Published at ACM Interactive Health 2026 (IH '26) · Porto, Portugal · REALIZE Lab, University of Michigan
PATHFinder
An LLM Agent for Tailored Prenatal Care
PATHFinder Agent turns the ACOG PATH tailored prenatal care guidelines into an individualized plan for every pregnancy — gathering medical history and social context through conversation, matching unmet needs to Michigan 211 community resources, and right-sizing visit frequency, testing, and telehealth, all under clinician oversight.
Abstract
Prenatal care is an important preventive service designed to improve outcomes for pregnant individuals. The American College of Obstetricians and Gynecologists (ACOG) recently introduced guidelines advocating tailored prenatal care, called PATH (Plan for Tailored Healthcare). We present PATHFinder Agent (Planner for Appropriate Tailored Healthcare), an end-to-end conversational agentic system that gathers patient health and social context through structured dialogue, curates individualized prenatal care plans aligned with PATH guidelines, and surfaces community resources from Michigan 211. The system features a four-stage workflow spanning patient intake, dynamic interaction, plan synthesis, and clinician oversight. We evaluate frontier large language models (LLMs) on expert-curated rubrics across five clinical dimensions, finding that GPT-5.2 achieves the highest average score (77.6%) while identifying key gaps in antenatal testing recommendations. We discuss future validation through human participant studies and randomized controlled trials.
Why prenatal care needs to be tailored
Prenatal care is the preventive care provided to pregnant and birthing people — medical visits, screening and antenatal testing, answers to questions, and help finding social support and community resources during pregnancy.
"Prenatal care can improve the detection and management of chronic conditions and pregnancy complications, particularly for individuals with an increased risk of adverse outcomes." — ACOG
Nearly four million pregnant people receive prenatal care each year in the United States, and over 80% of adverse pregnancy outcomes are preventable through essential prenatal care and management of unmet social needs. Yet the traditional 12–14 visit in-person schedule — largely unchanged for a century — is often inaccessible because of social drivers of health such as housing instability, limited transportation, and inflexible work schedules. Many patients end up underserved and feel unheard.
In 2025, ACOG issued Clinical Consensus No. 8 on tailored prenatal care delivery, built on the PATH (Plan for Appropriate Tailored Healthcare in Pregnancy) framework developed with the University of Michigan. PATH tackles two things at once: addressing unmet social needs and incorporating alternative care modalities such as telehealth and group care. But adopting PATH is a complex task for patients and clinicians alike — extra visits for social needs create burden without addressing root causes, and increased demands on clinicians to handle unmet social needs with insufficient resources are associated with burnout.
The PATHFinder approach
PATHFinder Agent was co-designed by a board-certified obstetrician-gynecologist and a certified nurse midwife alongside two computer scientists, so the clinical and technical requirements were validated jointly. It centralizes what is today a set of disjoint conversations into one grounded, guideline-aware dialogue — while keeping clinicians in the loop, so that tailored prenatal care planning is something every stakeholder can trust and clinician effort can be redirected toward care rather than paperwork.
The agent has three core abilities: (a) follow up with questions and clarifications requiring medical and conversational knowledge, (b) perform actions via tool calls — searching for community resources, generating reports, running clinical calculators — and (c) follow instructions to produce a comprehensive, individualized plan.
Under the hood
- 13 tools across 4 categories. Medical (a TOLAC calculator for trial of labor after cesarean; a structured "defer to clinician" referral), social needs resources (hierarchical Michigan 211 queries by group, category, subcategory, and ZIP code), personalization (a secondary LLM that generates context-aware follow-up questions), and reporting (patient summary, clinical summary, recommendations, resources, visit schedule).
- Grounded in guidelines. System instructions total roughly 14,000 tokens of domain knowledge drawn from the ACOG PATH guidelines, plus tool-use instructions, workflow rules, and safety policies that the agent retrieves as it works.
- Real community resource data. Social needs are matched against Michigan 211 data organized by category — food, housing, transportation, utilities, clothing — and queried by subcategory and ZIP code through dedicated tool calls.
- Built for oversight. A React frontend talks to a FastAPI backend routing to the agent (LLM plus tool executor), an oversight classifier, an FHIR service, and a report generator. Conversation state is persisted in a relational database so clinicians can audit what was asked and why.
How it works: the four-stage workflow
PATHFinder follows a four-stage workflow — the same flow you can step through in the interactive demo.
- Stage 1 — Patient intake. The patient completes a standardized form that mirrors existing clinical intake processes: demographics, medical history, gestational age from last menstrual period, and social factors. Where available, EHR data is imported rather than re-asked.
- Stage 2 — Dynamic interaction. Conditioned on those responses and on the guidelines, the agent asks personalized, non-redundant follow-up questions to elicit unmet social needs, preferences, and barriers. A second LLM reviews the questions and generates a form-based interface with buttons, check boxes, and sliders — reducing how much the patient has to type and preventing fatigue in long sessions.
- Stage 3 — Plan synthesis and patient review. The agent invokes its report tools to compose a patient-facing summary, a clinical summary, a visit schedule, and curated Michigan 211 resources. The patient reads the draft and can ask clarifying questions — "What does GTT mean?" — and get answers in place.
- Stage 4 — Clinician oversight. The clinician reviews the patient summary, report, and timeline on a dashboard showing conversation stage, risk flags, and escalation controls, then approves or edits before the plan is used in care.
Evaluation: how well do frontier LLMs plan prenatal care?
