PATHFinder Agent for Tailored Prenatal Care

Technical Summary: PATHFinder Agent for Tailored Prenatal Care

Problem Statement
Prenatal care is a critical preventive service, yet the traditional model of 12–14 in-person visits often fails to address the social drivers of health (e.g., housing instability, inflexible work) that prevent patients from accessing care. Furthermore, the American College of Obstetricians and Gynecologists (ACOG) has introduced new guidelines, the Plan for Appropriate Tailored Healthcare (PATH), which advocates for flexible, patient-specific care plans. However, implementing PATH is complex for clinicians and patients alike; clinicians face burnout from addressing unmet social needs with insufficient resources, while patients with social barriers may find additional visits burdensome rather than helpful. There is a need for a system that can operationalize these tailored guidelines, integrate social resource matching, and maintain clinical oversight without overburdening providers.

Methodology and System Design
The authors propose PATHFinder Agent, an end-to-end conversational agentic system designed to gather patient context, synthesize individualized care plans, and surface community resources. The system is co-designed by board-certified obstetricians, midwives, and computer scientists to ensure clinical validity and technical feasibility.

• Architecture: The system utilizes a React frontend communicating with a FastAPI backend. The core agent consists of a Large Language Model (LLM) paired with a tool executor, an oversight classifier, a FHIR service, and a report generator. Conversation state is persisted in a relational database.
• Tool Suite: The agent is equipped with 13 tools across four categories:
  • Medical: Includes a TOLAC (Trial of Labor After Cesarean) calculator and a "Defer to Clinician" action.
  • Social Needs: A hierarchical interface to query Michigan 211 resources (food, housing, transportation, etc.) by subcategory and ZIP code.
  • Personalization: Consults a secondary LLM to generate context-aware follow-up questions.
  • Reporting: Incrementally builds patient summaries, clinical summaries, visit schedules, and resource lists.
• Four-Stage Workflow:
  1. Intake: Patients complete a standardized form capturing demographics, medical history, gestational age, and social factors.
  2. Dynamic Interaction: The agent engages in structured, open-ended dialogue. It uses a "Form+Chat" interface where a secondary LLM generates intermediate UI elements (buttons, checkboxes) to reduce text entry fatigue while eliciting unmet social needs and preferences.
  3. Plan Synthesis: The agent generates a draft report containing a patient-facing summary, a clinical summary, visit schedule recommendations, and curated Michigan 211 resources.
  4. Clinician Oversight: A clinician reviews the draft, validates the plan, and can approve or edit the content before finalization.

Evaluation and Results
The authors evaluated frontier LLMs (GPT-5.2, GPT-4o, Gemini 2.5 Pro, Gemini 2.5 Flash) using synthetic patient profiles and expert-curated rubrics. The evaluation focused on five clinical dimensions:

1. Visit frequency
2. Services (testing, recommendations)
3. Timing for antenatal testing
4. Timing for growth ultrasound
5. Modality of care (in-person vs. telehealth)

Key Findings:

• Performance: GPT-5.2 achieved the highest average rubric score at 77.6%, followed by Gemini 2.5 Pro (71.5%), Gemini 2.5 Flash (62.0%), and GPT-4o (57.25%).
• Dimensional Variance: "Visit frequency" was the easiest dimension for all models to predict correctly. Conversely, "antenatal testing" and "services recommendation" were the most difficult, showing the widest performance gaps across models.
• Implications: The results indicate that while current models show promise, robust oversight measures (both LLM-based and human-driven) are essential to prevent errors in critical clinical recommendations.

Key Contributions

1. PATHFinder Agent: A novel agentic system that operationalizes ACOG's PATH guidelines, bridging the gap between medical guidelines and social resource integration.
2. Hybrid Interface Design: A "Form+Chat" interaction model that balances structured data intake with adaptive conversational follow-ups, designed to mitigate patient fatigue.
3. Clinician-in-the-Loop Workflow: A structured four-stage process that ensures all generated care plans undergo human validation, addressing safety and deployment concerns raised in recent medical AI literature.
4. Empirical Benchmarking: A comparative evaluation of state-of-the-art LLMs on specific prenatal care dimensions, highlighting current capabilities and limitations in complex medical reasoning.

Significance and Claims
The paper positions PATHFinder Agent as a potential enabler for the broad deployment of tailored prenatal care guidelines. The authors claim the system can support the shift from a rigid, visit-heavy model to a flexible, patient-centered approach that addresses both medical and social determinants of health.

The authors maintain a modest stance regarding deployment. They explicitly state that the system is designed to support, not replace, clinicians. The paper emphasizes that future work must include formal accuracy guarantees, human participant studies, and randomized controlled trials before the system can be considered for real-world clinical deployment. The current contribution is framed as a technical prototype and evaluation framework that identifies key gaps (particularly in antenatal testing recommendations) that must be addressed to ensure safety and efficacy.