Augmenting Clinical Workflow With a Structured Chatbot for Patient History Collection and Diagnostic Support (CHAT-HX)

Prospective Observational Evaluation of a Structured, Agentic AI Chatbot for Pre-Visit Patient History Collection and Diagnostic Support in Outpatient Clinical Workflows

Doctors spend much of each visit collecting the patient's medical history, and incomplete histories can delay diagnosis. This study tests an AI chatbot that interviews patients before their appointment and produces a structured summary of their history for the doctor. Adult patients at the first study site, a university gynecology outpatient clinic in Germany, answer the chatbot's questions on a tablet before the doctor visit. The doctor then takes the history and examines the patient as usual, without seeing the chatbot's output. Only after the doctor has finished planning care does the doctor review the chatbot's summary and rate its usefulness and accuracy. The chatbot does not change patient care. Patients also rate the ease of use and acceptability of the chatbot. The investigators will report how useful doctors find the summaries, how complete and correct the summaries are, and how patients experience the tool.

Study Overview

• Status: Not yet recruiting
• Conditions: Medical History Taking; Artificial Intelligence in Clinical Workflow; Outpatient Clinical Encounter
• Intervention / Treatment: Other: AI-Based Structured History-Taking Chatbot

Detailed Description

This is a prospective, observational study of an AI-based, structured history-taking tool used before a clinical visit. The tool uses large language models and follows a fixed interview plan with defined topic blocks, such as chief complaint, current history, past conditions, medications, allergies, specialty-specific history, and social and family history. It produces a structured history summary, a list of possible diagnoses, and suggested workup steps. In earlier benchmarking on more than 670 simulated and clinical cases, the system captured most clinically relevant history items and flagged clinically important warning signs.

The study is designed so that the tool does not influence care. Participants give written informed consent, then complete the AI interview on a study tablet using a pseudonymous study ID. Participants enter no name, birth date, or address. Participants then complete a short questionnaire on usability and acceptance. The treating physician takes a full history, examines the patient, and documents a plan without access to the AI output. Only afterward does the physician receive the AI output and complete a structured evaluation of its completeness, correctness, plausibility, and usefulness. No follow-up contact occurs after the visit.

The study is not randomized and has no control arm. The physician's independently obtained history serves as the reference standard. The first site is the Department of Gynecology at Leipzig University Hospital, with about 100 patients and 8 to 15 physicians. Analysis is descriptive, with subgroup comparisons by reason for visit. Additional sites may be added later. The study is a collaboration between Leipzig University Hospital and Massachusetts General Hospital and Harvard Medical School.

• Study Type: Observational
• Enrollment (Estimated): 100

Contacts and Locations

• Study Contact: Name: Sonu Subudhi, MBBS, PhD; Phone Number: 617-726-8051; Email: ssubudhi@mgh.harvard.edu

Study Locations

• Germany
  • Leipzig, Germany, 04103
    • Universität Leipzig
    • Contact: Bahriye Aktas; Phone Number: +49 341 97 23400; Email: bahriye.aktas@medizin.uni-leipzig.de

Participation Criteria

Eligibility Criteria

• Ages Eligible for Study: 18 years and older (Adult, Older Adult)
• Accepts Healthy Volunteers: No
• Genders Eligible for Study: Female

• Sampling Method: Non-Probability Sample

Study Population

Adult women presenting for the first time, or with a new complaint, to the general gynecology outpatient clinic, the gynecology emergency clinic, or as referrals to the gynecology outpatient clinic at the first study site (Leipzig University Hospital, Germany). Patients are recruited consecutively in the clinic waiting area by trained study staff. Additional sites may be added later.

Description

Inclusion Criteria:

Women 18 years or older

First presentation or new complaint at the gynecology outpatient clinic (self-presentation, emergency, or referral)

Sufficient German language skills to complete a structured patient interview independently

Able to complete study participation without external support

Written informed consent

Exclusion Criteria:

Acute emergency requiring immediate treatment (for example, impaired consciousness, hemodynamic instability, or acute bleeding requiring immediate care)

Lack of legal or decisional capacity to consent

Insufficient language skills

Severe cognitive, visual, or motor impairment preventing tablet use

Previous participation in this study

Refusal or withdrawal of consent

Study Plan

How is the study designed?

Design Details

• Observational Models: COHORT
• Time Perspectives: PROSPECTIVE

Number of groups / cohorts

1

Cohorts and Interventions

Group / Cohort

Intervention / Treatment

Adult Clinic Patients; Adult patients with a first presentation or new complaint at the participating outpatient clinic who consent to complete the AI-based pre-visit history interview and questionnaire. Their physician completes routine care blinded to the AI output, then evaluates the AI-generated summary.

