Research

Delfina at SMFM 2027: validating pregnancy risk models at Mayo Clinic and extending an HDP model to self-reported data

Authors
Authors
Authors
https://www.delfina.com/resource/delfina-at-smfm-2027

We’re proud to share that Delfina has been selected to present two abstracts at SMFM 2027. Both come from our data science and clinical teams, and both ask the same question: how do we make sure predictive tools stay accurate and useful for the patients and care teams who rely on them?

Pregnancy complications are often identified too late. Earlier, more reliable risk prediction gives care teams a chance to step in sooner. Here is a look at what we’re presenting.

1. Validating pregnancy risk models at Mayo Clinic

Clinical validation and workflow integration of predictive models for cardiometabolic pregnancy complications at Mayo Clinic

Authors: Isabel Fulcher (Presenter), Dennis Hu, Sara Sauer, Emanuel Trabuco, Regan Theiler, Tarek Khalife

Prediction models can lose accuracy as patient populations and care patterns change. To find out whether ours hold up, we revalidated three Delfina models on 14,149 Mayo Clinic Health System deliveries from 2022 to 2024. The models predict hypertensive disorders of pregnancy (HDP), gestational diabetes (GDM), and excessive gestational weight gain (eGWG).

All three performed consistently with the original 2017-2021 training cohort:

•  HDP: AUC 0.75

•  GDM: AUC 0.70

•  eGWG: AUC 0.81

A good model only matters if care teams can act on it. So a panel of 5 OB-GYNs, 4 registered nurses, and 1 certified nurse midwife defined risk tiers and matched each one to evidence-based interventions, including remote blood pressure monitoring, aspirin, dietitian support, and group classes. Our modeling estimates that 67% of patients would be targeted for at least one preventive intervention.

The takeaway: validated models can stay stable over time. Paired with workflows that clinicians help design, they give health systems a scalable way to deliver precision pregnancy care.

2. Extending an HDP model to self-reported data

Transfer Learning Improves Generalization of an EHR-Trained Hypertensive Disorders of Pregnancy Model to Self-Reported Data

Authors: Dennis Hu (Presenter), Isabel Fulcher, Sara Sauer

Most clinical prediction models are trained on one health system’s electronic health record (EHR) data. That limits how well they work for patients whose information comes from somewhere else, including the self-reported data that digital care platforms collect between visits.

To close that gap, we fine-tuned our HDP model using Bayesian transfer learning. We combined updated EHR data, EHR data from integrated prenatal clinics, and self-reported survey data from a digital pregnancy-care platform, across 20,536 pregnancies.

Overall performance matched our predecessor model (AUC 0.74) and stayed consistent across data sources:

•  EHR only: AUC 0.76

•  EHR and self-report: AUC 0.73

•  Self-report only: AUC 0.72

We also compared the model with the ACOG checklist for aspirin initiation, with two results per 100 patients:

•  At matched specificity, the model correctly identified 18 more patients with HDP.

•  At matched sensitivity, it recommended aspirin unnecessarily for 21 fewer patients.

The takeaway: transfer learning offers a scalable way to extend validated clinical models to digital health populations with mixed data sources.

‍

Let’s connect at SMFM 2027

If you’re interested in following up on this work, exploring a collaboration, or meeting with our team during SMFM, we’d love to hear from you. Please email us at partner@delfina.com.

Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.
No items found.
Research

Delfina at SMFM 2027: validating pregnancy risk models at Mayo Clinic and extending an HDP model to self-reported data

Authors
Authors
Authors
https://www.delfina.com/resource/delfina-at-smfm-2027

We’re proud to share that Delfina has been selected to present two abstracts at SMFM 2027. Both come from our data science and clinical teams, and both ask the same question: how do we make sure predictive tools stay accurate and useful for the patients and care teams who rely on them?

Pregnancy complications are often identified too late. Earlier, more reliable risk prediction gives care teams a chance to step in sooner. Here is a look at what we’re presenting.

