since admission
• Explicit Time Variables: Including time-since-admission as an explicit predictor in your models
• Dynamic Weighting: Adjusting the importance of predictors based on their observed stability over time

Data Quality and Frequency

The study underscores that reliable risk assessment requires regular data capture. For PreeclampsiaWatch development, this means:

• Optimizing Data Collection: Designing intuitive interfaces that make it easy to enter updated measurements without disrupting clinical workflow
• Automated Data Pulls: Integrating with EHR systems to automatically incorporate new labs and vitals as they become available
• Missing Data Handling: Implementing sophisticated strategies for handling missing or delayed measurements
• Temporal Visualization: Providing clinicians with clear visualization of how predictors and risk scores change over time

Balancing Model Complexity & Clinical Workflow

Complex models that require extensive data entry aren’t helpful if they impede clinical care. The PreeclampsiaWatch application needs to:

• Prioritize Simplicity: Focus on the most discriminative predictors rather than requiring exhaustive data entry
• Emphasize Interpretability: Ensure clinicians understand why risk estimates change
• Provide Context: Clearly communicate the model’s strengths and limitations at different time points
• Support Decision-Making: Translate risk scores into concrete clinical recommendations while respecting clinician autonomy

Specific Implications for the PreeclampsiaWatch App

Based on the study findings, here are specific recommendations for our PreeclampsiaWatch web and mobile applications:

Serial Monitoring Requirements

The app should be designed from the ground up to support repeated measurements:

• Longitudinal Data Storage: Maintain complete histories of all measurements to enable trend analysis
• Intelligent Prompting: Remind users to update key measurements at clinically appropriate intervals
• Risk Trajectory Display: Visualize how risk changes over time rather than presenting only the current estimate
• Alert Calibration: Implement time-aware thresholds that account for the changing reliability of predictions

Caution in Consecutive Predictions

The app should explicitly address the performance drift issue:

• Confidence Intervals: Display widening confidence intervals for predictions beyond 48 hours
• Time-Since-Admission Indicator: Clearly show how much time has elapsed since initial assessment
• Explicit Disclaimers: Include appropriate cautions about the decreasing reliability of extended predictions
• Clinical Override: Provide mechanisms for clinicians to document their reasoning when they diverge from model recommendations

Adaptive or Dynamic Updating

While implementing fully dynamic models would require additional research, the app can incorporate elements that mitigate performance drift:

• Trend-Based Risk Modification: Adjust risk estimates based on the trajectory of key measurements rather than just their current values
• Modular Prediction: Use different models or parameters based on time elapsed and clinical context
• Continuous Learning: If institutional approval allows, implement privacy-preserving mechanisms to learn from local outcomes and improve prediction
• Research Integration: Design the app architecture to facilitate seamless integration of new models as research advances

User Interface & Patient Engagement

The interface should reflect the nuanced nature of preeclampsia risk:

• Contextual Guidance: Provide different clinical recommendations based not just on risk level but on how reliable that estimate is
• Patient-Facing Tools: Develop appropriate visualizations and explanations for shared decision-making
• Action Thresholds: Integrate clinical guidelines to suggest appropriate next steps when risk changes significantly
• Documentation Support: Help clinicians document their risk assessment and clinical reasoning in the EHR

Implementation Roadmap for Clinician-Developers

Translating these insights into a production-ready application requires careful planning. Here’s a proposed roadmap:

Data Infrastructure

• EHR Integration: Develop robust APIs for bidirectional data exchange with major EHR systems
• FHIR Compatibility: Implement FHIR resources for seamless interoperability
• Local Data Storage: Enable offline functionality with appropriate synchronization
• Privacy Protection: Implement end-to-end encryption and compliance with healthcare privacy regulations

Model Monitoring & Recalibration

• Performance Tracking: Implement analytics to monitor how well predictions match outcomes
• Calibration Tooling: Develop administrative interfaces for authorized users to review and adjust model parameters
• Versioning System: Maintain clear documentation of model versions and changes
• Feedback Loops: Create mechanisms for clinicians to report perceived inaccuracies

Clinical Validation & Pilot Testing

• Simulation Testing: Test with retrospective data before deployment
• Limited Rollout: Begin with a small group of interested clinicians
• Structured Evaluation: Define clear metrics for success and monitoring protocols
• Iterative Refinement: Plan for regular updates based on user feedback and observed performance

Future Developments in Preeclampsia Prediction

Looking beyond our initial release, several promising approaches could enhance PreeclampsiaWatch:

Dynamic Modeling

Recent advances in machine learning offer potential paths to more robust consecutive prediction:

• Recurrent Neural Networks: These architectures can explicitly model temporal dependencies
• Attention Mechanisms: These can help models focus on the most relevant historical data points
• Transfer Learning: Models pretrained on large datasets can be fine-tuned for specific populations

Remote Monitoring Integration

Wearable technology opens new possibilities for continuous monitoring:

• Automated BP Measurement: Integration with validated home blood pressure monitors
• Activity Tracking: Incorporating activity and sleep patterns as potential risk modifiers
• Symptom Logging: Patient-reported symptoms via mobile interfaces
• Alert Thresholds: Personalized thresholds for remote monitoring alerts

Personalization

The future of preeclampsia prediction lies in hyper-personalized approaches:

• Genomic Integration: Incorporating genetic risk factors when available
• Pregnancy History: Weighting models based on previous pregnancy outcomes
• Biomarker Discovery: Integration of novel biomarkers as they become clinically validated
• Patient Preferences: Adjusting recommendations based on documented patient values and priorities

Conclusion: Advancing Safer Maternal Care Through Smart Software

The new research on consecutive prediction in preeclampsia highlights both promise and pitfall. The PIERS-ML and fullPIERS models represent significant advances in risk stratification, but their limitations when used serially must inform how we develop and deploy tools like PreeclampsiaWatch.

For clinician-developers building maternal health applications, these findings underscore the need for humility and transparency. Our models are powerful but imperfect, and our software should reflect both their strengths and limitations. By designing for performance drift, incorporating serial data thoughtfully, and maintaining rigorous validation cycles, we can develop tools that genuinely enhance clinical decision-making rather than offering a false sense of precision.

Call to Action: Join the PreeclampsiaWatch Development Community

Preeclampsia remains one of the most significant challenges in maternal care, and addressing it effectively requires collaboration between clinicians, data scientists, and software developers. We invite you to join us in developing the next generation of PreeclampsiaWatch:

• Clinical Partners: We’re seeking obstetric practices interested in pilot testing
• Technical Contributors: Developers with healthcare experience can contribute to our open-source components
• Data Scientists: Help us develop and validate more robust consecutive prediction algorithms
• User Experience: Provide feedback on interface designs and clinical workflows

Together, we can build tools that translate cutting-edge research into everyday clinical practice, helping physicians make more informed decisions and ultimately improving outcomes for mothers and babies around the world.

Ready to make a difference? Contact us at partners@preeclampsiawatch.com to learn how you can contribute to this vital effort and be among the first to implement PreeclampsiaWatch in your practice.