Other: AI-Based Structured History-Taking Chatbot; A large language model-based agentic chatbot that conducts a structured, pre-visit history interview on a study tablet and generates a history summary, differential diagnosis list, and workup suggestions. The output is released to the treating physician only after the physician has completed an independent history, examination, and care plan, and does not influence patient care.

What is the study measuring?

Primary Outcome Measures

Outcome Measure	Measure Description	Time Frame
Proportion of AI history summaries rated useful by the evaluating physician	After completing routine care, the treating physician rates the usefulness of the AI-generated history summary on a 5-point Likert scale (1 = not useful at all, 5 = very useful). The outcome is the proportion of summaries rated 4 or higher, reported with a 95% Wilson confidence interval.	Same day as the enrollment visit, after routine care is complete (up to approximately 1 day)

Secondary Outcome Measures

Outcome Measure	Measure Description	Time Frame
Completeness of the AI-generated history	Percentage of items from a standardized history grid that the AI captured correctly, reported overall and by history domain.	Same day as the enrollment visit, after routine care is complete (up to approximately 1 day)
Correctness of the AI-generated history	Proportion of factually incorrect or missing clinically relevant items in the AI history, per physician review.	Same day as the enrollment visit, after routine care is complete (up to approximately 1 day)
Plausibility of AI differential diagnoses	Physician rating of plausibility on a 1-5 Likert scale, and concordance of the AI differential with the physician's working diagnosis.	Same day as the enrollment visit, after routine care is complete (up to approximately 1 day)
Plausibility of AI workup recommendations	Physician rating of the AI's suggested diagnostic workup on a 1-5 Likert scale.	Same day as the enrollment visit, after routine care is complete (up to approximately 1 day)
Identification of clinically relevant red flags	Proportion of clinically relevant red flags correctly recognized by the AI versus not recognized, per physician review.	Same day as the enrollment visit, after routine care is complete (up to approximately 1 day)
Patient usability (System Usability Scale)	Patient-completed System Usability Scale, score 0-100, higher scores indicating better usability.	Same day as the enrollment visit, immediately after the AI interview or after the physician consultation (up to approximately 1 day)
Patient acceptance and willingness to use the tool again	Patient-reported acceptance, comprehensibility, empathy, comfort, trust, and willingness to use again, each on a 1-5 Likert scale.	Same day as the enrollment visit, immediately after the AI interview or after the physician consultation (up to approximately 1 day)
Duration of the AI interaction	Time in minutes for the patient to complete the AI interview, recorded automatically by the system.	During the AI interview at the enrollment visit (approximately 15 minutes)
Physician-estimated time savings	Physician's estimate, in minutes, of the time the AI history would save in clinical workflow.	Same day as the enrollment visit, after routine care is complete (up to approximately 1 day)

Collaborators and Investigators

• Sponsor: Massachusetts General Hospital
• Collaborators: University of Leipzig

Publications and helpful links

General Publications

• Steinbuch S, de Vos-Hillebrand L, O'Neill-Dee C, Wu I, Steinbuch H, Kulcsar Z, Ranjan A, Verma A, Landsberg J, Dietrich D, Hardin CC, Jain RK, Subudhi S. Comparative Performance of agentic AI and Physicians in Taking Clinical History across Leading Large Language Models (LLMs). medRxiv [Preprint]. 2026 Jan 25:2026.01.23.26344723. doi: 10.64898/2026.01.23.26344723.

Study record dates

Study Major Dates

• Study Start (Estimated): October 2, 2026
• Primary Completion (Estimated): February 28, 2027
• Study Completion (Estimated): October 1, 2027

Study Registration Dates

• First Submitted: September 30, 2026
• First Submitted That Met QC Criteria: September 30, 2026
• First Posted (Actual): October 6, 2026

Study Record Updates

• Last Update Posted (Actual): October 7, 2026
• Last Update Submitted That Met QC Criteria: October 5, 2026
• Last Verified: September 1, 2026

More Information

Terms related to this study

• Keywords: Digital Health; Usability; Chatbot; Clinical decision support; Diagnostic support; Large language model; Agentic AI; Patient History Collection; Pre-visit intake; Structured Interview
• Other Study ID Numbers: 2026P001764; 201/26-ek (Other Identifier: Ethics Committee, Medical Faculty, Leipzig University)

Plan for Individual participant data (IPD)

• Plan to Share Individual Participant Data (IPD)?: NO

IPD Plan Description

Individual participant data will not be shared with outside researchers. Participants consented under a German ethics-approved form that permits only fully anonymized, aggregated data to be shared with collaborating study partners for analysis and publication. Data are pseudonymized, are classified as health data under GDPR, and remain on servers of the clinical site. De-identified aggregate results will be reported in publications.

Drug and device information, study documents

• Studies a U.S. FDA-regulated drug product: No
• Studies a U.S. FDA-regulated device product: No