We evaluate PATHFinder Agent on synthetic patient profiles spanning diverse medical histories — including high-risk pregnancies and TOLAC candidates. Each profile is paired with an expert-curated rubric specifying the correct expectations along five clinical dimensions: visit frequency, services (testing and recommendations), antenatal testing (for example, timing of growth ultrasound), delivery timing, and modality of care (mandatory in-person, mixed in-person, group health). We use LLM-as-judge scoring, with each final score normalized to 1.
| Model | Avg. rubric score |
|---|---|
| GPT-5.2 | 77.60% |
| Gemini 2.5 Pro | 71.50% |
| Gemini 2.5 Flash | 62.00% |
| GPT-4o | 57.25% |
What this means for deployment. Even the strongest model leaves roughly a fifth of the rubric unmet, and the gaps cluster in exactly the recommendations that matter clinically — antenatal testing and services. These results argue that robust oversight measures, both LLM-driven and human, must be built into agentic health systems with appropriate communication so that deployed models do not make mistakes. That is why clinician review is a stage of the workflow rather than an optional add-on.
Frequently asked questions
What is PATHFinder Agent?
PATHFinder Agent (Planner for Appropriate Tailored Healthcare) is an end-to-end conversational AI agent, built on large language models, that produces individualized prenatal care plans. It collects a patient's medical history and social context through structured dialogue, applies the ACOG PATH tailored prenatal care guidelines, matches identified social needs to Michigan 211 community resources, and routes every plan through a clinician for review before it is used in care.
What are the ACOG PATH guidelines for tailored prenatal care?
PATH stands for Plan for Appropriate Tailored Healthcare in Pregnancy. Developed by an independent expert panel convened by the American College of Obstetricians and Gynecologists and the University of Michigan, PATH replaces the century-old fixed 12–14 visit schedule with a plan built around each patient's medical needs, social and structural risk factors, and preferences. It recommends a core set of in-person visits plus flexible use of telehealth and group care. PATHFinder Agent operationalizes these guidelines in software.
How does PATHFinder keep clinicians in the loop?
PATHFinder never delivers a plan to a patient without oversight. Its fourth workflow stage is a clinician review dashboard showing the patient summary, the clinical summary, the proposed visit timeline, risk flags, and escalation controls. The clinician can approve or edit any recommendation. A separate oversight classifier and a dedicated "defer to clinician" tool guard against unintended agent behavior, and full conversation state is persisted so decisions can be audited.
How does PATHFinder address social determinants of health?
The intake and dialogue stages capture unmet social needs — housing instability, limited transportation, food insecurity, inflexible work schedules, language access, and support systems. The agent then queries a hierarchical Michigan 211 resource index by category, subcategory, and ZIP code through dedicated tool calls, producing a resource assistance plan alongside the medical plan. Social needs are treated as first-class inputs to the care plan rather than an afterthought.
Which large language models were evaluated, and how well did they do?
Frontier models were scored against expert-curated rubrics on five clinical dimensions. GPT-5.2 scored highest at 77.60%, followed by Gemini 2.5 Pro at 71.50%, Gemini 2.5 Flash at 62.00%, and GPT-4o at 57.25%. Visit frequency was the easiest dimension for every model; antenatal testing and service recommendations showed the widest performance gaps — which is precisely why human and automated oversight remain necessary.
Is PATHFinder safe to deploy in a clinic today?
Not yet as an autonomous system. The published evaluation uses synthetic patient profiles and LLM-as-judge rubric scoring, and the results show meaningful gaps in antenatal testing recommendations. We are actively working on extensive evaluation and human-AI studies, and the design assumes clinician review of every plan. For more, contact vballoli@umich.edu.
Cite this paper
ACM Reference Format
Vaibhav Balloli, Carissa Samuel, Samia Abdelnabi, Alex Peahl, and Elizabeth Bondi-Kelly. 2026. PATHFinder Agent for Tailored Prenatal Care. In Interactive Health Conference (IH '26), July 05–08, 2026, Porto, Portugal. Association for Computing Machinery, New York, NY, USA, Article 75, 5 pages. https://doi.org/10.1145/3786579.3804996
BibTeX
@inproceedings{balloli2026pathfinder,
author = {Balloli, Vaibhav and Samuel, Carissa and Abdelnabi, Samia and Peahl, Alex and Bondi-Kelly, Elizabeth},
title = {PATHFinder Agent for Tailored Prenatal Care},
year = {2026},
isbn = {979-8-4007-2422-0},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
url = {https://doi.org/10.1145/3786579.3804996},
doi = {10.1145/3786579.3804996},
booktitle = {Proceedings of the 2026 ACM Interactive Health Conference},
articleno = {75},
numpages = {5},
keywords = {Large Language Models, Large Language Model Agents, Health,
Reproductive Health, Maternal Health, Healthcare Agents,
LLM-in-the-loop},
location = {Porto, Portugal},
series = {IH '26}
}
Team
Authors: Vaibhav Balloli, Carissa Samuel, Samia Abdelnabi, Alex Peahl, and Elizabeth Bondi-Kelly · REALIZE Lab, University of Michigan, Ann Arbor, USA.
Contact: vballoli@umich.edu · ecbk@umich.edu
Acknowledgements: This work was partially supported by funding from Google and the University of Michigan, including the Raoul Wallenberg Institute, E-Health and Artificial Intelligence, and the Center for Academic Innovation.
Personal Information
PATHFinder Follow-up
V B
Draft Report Review
V B
Your Personalized Care Report
Ask About Your Report
Shared Review — Clinician & Patient
Collaborative visit