ekacare/ekacare_medical_history_summarisation

Medical History Summarization Dataset

This dataset contains 58 diverse medical cases from EkaCare's internal medical team and the company's employees and their family members. Each case provides a complete view of a patient's medical journey, including vital sign trends and historical health context. The dataset has been designed to evaluate the ability to summarize complex medical histories and extract clinically relevant insights. For details read this blog.

Dataset Overview

Each case represents a real-world medical scenario with comprehensive patient information spanning multiple visits, vital signs, and medical history. The primary task is to generate concise, clinically relevant summaries that highlight the most important information from the patient's in most recent six months of their medical journey.

Task Description

Objective: Generate comprehensive medical summaries that capture:

• Key clinical developments from the most recent 6 months
• Significant historical medical events that impact current care
• Critical patterns in symptoms, diagnoses, and treatments
• Important changes in patient condition or medication regimens

Evaluation Framework

Rubric-Based Assessment

Following the methodology established by Arora et al. (2025) in HealthBench, this dataset employs objective evaluation rubrics created by medical professionals. Each case includes multiple evaluation rubric that assess multiple aspects of the summary. The rubrics are meant to ensure the evaluation does not need any subjective decision, any information required to evaluate the rubric is self-contained.

Rubric Structure

Each rubric contains:

• Criterion: Specific aspect being evaluated
• Points: Scoring weight - in this case it's either 1 or 0
• Evaluation Logic: Average of the points from the given rubrics per case

Data Structure

The dataset contains the following key components:

Core Fields

• ideal_completions_data: Expert-generated reference summaries for evaluation, with an emphasis tag on conditions that need to be highlighted.
• rubrics: Objective evaluation criteria developed basis summary curated by medical professionals.
• prompt_id: Unique identifier for each medical case.
• prompt: Complete patient information requiring summarization.

Patient Information Structure

Each case includes comprehensive medical data:

{
  "demographic_details": {
    "gender": "Patient gender",
    "dob": "Date of birth",
  },
  "visits": [
    {
      "date": "Visit date",
      "diagnosis": "Primary and secondary diagnoses",
      "symptoms": "Reported symptoms",
      "medications": "Prescribed medications",
      "notes": "Clinical notes and observations",
      "lab_tests": "Ordered laboratory tests",
      "vital_signs": "Recorded vital measurements"
    }
  ],
  "vitals": [
    {
      "date": "Measurement date",
      "type": "Vital sign type",
      "value": "Recorded value",
      "unit": "Measurement unit"
    }
  ],
  "medical_history": {
    "past_conditions": "Previous medical conditions",
    "allergies": "Food and drug allergies",
    "current_medications": "Ongoing treatments",
    "family_history": "Relevant family medical history",
    "lifestyle": "Smoking, alcohol, exercise habits",
    "procedures": "Past surgical procedures",
    "vaccinations": "Immunization history"
  },
  "doctor_specialisation": "The treating doctor's specalisation"
}

From the above schema any of the keys could be missing depending on the case.

Performance Benchmarks

Model	Rubric average
OpenAI - o3	67.8%
OpenAI - GPT-4o	66.4%
OpenAI - GPT-4o-mini	51.8%
Anthropic - Bedrock - Sonnet V4	74.6%
DocAssist Chat	72.2%

DocAssist Chat is EkaCare's clinical AI assistant for doctors.

Use Cases

This dataset is valuable for:

• Medical AI Development: Training and evaluating models for clinical summarization
• Healthcare Technology: Developing tools for medical record processing
• Clinical Decision Support: Creating systems that highlight important patient information
• Medical Education: Teaching effective medical summarization techniques
• Research: Studying AI performance in medical text understanding

Key Challenges

The dataset addresses several important challenges in medical AI:

• Information Synthesis: Combining data from multiple sources and time points
• Clinical Prioritization: Identifying the most relevant information for patient care
• Temporal Understanding: Tracking disease progression and treatment responses
• Medical Terminology: Accurate handling of complex medical language
• Safety Considerations: Ensuring summaries don't omit critical information

Dataset Statistics

• Total Cases: 58 diverse medical scenarios
• Language: English
• License: MIT

Evaluation Metrics

Models can be evaluated using:

• Rubric Scores: Objective assessment based on medical professional criteria

Getting Started

1. Load the Dataset: Import the JSON structure containing all patient cases
2. Understand the Format: Review the data structure and example cases
3. Implement Summarization: Develop your approach to generate medical summaries
4. Evaluate Results: Use the provided rubrics to assess summary quality

List of Doctors Curated the dataset

• Dr Anushree Rana
• Dr Arun Kumar R
• Dr Sanjana SN

License

This dataset is released under the MIT License, enabling broad use while maintaining attribution requirements.