1. Validating pregnancy risk models at Mayo Clinic

Clinical validation and workflow integration of predictive models for cardiometabolic pregnancy complications at Mayo Clinic

Authors: Isabel Fulcher (Presenter), Dennis Hu, Sara Sauer, Emanuel Trabuco, Regan Theiler, Tarek Khalife

Prediction models can lose accuracy as patient populations and care patterns change. To find out whether ours hold up, we revalidated three Delfina models on 14,149 Mayo Clinic Health System deliveries from 2022 to 2024. The models predict hypertensive disorders of pregnancy (HDP), gestational diabetes (GDM), and excessive gestational weight gain (eGWG).

All three performed consistently with the original 2017-2021 training cohort:

•  HDP: AUC 0.75

•  GDM: AUC 0.70

•  eGWG: AUC 0.81

A good model only matters if care teams can act on it. So a panel of 5 OB-GYNs, 4 registered nurses, and 1 certified nurse midwife defined risk tiers and matched each one to evidence-based interventions, including remote blood pressure monitoring, aspirin, dietitian support, and group classes. Our modeling estimates that 67% of patients would be targeted for at least one preventive intervention.

The takeaway: validated models can stay stable over time. Paired with workflows that clinicians help design, they give health systems a scalable way to deliver precision pregnancy care.

2. Extending an HDP model to self-reported data

Transfer Learning Improves Generalization of an EHR-Trained Hypertensive Disorders of Pregnancy Model to Self-Reported Data

Authors: Dennis Hu (Presenter), Isabel Fulcher, Sara Sauer

Most clinical prediction models are trained on one health system’s electronic health record (EHR) data. That limits how well they work for patients whose information comes from somewhere else, including the self-reported data that digital care platforms collect between visits.

To close that gap, we fine-tuned our HDP model using Bayesian transfer learning. We combined updated EHR data, EHR data from integrated prenatal clinics, and self-reported survey data from a digital pregnancy-care platform, across 20,536 pregnancies.

Overall performance matched our predecessor model (AUC 0.74) and stayed consistent across data sources:

•  EHR only: AUC 0.76

•  EHR and self-report: AUC 0.73

•  Self-report only: AUC 0.72

We also compared the model with the ACOG checklist for aspirin initiation, with two results per 100 patients:

•  At matched specificity, the model correctly identified 18 more patients with HDP.

•  At matched sensitivity, it recommended aspirin unnecessarily for 21 fewer patients.

The takeaway: transfer learning offers a scalable way to extend validated clinical models to digital health populations with mixed data sources.

‍

Let’s connect at SMFM 2027

If you’re interested in following up on this work, exploring a collaboration, or meeting with our team during SMFM, we’d love to hear from you. Please email us at partner@delfina.com.

Research

Delfina at SMFM 2027: validating pregnancy risk models at Mayo Clinic and extending an HDP model to self-reported data

Authors
Authors
Authors
https://www.delfina.com/resource/delfina-at-smfm-2027

We’re proud to share that Delfina has been selected to present two abstracts at SMFM 2027. Both come from our data science and clinical teams, and both ask the same question: how do we make sure predictive tools stay accurate and useful for the patients and care teams who rely on them?

Pregnancy complications are often identified too late. Earlier, more reliable risk prediction gives care teams a chance to step in sooner. Here is a look at what we’re presenting.

1. Validating pregnancy risk models at Mayo Clinic

Clinical validation and workflow integration of predictive models for cardiometabolic pregnancy complications at Mayo Clinic

Authors: Isabel Fulcher (Presenter), Dennis Hu, Sara Sauer, Emanuel Trabuco, Regan Theiler, Tarek Khalife

Prediction models can lose accuracy as patient populations and care patterns change. To find out whether ours hold up, we revalidated three Delfina models on 14,149 Mayo Clinic Health System deliveries from 2022 to 2024. The models predict hypertensive disorders of pregnancy (HDP), gestational diabetes (GDM), and excessive gestational weight gain (eGWG).

