About
Asquire is a data collection application designed to collect voice data related to sustained phonations, breath sounds, and cough sounds. The objective of this project is to analyze the sound characteristics of the collected voice data (recorded with smartphones) using data-driven and signal processing methods. The ultimate goal is to leverage machine learning (ML) techniques for the diagnosis and monitoring of asthmatic patients.
The current clinical methods for diagnosing asthma are often tedious, expensive, and time-consuming. The Asquire project was developed with the motivation to provide an easy, effective, and fast diagnosis method for asthma using vocal sounds and ML algorithms. By utilizing the power of machine learning and signal processing techniques, this project aims to overcome the limitations of traditional diagnostic methods and provide a more efficient and accurate solution.
System design
Data Collection and Processing
Asquire collects voice data from users through a web application. The users are instructed to perform sustained phonations, provide breath sounds, and cough sounds. The application then processes this voice data using data-driven and signal processing methods. Various sound characteristics such as frequency, intensity, duration, and patterns are extracted from the voice samples.
The processed voice data is then used to train a machine learning model. The ML model learns from the collected data to recognize patterns and associations between sound characteristics and asthma conditions. Once the model is trained, it can be used to analyze new voice samples and provide a diagnosis or monitoring feedback for asthmatic patients.
Steps to contribute
To contribute to the Asquire project:
- Visit the Asquire web-app https://asquire.web.app
- Read the consent form carefully
- Create a new user ID to get started.
- Fill in your details and respond to the survey provided.
- Proceed to the recording section and carefully follow the instructions to record your voice.
- Once you have completed the recording, submit your data.
- Our team will verify the submitted data, and for successful contributions, you will be compensated accordingly.
- Bonus! There is a fun voice changing interface you can play with!
Tech stack
React, Firebase, Affinity Designer, Python Pandas, Machine Learning
App screenshots
Challenges
Data Quality and Consistency: Ensuring the quality and consistency of voice data collected from various users posed a significant challenge. Differences in recording environments, microphone quality, and user proficiency could introduce variability in the collected data, impacting the effectiveness of the machine learning models. So we had to implement a robust data validation and preprocessing pipeline to address these challenges. Along with we created in-app video + audio instructions to guide the users through the recording process.
Model Generalization and Robustness: Developing machine learning models that generalize well across diverse populations and are robust to variations in voice characteristics presented challenges. Ensuring that the models could accurately diagnose asthma across different age groups, genders, and ethnicities while avoiding overfitting to specific data distributions required rigorous validation and testing.
Feature Extraction and Representation: Extracting relevant features from voice data for use in machine learning models required careful consideration. Identifying informative features related to asthma diagnosis, such as frequency characteristics of breath sounds or cough patterns, while filtering out irrelevant noise, was a complex task that required expertise in signal processing.
Privacy and Ethical Considerations: Addressing privacy concerns and ensuring ethical use of collected voice data was a key challenge. Implementing robust data anonymization techniques and obtaining informed consent from users. We obtained ethical clearance from the Institutional Ethics Committee (IEC) of St.John's Hospital, Bangalore, under the guidance of pulmonologists to ensure that the project adhered to ethical guidelines and data protection regulations.
Integration and Scalability: Integrating the various components of the Asquire system, including the web application, data processing pipeline, machine learning models, and user interface, while ensuring scalability and reliability, presented challenges. Balancing system performance with resource constraints and accommodating potential future growth required careful architectural design and optimization.
User Engagement and Participation: Encouraging user engagement and participation in data collection activities was crucial for the success of the project. Designing intuitive user interfaces, providing clear instructions, and offering incentives for participation were strategies employed to overcome this challenge and ensure a sufficient volume of high-quality voice data for model training. We designed a student volunteer scheme where students could contribute to the project and earn by bringing in more users.
Impact
We collected around 10 hours of data from 200+ users. The data was annotated using the annotation scheme designed for asquire. The data collected is being used to develop machine learning models for asthma diagnosis and monitoring.
Gender distribution
Age distribution
For further details, refer
- Asquire Corpus: A detailed description of the data collected and the annotation scheme used for the Asquire project.
- Asquire Tako: A web-app designed to record vocal breath sound simultaneously from multiple devices.
- Asquire VAD: Automatic validation of asquire data from crowd sourced data.