Research

If you have a question about our research, why not just ask?

Prototype Lead

Project Team

Noah, medical student and AlloPatient project lead.

Noah Kreiling

Co-Founder

MD Candidate, Drexel University College of Medicine

Kabilan Balasubramani

Co-Founder

Software Engineer

Measurement Model

What AlloPatient Measures Today

AlloPatient is organized around physician-vetted vignettes that define the clinical scenario, expected information, scoring criteria, safety issues, and AI-use risks before learners begin.

Those vignettes can be assigned to controlled learner cohorts, producing structured assessment attempts that capture both final clinical output and the process used to get there.

Vignette Builder

Clinician-reviewed case + rubric

Press to flip

Vignette Builder

A physician-vetted vignette defines the clinical scenario, expected information, scoring criteria, safety issues, and AI-use risks before learners begin.

Press to return

Controlled Cohort

Assigned learners or study groups

Press to flip

Controlled Cohort

The same vignette can be assigned to a defined learner group or study condition, making attempts comparable across students, cohorts, or training arms.

Press to return

Assessment Attempt

Patient, chart, and AI interaction

Press to flip

Assessment Attempt

During the attempt, AlloPatient captures how learners interview the patient, use the chart, query AI, revise work, and respond to safety concerns.

Press to return

Structured Evidence

Process data + final clinical output

Press to flip

Structured Evidence

Each attempt produces reviewable evidence from both the final clinical answer and the learner’s reasoning process, AI-use behavior, and calibration inputs.

Press to return

Research Insights

Cohort-level clinical AI readiness patterns

Press to flip

Research Insights

Across repeated attempts, these records can reveal cohort-level patterns in clinical reasoning, AI reliance, verification behavior, and readiness for AI-assisted care.

Press to return

Clinical AI Readiness Profile

Five scored domains, one weighted composite

The composite score is a weighted average of these domain scores. It is one layer of the research model: the broader product links vignette design, controlled cohort attempts, and reviewable process/output evidence for cohort-level analysis.

Critical Information

25%

Weighted clinical facts the learner elicits or surfaces through the patient interview, chart review, final note, and supporting evidence.

Safe AI Use

25%

A deduction-based score for privacy-aware prompting, supervising AI output, correcting seeded AI errors, and completing disclosure/supervision steps.

Final Clinical Quality

30%

Weighted rubric items in the final note and plan, including diagnosis, management, medications, contraindications, safety netting, and follow-up.

Efficiency

10%

Time and interaction-count signals, including AI queries, chart checks, patient questions, irrelevant questions, and post-review revisions.

Calibration

10%

Learner confidence ratings compared with measured performance across diagnosis, management, safety, and AI-use domains.

In practice, the profile separates what the learner produced from how they got there. Critical Information and Final Clinical Quality reflect clinical performance; Safe AI Use reflects supervision, privacy, disclosure, and correction behavior; Efficiency captures workflow patterns; and Calibration compares confidence with measured performance. This makes the score more interpretable than a single pass/fail result.

From Vignette Attempts to Research Data

Each assessment attempt connects a physician-vetted vignette to a learner's workflow and final output. When the same vignette is assigned across a cohort, these structured records can support research into clinical reasoning, AI reliance, verification behavior, documentation quality, and learner calibration.

Research layerWhat it captures

Vignette configuration

Physician-vetted scenario design, expected facts, rubric items, safety flags, assessment goals, and anticipated AI-use risks.

Learner workflow

Patient questions, chart review, AI queries, revisions, safety-review behavior, and other steps taken during the attempt.

Final clinical output

Diagnosis, management plan, documentation quality, safety-netting, follow-up, and patient-facing explanation quality.

AI-use behavior

Appropriate AI use, output verification, error correction, privacy and minimum-necessary disclosure concerns, and supervision/disclosure.

Confidence and calibration

Learner confidence compared with measured performance across diagnosis, management, safety, and AI use.

Cohort comparison

Performance patterns across learners, training groups, AI-access conditions, or repeated pilot cohorts assigned the same vignette.

Research Status

AlloPatient is designed to make learner reasoning and AI use easier to study across structured vignette attempts. Its current role is to support disciplined pilot work, not to overclaim validation.

AlloPatient is an early-stage clinical AI readiness prototype, not a fully validated assessment instrument.

The current research goal is to support structured pilot studies, clinician review, and iterative validity evidence.

Some criteria may be supported by AI-assisted judges, but the research model emphasizes transparent vignette design and reviewable learner evidence.

AlloPatient is currently in an early prototype stage. Clinicians, educators, students, and researchers interested in providing feedback or collaborating on pilot studies can contact: hello@allopatient.com