All three performed consistently with the original 2017-2021 training cohort:

•  HDP: AUC 0.75

•  GDM: AUC 0.70

•  eGWG: AUC 0.81

A good model only matters if care teams can act on it. So a panel of 5 OB-GYNs, 4 registered nurses, and 1 certified nurse midwife defined risk tiers and matched each one to evidence-based interventions, including remote blood pressure monitoring, aspirin, dietitian support, and group classes. Our modeling estimates that 67% of patients would be targeted for at least one preventive intervention.

The takeaway: validated models can stay stable over time. Paired with workflows that clinicians help design, they give health systems a scalable way to deliver precision pregnancy care.

2. Extending an HDP model to self-reported data

Transfer Learning Improves Generalization of an EHR-Trained Hypertensive Disorders of Pregnancy Model to Self-Reported Data

Authors: Dennis Hu (Presenter), Isabel Fulcher, Sara Sauer

Most clinical prediction models are trained on one health system’s electronic health record (EHR) data. That limits how well they work for patients whose information comes from somewhere else, including the self-reported data that digital care platforms collect between visits.

To close that gap, we fine-tuned our HDP model using Bayesian transfer learning. We combined updated EHR data, EHR data from integrated prenatal clinics, and self-reported survey data from a digital pregnancy-care platform, across 20,536 pregnancies.

Overall performance matched our predecessor model (AUC 0.74) and stayed consistent across data sources:

•  EHR only: AUC 0.76

•  EHR and self-report: AUC 0.73

•  Self-report only: AUC 0.72

We also compared the model with the ACOG checklist for aspirin initiation, with two results per 100 patients:

•  At matched specificity, the model correctly identified 18 more patients with HDP.

•  At matched sensitivity, it recommended aspirin unnecessarily for 21 fewer patients.

The takeaway: transfer learning offers a scalable way to extend validated clinical models to digital health populations with mixed data sources.

‍

Let’s connect at SMFM 2027

If you’re interested in following up on this work, exploring a collaboration, or meeting with our team during SMFM, we’d love to hear from you. Please email us at partner@delfina.com.

Research

Delfina at SMFM 2027: validating pregnancy risk models at Mayo Clinic and extending an HDP model to self-reported data

Authors
Authors
Authors
https://www.delfina.com/resource/delfina-at-smfm-2027

We’re proud to share that Delfina has been selected to present two abstracts at SMFM 2027. Both come from our data science and clinical teams, and both ask the same question: how do we make sure predictive tools stay accurate and useful for the patients and care teams who rely on them?

Pregnancy complications are often identified too late. Earlier, more reliable risk prediction gives care teams a chance to step in sooner. Here is a look at what we’re presenting.

1. Validating pregnancy risk models at Mayo Clinic

Clinical validation and workflow integration of predictive models for cardiometabolic pregnancy complications at Mayo Clinic

Authors: Isabel Fulcher (Presenter), Dennis Hu, Sara Sauer, Emanuel Trabuco, Regan Theiler, Tarek Khalife

Prediction models can lose accuracy as patient populations and care patterns change. To find out whether ours hold up, we revalidated three Delfina models on 14,149 Mayo Clinic Health System deliveries from 2022 to 2024. The models predict hypertensive disorders of pregnancy (HDP), gestational diabetes (GDM), and excessive gestational weight gain (eGWG).

All three performed consistently with the original 2017-2021 training cohort:

•  HDP: AUC 0.75

•  GDM: AUC 0.70

•  eGWG: AUC 0.81

A good model only matters if care teams can act on it. So a panel of 5 OB-GYNs, 4 registered nurses, and 1 certified nurse midwife defined risk tiers and matched each one to evidence-based interventions, including remote blood pressure monitoring, aspirin, dietitian support, and group classes. Our modeling estimates that 67% of patients would be targeted for at least one preventive intervention.

The takeaway: validated models can stay stable over time. Paired with workflows that clinicians help design, they give health systems a scalable way to deliver precision pregnancy care.

2. Extending an HDP model to self-reported data

Transfer Learning Improves Generalization of an EHR-Trained Hypertensive Disorders of Pregnancy Model to Self-Reported Data

Authors: Dennis Hu (Presenter), Isabel Fulcher, Sara Sauer

Most clinical prediction models are trained on one health system’s electronic health record (EHR) data. That limits how well they work for patients whose information comes from somewhere else, including the self-reported data that digital care platforms collect between visits.

