Page 201 - AI for Good-Innovate for Impact Final Report 2024

AI for Good-Innovate for Impact

data collection from the country's partnered hospital/medical college. The collected data from
primary sources will be analyzed, and a deep learning-based AI architecture will be constructed
for data training and testing. A detailed description of how the data collection and model
development phases will be carried out is provided here. 46 - BOU

The data collection phase will contain the following working steps:
1. Expand data sources: We will collaborate with hospitals to collect data on a larger and
more diverse population of newborns, including data from different ethnicities, gestational
ages, and birth complications. This process will also explore collecting data from pre-birth
stages (fetal heart rate, maternal health data) to identify potential risk factors.
2. Simulate data variations: Further, we will develop methods to simulate variations in
crying sounds due to background noise, microphone quality, and different recording
environments.

The model development phase will contain the following working steps:
1. Advanced AI models: The development phase will try to implement deep learning
architectures like convolutional neural networks (CNNs) and recurrent neural networks
(RNNs) to analyze and handle sequential data like cry sounds. It will also explore transfer
learning by leveraging pre-trained models on similar audio classification tasks. Moreover,
for global accessibility, it will explore methods for developing low-cost and low-power
solutions that can be used in resource-limited settings. In addition, we will focus on the
integration of generative AI (GenAI) model development, depending on the outcome of
the initial stage of model development.

46�3 Use case Requirement

• REQ-01: It is critical that the solution must provide an AI-powered tool for analyzing
newborns' cries via smartphones, accessible in resource-limited settings, and deployable
without specialized equipment, ensuring continuous data collection for AI model training
with minimal internet dependency.
• REQ-02: It is critical that the solution must have a user-friendly interface suitable for
healthcare professionals, designed to integrate seamlessly with existing healthcare
systems, and address barriers to adoption, such as smartphone functionality and internet
access.
• REQ-03: It is critical that the solution should support ongoing research and development,
be scalable and adaptable to diverse populations, and contribute to the UN Sustainable
Development Goal 3 by improving health outcomes for at-risk newborns.

185