PRISM-Q (Patient Real-World Internet Search Medical Questions) Dataset

PRISM-Q is a curated dataset of 500 patient-phrased medical questions sourced from real-world internet searches. It is designed to support research in Large Language Models and clinical reasoning by capturing the authentic language and concerns of patients seeking medical information online.

Dataset Overview

The dataset is evenly divided into five clinically relevant domains (100 questions each):

• Symptom Management & Treatment
• Acute Emergency Scenarios
• Medication Safety & Drug Interactions
• Mental Health & Psychiatric Conditions
• Diagnostic Test & Laboratory Result Interpretation

Each question preserves the original, natural phrasing as typed by users (e.g., “I can’t stop worrying do I have anxiety?” or “Should I go to the ER if my child swallowed a battery?”), enabling models to learn from unfiltered, real-world patient language.

Annotation Information

Every question in PRISM-Q is annotated along the following dimensions:

Domain Category

Indicates the clinical theme or context of the question:

• Symptom Management & Treatment
• Acute Emergency Scenarios
• Medication Safety & Drug Interactions
• Mental Health & Psychiatric Conditions
• Diagnostic Test & Laboratory Result Interpretation

Clinical Area/Diagnosis

Represents the relevant clinical area or diagnosis:

• Includes both general conditions (e.g., “Abdominal Pain”, “ADHD”, “Borderline Personality Disorder”) and medical speciality

Intent Category

Defines the underlying purpose of the question:

• Symptom Interpretation – e.g., “Why do my fingers feel stiff?”
• Diagnosis Interpretation – e.g., “My doctor says I have ADHD, what does that mean?”
• Treatment – e.g., “How do I treat my spring allergies?”
• Emergency Interpretation – (event vs. symptom-specific)

Test Results (Diagnosis & Lab/Imaging Test Result Only)

• Diagnostic interpretation entries are labeled with explicit test result status:
  • Positive, Negative, Inconclusive, Normal, Abnormal

Triage Level (Emergency Scenarios Only)

Acute emergency questions are labeled using a five-point triage scale, adapted from ER protocols:

Level	Description	Examples
1	Immediate – Immediate, life-saving intervention required without delay.	Conditions like cardiac arrest; unresponsive patient; or severe respiratory distress.
2	Emergent – High risk of deterioration	Conditions and symptoms that are potentially life-threatening and require rapid treatment, such as heart attack, or stroke.
3	Urgent – Stable but requires multiple resource types (e.g., labs + imaging).	Conditions that need prompt attention but are not immediately life-threatening, like deep cuts, moderate pain, or signs of infection.
4	Semi-Urgent – Stable with only one resource anticipated (e.g., x-ray or sutures)	Conditions that are not life-threatening and can wait for treatment, such as sprains, minor cuts, or earaches.
5	Non-Urgent – Stable, with no or minimal intervention and resources anticipated	Conditions that just require oral/topical medications or prescriptions.

Drug Interaction Categorization (Medication Questions Only)

For medication safety questions, the following additional metadata is included:

• Interaction Severity

  • No Interaction, Minor, Moderate, Severe

• Pharmacological Classes

  • Each agent is mapped to its drug class (e.g., ACE inhibitor, SSRI, Beta-blocker, NSAID, Herbal supplement)

File Contents

• prism_q_dataset.csv – Main dataset with question text and annotations