To close that gap, we fine-tuned our HDP model using Bayesian transfer learning. We combined updated EHR data, EHR data from integrated prenatal clinics, and self-reported survey data from a digital pregnancy-care platform, across 20,536 pregnancies.

Overall performance matched our predecessor model (AUC 0.74) and stayed consistent across data sources:

•  EHR only: AUC 0.76

•  EHR and self-report: AUC 0.73

•  Self-report only: AUC 0.72

We also compared the model with the ACOG checklist for aspirin initiation, with two results per 100 patients:

•  At matched specificity, the model correctly identified 18 more patients with HDP.

•  At matched sensitivity, it recommended aspirin unnecessarily for 21 fewer patients.

The takeaway: transfer learning offers a scalable way to extend validated clinical models to digital health populations with mixed data sources.

‍

Let’s connect at SMFM 2027

If you’re interested in following up on this work, exploring a collaboration, or meeting with our team during SMFM, we’d love to hear from you. Please email us at partner@delfina.com.

Research

Delfina at SMFM 2027: validating pregnancy risk models at Mayo Clinic and extending an HDP model to self-reported data

Authors
Authors
Authors
https://www.delfina.com/resource/delfina-at-smfm-2027

We’re proud to share that Delfina has been selected to present two abstracts at SMFM 2027. Both come from our data science and clinical teams, and both ask the same question: how do we make sure predictive tools stay accurate and useful for the patients and care teams who rely on them?

Pregnancy complications are often identified too late. Earlier, more reliable risk prediction gives care teams a chance to step in sooner. Here is a look at what we’re presenting.

1. Validating pregnancy risk models at Mayo Clinic

Clinical validation and workflow integration of predictive models for cardiometabolic pregnancy complications at Mayo Clinic

Authors: Isabel Fulcher (Presenter), Dennis Hu, Sara Sauer, Emanuel Trabuco, Regan Theiler, Tarek Khalife

Prediction models can lose accuracy as patient populations and care patterns change. To find out whether ours hold up, we revalidated three Delfina models on 14,149 Mayo Clinic Health System deliveries from 2022 to 2024. The models predict hypertensive disorders of pregnancy (HDP), gestational diabetes (GDM), and excessive gestational weight gain (eGWG).

All three performed consistently with the original 2017-2021 training cohort:

•  HDP: AUC 0.75

•  GDM: AUC 0.70

•  eGWG: AUC 0.81

A good model only matters if care teams can act on it. So a panel of 5 OB-GYNs, 4 registered nurses, and 1 certified nurse midwife defined risk tiers and matched each one to evidence-based interventions, including remote blood pressure monitoring, aspirin, dietitian support, and group classes. Our modeling estimates that 67% of patients would be targeted for at least one preventive intervention.

The takeaway: validated models can stay stable over time. Paired with workflows that clinicians help design, they give health systems a scalable way to deliver precision pregnancy care.

2. Extending an HDP model to self-reported data

Transfer Learning Improves Generalization of an EHR-Trained Hypertensive Disorders of Pregnancy Model to Self-Reported Data

Authors: Dennis Hu (Presenter), Isabel Fulcher, Sara Sauer

Most clinical prediction models are trained on one health system’s electronic health record (EHR) data. That limits how well they work for patients whose information comes from somewhere else, including the self-reported data that digital care platforms collect between visits.

To close that gap, we fine-tuned our HDP model using Bayesian transfer learning. We combined updated EHR data, EHR data from integrated prenatal clinics, and self-reported survey data from a digital pregnancy-care platform, across 20,536 pregnancies.

Overall performance matched our predecessor model (AUC 0.74) and stayed consistent across data sources:

•  EHR only: AUC 0.76

•  EHR and self-report: AUC 0.73

•  Self-report only: AUC 0.72

We also compared the model with the ACOG checklist for aspirin initiation, with two results per 100 patients:

•  At matched specificity, the model correctly identified 18 more patients with HDP.

