This project creates a benchmark multilingual multimodal QA dataset evaluating frontline worker (FLW) focused AI assistants and agents.This benchmark QA dataset was collected from Nigeria, across 3 regions (North-central, Southwest, South-south), covering 5 languages - English, Hausa, Igbo, Pidgin, and Yoruba.

The project seeks to enhance healthcare delivery by addressing shortage of physicians in Low- and Middle-Income Countries (LMICs), especially in rural areas, by using AI-powered tools to provide reliable medical guidance to Community Health Workers (CHWs) or Community Health Extension Workers (CHEWs) who often face situations, scenarios, or challenges beyond their training and may not be able to reach a doctor in time.

Key Objectives / Goals

Dataset Creation: Create a benchmark dataset of 5,000 multilingual, situational vignette-style-questions and answers. A comprehensive and diverse dataset is essential for evaluating the performance of LLMs across various medical scenarios and languages. This dataset will serve as a critical tool for benchmarking the effectiveness of LLMs in providing medical assistance.
Benchmarking: Assess the performance of leading LLMs such as GPT-4, Claude, and Gemini in providing medical assistance. Different LLMs may exhibit varying degrees of accuracy and usefulness in medical contexts. Evaluate multiple leading LLMs to identify the most effective models for supporting CHEWs.
Dissemination: Inform the community on the practical application of LLMs in enhancing the healthcare system by sharing the findings of this project with the broader healthcare and academic community, which is crucial for driving innovation and adoption of AI technologies in healthcare. It will also help build trust and acceptance among healthcare professionals and policymakers.

Content Warning & Disclaimer

Warning This dataset contains text, audio, image, and video materials related to various health conditions, some of which may be graphic or distressing. The visual and auditory content may include depictions of medical procedures, injuries, or illnesses that could be unsettling, particularly for individuals without medical training.

Viewer Discretion Advised: The content is intended for educational and research purposes. Individuals who are sensitive to medical imagery or discussions should proceed with caution. By accessing or using this dataset, you acknowledge that you have read and understood this disclaimer and agree to use the content responsibly.

Funding/Support

Collaborating Organizations:

Usage Instructions

Accessing the Dataset: The dataset can be accessed through Hugging Face:

from datasets import load_dataset
afrispeech_dialog = load_dataset("intronhealth/afrihealth-qa-flw-nig")

AfriHealth-QA stats

Dataset size = 9,303
Total number of questions = 7289
Total number of answers = 9303
Total number of contributors = 107
Total number of doctors = 68

Contributions per regions

South West Nigeria (Oyo) 3419
North Central Nigeria (Jos) 3705
South South Nigiera (Bayelsa) 2179

Language distribution:

Language	Count
English	7246
Yoruba	964
Hausa	653
Pidgin	412
Igbo	28

Question composition

Non-English questions: 2072
Translated questions: 1487
Transcribed audio questions: 1147
Audio scenario+question descriptions: 1147
questions with images: 793
questions with videos: 6
questions with audio: 1
Number of questions with multiple answers = 1435
Largest number of answers to a single question = 12
Lowest number of answers to a single question = 1
Number of questions with a single answer = 5854

Difficulty distribution

Difficulty	Count
Easy	1835
Medium	4860
Hard	2601

Version distribution

Version	Count	Description
v1	600	preliminary analysis
v2	5699	English only
v3	3004	Multilingual / Multimodal only

Category distribution

Category	Count
Unique question categories	57
Unique category-specific scenario prompts	362

Data column descriptions

answer_id [string]: unique identifier for each answer
question_id [string]: unique identifier for each question
category [string]: Topics in CHW curriculum
SNOMED_code [int]: A unique identifier within the Systematized Nomenclature of Medicine. Each SNOMED code corresponds to a specific medical concept of the corresponding category
prompt [string]: Example conditions under each category adapted from CHW training curriculum, provided to CHWs to broaden scenario diversity
project_name [string]: data collection city/location
scenario [string]: patient situation, presentation, or context for question
scenario_len [int]: number of characters in scenario, a measure of the level of detail provided in the scenario submitted
prompt_scenario_similarity [float]: measure of similarity between scenario and scenario prompt. A quality measure that determines if scenario was sufficiently detailed, not just an adaptation of the prompt
question [string]: question or query posed by CHW to doctor or AI based on the scenario described
question_len [int]: number of characters in the question, a measure of detail
answer [string]: free text answer provided by doctor based on information available in the scenario and question
answer_len [int]: number of characters in the answer, a measure of detail/quality
difficulty [string]: a subjective measure of question difficulty [Easy, Medium, Hard] as determined by the CHW
language [string]: scenario language. Language in scenario, question, and answer may be different. Most questions and answers are in English. But scenarios are in multiple local languages.
audio_duration [float]: number of seconds of audio recording, where a question was asked via voice. Audio scenarios could be spoken in English or local languages. Most audio questions were human transcribed.
src_lang [string]: scenario language
src_text [string]: scenario text in source language
target_lang [string]: scenario target language
target_text [string]: translated scenario text, e.g. English to Yoruba
transcribed_audio_language [string]: language of audio translation or transcription
audio_transcribed_duration [float]: length in seconds of audio translations
data_audio_recording_path [string]: path to audio recording of scenario
data_audio_path [string]: path to attached audio clip for question (e.g. if a question was about a strange cough, an audio recording was included separate from the audio description of the scenario)
data_image_path [string]: path to attached image
data_video_path [string]: path to attached video
data_audio_path_transcribed [string]: path to audio translation

Downloads last month: 8

Total file size:

881 MB