Delfina

Pregnancy risk prediction is an important part of a broader maternal care service. The platform combines clinical records and patient-entered measurements to help identify people who may benefit from earlier intervention. Clinicians review insights through a dashboard, while the app supports ongoing symptom and vital-sign collection. However, monitoring devices, education, care guides, doula support, and separate medical services also provide value without a predictive model.

AI therefore contributes to personalization and risk identification rather than constituting the whole offering. Its effect depends on how care teams interpret the predictions and connect them with the human support available through each program.

Clinicians decide which patients receive monitoring devices and may adjust measurement frequency. The provider dashboard highlights out-of-range readings, patients awaiting intervention, and recent inputs, giving staff a defined review role. Guides monitor participation and help patients use the program, while medical questions remain with the clinical team. Patient terms caution against relying on the platform for active emergency monitoring.

Exact response thresholds, coverage hours for clinical review, escalation timing, and recovery after a missed signal remain unclear. The workflow therefore combines patient data collection with professional oversight, without a complete commitment that every concerning reading receives an immediate response.

The weight-gain research identifies a gradient boosting classifier using clinical and self-reported measurements through 24 weeks of pregnancy. Inputs include weight trajectory, prepregnancy BMI, age, parity, chronic hypertension, and pregestational diabetes. Other pregnancy modeling research also describes conventional machine-learning approaches, giving more technical substance than a generic predictive AI label.

However, production model versions, their training lineage, and the relationship between the research models and current dashboard functions remain unspecified. Model update timing and release review procedures are also unclear. The known research method therefore explains a particular analysis without fully identifying the deployed prediction stack.

Proprietary pregnancy predictions sit alongside patient communication, support, and other app functions. Privacy terms describe categories of service providers and research partners, but do not map each clinical model to its infrastructure and processors. The full list of model providers, their access to patient information, and downstream retention or training obligations remain unspecified. Named analytics or communication tools cannot establish the complete inference chain for a pregnancy risk prediction.

A 2026 conference analysis included 1,088 pregnancies from four community clinics using data collected from 2021 to 2025. A gradient boosting model predicted gestational weight-gain categories using five-fold cross-validation and achieved an overall AUC of 0.85. The study combined clinical and self-reported weights and excluded preterm births, multiple gestations, and pregnancies with insufficient weight data.

This supplies a defined sample, method, and outcome rather than a testimonial alone. It remains an internal research evaluation, with independent validation and a clear mapping to the current production model unspecified. The result does not establish performance for every pregnancy complication.

The patient FAQ limits research to anonymized information and says identifying information is erased when a participant leaves. The broader privacy policy permits predictive-model development and improvement using information within its scope, while secure clinical records follow separate medical-group terms. Those boundaries prevent treating the policy as a blanket permission to train on all PHI.

Fixed retention periods, the exact anonymization method, and complete training restrictions across the service remain unclear. Clinical review and monitoring support oversight, but detailed hallucination controls and model safety incident procedures are unspecified. The research-use commitment is therefore meaningful but narrower than complete lifecycle assurance.

The service claims HIPAA compliance with employee training, restricted access to health information, and breach notification procedures. Its privacy policy covers the app, software, devices, and nonclinical support, making it more substantive than a website-only notice. It separately places information in the secure clinical platform under HIPAA and the medical group's applicable privacy notice, with restrictions on advertising and data mining.

A readable BAA and its product and subcontractor scope remain unavailable. The separation between Delfina's administrative platform and independent medical practices is important because different notices and responsibilities apply to different portions of a patient's information.

SOC 2 Type II coverage is claimed for security, confidentiality, and availability. The security description also includes encryption at rest and during transmission, mandatory employee multifactor authentication, third-party security audits, restricted health-data access, and continuous security monitoring. These are substantive control descriptions alongside a recognized assurance claim.

A current report with its audit period, covered service boundaries, exceptions, and independent opinion remains unavailable for direct confirmation. HITRUST is described as forthcoming rather than an achieved certification. The visible material supports the stated security posture without establishing the exact current audit scope for every predictive and clinical service component.

Current product claims include predicting pregnancy complications and enabling proactive clinical interventions. These functions can influence care decisions beyond simple storage or administrative messaging. The consumer legal terms instead describe informational and educational software, separate medical practices, and limits on reliance for medical decisions. A clear regulatory classification and any applicable clearance scope for the current predictive functions remain unspecified.

The existence of separate clinicians does not by itself resolve the status of the software producing risk information. This leaves a material gap between the clinical role described for the product and the narrower role stated in its legal terms.

The 2026 weight-gain research compares AUC across race and body mass index groups using DeLong testing with a correction for multiple comparisons. It reports no statistically significant differences in those comparisons. That is a useful research analysis, but absence of statistical significance is not proof of equivalent performance in every group.

The study is not mapped to the current deployed model version, and ongoing production fairness monitoring or release acceptance criteria remain unspecified. Consequently, the research finding cannot establish the governance of every current prediction function. The available assurance is more limited than a version-linked evaluation of the production service.

The health-plan offering presents a minimum same-year return guarantee of three times investment. That is a falsifiable commercial commitment rather than a general promise to improve outcomes. Its calculation, covered costs, exclusions, and remedy if the result is missed remain unspecified. It also concerns program economics rather than guaranteed accuracy of an individual pregnancy prediction.

Consumer terms separately limit platform liability and require individual dispute resolution; they do not establish the terms negotiated with a health plan. The available commitment therefore has substance, but does not supply a defined output-error indemnity or an explicit reimbursement procedure.

An external 2024 technical review describes an Epic-embedded visualization tool for patient measurements, including blood pressure, glucose, and weight. The current provider workflow also brings patient inputs and risk information into a dashboard for the obstetric team. This is a named clinical-system connection rather than only a generic integration claim. Current production coverage, supported versions, writeback behavior, and the scope of automatic exchange remain unclear.

Reviewing monitoring information inside an EHR does not necessarily mean that every generated prediction or care-plan change writes back to the source record. The strongest evidence concerns embedded access to patient monitoring data.

Patients use a mobile or web app, while clinical teams access a hosted dashboard. The privacy policy allows information within its scope to move to company and service-provider offices or servers in the United States and other countries. Secure clinical information is governed separately through the medical-group framework. Specific hosting regions, tenant isolation, backup locations, and the geography of clinical model inference remain unspecified, leaving the complete service residency boundary unclear.

Participating practices are offered the program at $0, with funding potentially supplied by health plans, grants, or government sources. The arrangement is developed around jointly defined goals and the funding available for the practice's pregnant patients. This is a concrete purchasing basis for a specific participant group rather than a universal free-service claim.

Health plan contract prices, funding eligibility, implementation charges, and the financial terms for other organizational buyers remain unspecified. The no-cost practice arrangement also does not establish the cost borne by the funding organization. Program services and postpartum support can vary with the payer agreement.

The service focuses on prenatal and postpartum care, combining obstetric practices with health plan and community support. Named practice experience includes A&A Women's Health, Center for Women's Health in Houston, and Mayo Clinic Mankato. The program can continue for up to 12 months postpartum and includes monitoring, education, and several support services. A weight-gain study spans four community clinics, giving a concrete research setting.

However, its exclusions limit conclusions for preterm births and multiple gestations. Evidence from those clinics also does not establish equivalent performance for every complication or the full range of payer, employer, and health-system populations targeted by the service.