•  At matched sensitivity, it recommended aspirin unnecessarily for 21 fewer patients.

The takeaway: transfer learning offers a scalable way to extend validated clinical models to digital health populations with mixed data sources.

‍

Let’s connect at SMFM 2027

If you’re interested in following up on this work, exploring a collaboration, or meeting with our team during SMFM, we’d love to hear from you. Please email us at partner@delfina.com.

Research

Delfina at SMFM 2027: validating pregnancy risk models at Mayo Clinic and extending an HDP model to self-reported data

https://www.delfina.com/resource/delfina-at-smfm-2027

We’re proud to share that Delfina has been selected to present two abstracts at SMFM 2027. Both come from our data science and clinical teams, and both ask the same question: how do we make sure predictive tools stay accurate and useful for the patients and care teams who rely on them?

Pregnancy complications are often identified too late. Earlier, more reliable risk prediction gives care teams a chance to step in sooner. Here is a look at what we’re presenting.

1. Validating pregnancy risk models at Mayo Clinic

Clinical validation and workflow integration of predictive models for cardiometabolic pregnancy complications at Mayo Clinic

Authors: Isabel Fulcher (Presenter), Dennis Hu, Sara Sauer, Emanuel Trabuco, Regan Theiler, Tarek Khalife

Prediction models can lose accuracy as patient populations and care patterns change. To find out whether ours hold up, we revalidated three Delfina models on 14,149 Mayo Clinic Health System deliveries from 2022 to 2024. The models predict hypertensive disorders of pregnancy (HDP), gestational diabetes (GDM), and excessive gestational weight gain (eGWG).

All three performed consistently with the original 2017-2021 training cohort:

•  HDP: AUC 0.75

•  GDM: AUC 0.70

•  eGWG: AUC 0.81

A good model only matters if care teams can act on it. So a panel of 5 OB-GYNs, 4 registered nurses, and 1 certified nurse midwife defined risk tiers and matched each one to evidence-based interventions, including remote blood pressure monitoring, aspirin, dietitian support, and group classes. Our modeling estimates that 67% of patients would be targeted for at least one preventive intervention.

The takeaway: validated models can stay stable over time. Paired with workflows that clinicians help design, they give health systems a scalable way to deliver precision pregnancy care.

2. Extending an HDP model to self-reported data

Transfer Learning Improves Generalization of an EHR-Trained Hypertensive Disorders of Pregnancy Model to Self-Reported Data

Authors: Dennis Hu (Presenter), Isabel Fulcher, Sara Sauer

Most clinical prediction models are trained on one health system’s electronic health record (EHR) data. That limits how well they work for patients whose information comes from somewhere else, including the self-reported data that digital care platforms collect between visits.

To close that gap, we fine-tuned our HDP model using Bayesian transfer learning. We combined updated EHR data, EHR data from integrated prenatal clinics, and self-reported survey data from a digital pregnancy-care platform, across 20,536 pregnancies.

Overall performance matched our predecessor model (AUC 0.74) and stayed consistent across data sources:

•  EHR only: AUC 0.76

•  EHR and self-report: AUC 0.73

•  Self-report only: AUC 0.72

We also compared the model with the ACOG checklist for aspirin initiation, with two results per 100 patients:

•  At matched specificity, the model correctly identified 18 more patients with HDP.

•  At matched sensitivity, it recommended aspirin unnecessarily for 21 fewer patients.

The takeaway: transfer learning offers a scalable way to extend validated clinical models to digital health populations with mixed data sources.

‍

Let’s connect at SMFM 2027

If you’re interested in following up on this work, exploring a collaboration, or meeting with our team during SMFM, we’d love to hear from you. Please email us at partner@delfina.com.

Other videos

View all
No items found.
Research

Delfina at SMFM 2027: validating pregnancy risk models at Mayo Clinic and extending an HDP model to self-reported data

Guests
No items found.
https://www.delfina.com/resource/delfina-at-smfm-2027