See us at AAFP 2026 HIPAA compliant · BAA available Works with Athena, eClinicalWorks, Epic & more
Case Studies

Case Study

Primary adrenal insufficiency presenting with fatigue and constitutional symptoms

A synthetic case run through three clinical AI tools — ChatGPT, OpenEvidence, and Mari — to evaluate differential output from the same patient presentation. Two tools were given the patient's self-reported narrative. Mari completed a structured interview before generating a differential.

Synthetic case Physician-verified Primary care Differential diagnosis Autoimmune

The patient

A 42-year-old woman with an evolving picture.

The case was designed with layered complexity: a plausible primary complaint, prior diagnoses that provide reasonable cover for the symptoms, and a family history that — with the right questions — points clearly toward the correct diagnosis.

Clinical presentation

42-year-old woman with Hashimoto's thyroiditis, vitiligo, and chronic iron-deficiency anemia presenting with nine months of progressive fatigue, night sweats attributed to perimenopause, nausea, and occasional mild abdominal discomfort. She was started on escitalopram for presumed depression with no improvement. She has lost 8 lbs, which she attributes to decreased appetite from nausea. Periods have become more irregular.

Family history: Mother with type 1 diabetes. Brother with celiac disease.

Duration
9 months progressive fatigue
Attributed to
Perimenopause · Depression · Escitalopram
Weight
8 lb unintentional loss
Prior diagnoses
Hashimoto's · Vitiligo · Iron-deficiency anemia
Family history
T1DM (mother) · Celiac (brother)
Current treatment
Escitalopram — no improvement
Symptom timeline · 9 months
Onset: fatigue, night sweats
Escitalopram started: no improvement
Weight loss, nausea onset
Presents to clinic
Month 0Month 9

Method

Differential output across three tools

ChatGPT and OpenEvidence were given the presenting complaint and background in natural language, as the patient would share it. Mari's structured interview was completed by a simulated patient as would occur before a real appointment.

ChatGPT
LeadingCeliac disease
ConsiderIron-deficiency anemia
ConsiderPerimenopause
ConsiderDepression / SSRI effect
ConsiderThyroid dysfunction
Input: patient narrative as shared by the patient
OpenEvidence
LeadingCeliac disease
ConsiderIron-deficiency anemia
ConsiderPerimenopause
ConsiderHypothyroidism
ConsiderDepression
Input: patient narrative as shared by the patient
Mari
LeadingPrimary adrenal insufficiency (Addison's disease)
ConsiderCeliac disease
ConsiderIron-deficiency anemia
ConsiderPerimenopause
ConsiderAutoimmune polyglandular syndrome
Input: structured patient interview, physician-designed pathways

Both ChatGPT and OpenEvidence returned clinically reasonable differentials. The correct diagnosis was not among them. Mari's differential was verified as correct by physicians and clinical AI tools.

History-taking

Findings surfaced by structured interview

The structured interview surfaced three findings that did not appear in the patient's self-reported narrative. Each is clinically significant for primary adrenal insufficiency.

Skin darkening

Hyperpigmentation — particularly of skin folds, scars, and mucous membranes — is a hallmark of primary adrenal insufficiency. Not volunteered. Asked directly.

Orthostatic symptoms

Lightheadedness on standing indicates postural hypotension — a direct consequence of cortisol deficiency affecting vascular tone. Not volunteered. Asked directly.

Weight loss in context

The patient attributed her 8 lb loss to nausea. Mari reframed it alongside decreased appetite and fatigue together — a pattern consistent with adrenal insufficiency, not GI illness.

Mari interview screenshot — targeted question Mari asking about skin darkening or orthostatic symptoms. Anonymized real product UI, patient-portal mobile frame.

Why these questions point to Addison's disease

The patient already had two autoimmune conditions — Hashimoto's thyroiditis and vitiligo. Her mother had type 1 diabetes. Her brother had celiac disease. Autoimmune polyglandular syndrome clusters these conditions: a patient with two autoimmune diagnoses and a family history of others carries meaningful risk for primary adrenal insufficiency. The structured interview is designed to surface this pattern through targeted questioning.

Discussion

History completeness and differential accuracy

ChatGPT and OpenEvidence returned clinically reasonable differentials from the information they were given. The input in both cases was a patient's self-reported narrative — which is characteristically incomplete. Patients contextualize their symptoms through prior diagnoses, family patterns, and plausible explanations. In this case: perimenopause, escitalopram side effects, and a brother with celiac disease. That framing is understandable, and it shaped what each tool returned.

The three findings that changed the differential — skin darkening, orthostatic symptoms, and the weight loss pattern — were not volunteered. They were present. Eliciting them required targeted questions directed at the specific diagnostic possibilities implied by the patient's autoimmune history and family history.

"The difference was not a better algorithm. It was a better history."

<30%
of women with primary adrenal insufficiency are diagnosed within the first 6 months of symptoms
The median diagnostic delay for Addison's disease is measured in years — not because physicians miss it, but because the history that points to it is rarely complete when it first presents. This case was designed to replicate exactly that scenario.
Mari physician portal — case output Full differential: Addison's disease as leading dx, Do-Not-Miss flag, and key history details. Label: "42F · Synthetic case" — not real patient data.

Methodology

How this case was run

Case type

This is a synthetic test case — not a real patient. The presenting scenario was constructed to test Mari's history-taking against published clinical benchmarks for primary adrenal insufficiency. No real patient data was used.

Comparison methodology

ChatGPT and OpenEvidence were each given the patient's presenting complaint and background in natural language, as a patient would share it in a clinical encounter or online query. Mari's structured interview was completed by a simulated patient responding to Mari's questions in the patient portal — as would occur before a real appointment.

Verification

Mari's output — with primary adrenal insufficiency as the leading diagnosis — was reviewed and confirmed as correct by physicians and by the same clinical AI tools used in the comparison. This constitutes structured evaluation with physician verification. It is not a clinical trial or peer-reviewed study.

Evidence level

Level 3 — early quantitative evidence from a structured single-case evaluation with physician verification. Hekavor is building toward larger-scale clinical validation. This case is one data point, not a population study.

Try Mari free for 60 days.

No EHR required. No long-term commitment.

Start your free trial Email us first

Trial link and onboarding guide delivered within 24 hours.