FHI2020_Stratified_Sampling - Documentation

Back to Dataset View on GitHub

FHI2020 Stratified Sampling Dataset

Welcome to the FHI2020 Stratified Sampling Dataset, a comprehensive collection of Paper Analytical Devices (PADs) used in the FHI2020 and FHI2022 projects. This dataset was utilized to train some machine learning models, as listed below:

Model Name	Endpoint URL	Dataset Name	Training Dataset	Test Dataset
24fhiNN1classifyAPI	https://pad.crc.nd.edu/api/v2/neural-networks/16	FHI2020_Stratified_Sampling	metadata_dev.csv	metadata_test.csv
24 fhi NN1 API concentration	https://pad.crc.nd.edu/api/v2/neural-networks/19	FHI2020_Stratified_Sampling	metadata_dev.csv	metadata_test.csv
24fhiNN1concAPI	https://pad.crc.nd.edu/api/v2/neural-networks/17	FHI2020_Stratified_Sampling	metadata_dev.csv	metadata_test.csv
24fhiPLS1conc	https://pad.crc.nd.edu/api/v2/neural-networks/18	FHI2020_Stratified_Sampling	metadata_dev.csv	metadata_test.csv

---

Table of contents

• FHI2020 Stratified Sampling Dataset
  • Description
    • Data Distribution
      • All Dataset (8001 images)
      • Development set (5,923 images)
      • Test set (2078 images)
      • Camera and Drug Heatmap
    • Directory Structure

---

Description

This dataset draws from two projects, FHI2020 and FHI2022. From Project FHI2020, it includes images of Paper Analytical Devices (PADs) that represent specific Active Pharmaceutical Ingredients (APIs) from a list of 20 APIs, available in one of four concentrations (20%, 50%, 80%, 100%), and PADs for lactose at 100% concentration. From Project FHI2022, the dataset includes PADs categorized as swiped but not run and blank, with a nominal quantity of 0% assigned for uniformity.

The collection features multiple images of each PAD, taken under various lighting conditions and with different devices. Each image has a unique id, and all images from the same PAD (same card) share a sample_id.

To address the challenge posed by its unbalanced nature regarding API names and quantities, a stratified sampling approach was employed for splitting the data into development (dev) and test sets.

Specifically, the dataset is divided considering PADs for each API along with each quantity category, ensuring a proportional representation across both developing and testing subsets under the splitting proportion of 25% and 75%, respectively.

This method maintains the balance of API types and quantities. It ensures that all photographs of the same PAD are kept together in the development or test set to prevent data leakage and promote model generalizability.

Data Distribution

All Dataset (8001 images)

Development set (5,923 images)

Test set (2078 images)

Camera and Drug Heatmap

Directory Structure

The dataset is organized within the datasets/FHI2020_Stratified_Sampling directory. Below is an overview of its structure and contents:

• Development Set: The images for the development set are stored in the dev_images directory, which contains 5,923 images.

• Test Set: The images for the test set are located in the test_images directory, with a total of 2,078 images.

• Metadata: Metadata files are provided for both the development and test sets:
  • metadata_dev.csv contains metadata for the development set.
  • metadata_test.csv contains metadata for the test set.
  • labels.csv contains a list of categories/classes in the dataset.

Here is the visual representation of the directory structure:

datasets/FHI2020_Stratified_Sampling
├── labels.csv
├── metadata_dev.csv
├── metadata_test.csv
├── dev_images [5923 images]
└── test_images [2078 images]