AI software for classifying VOC patterns from breath, with a prototype gas sensor box, for research by a government environmental agency in Thailand

How the brief came about

The brief for this case was a tool to collect and classify VOC patterns from volunteers' breath, for research by the Environmental Research and Training Centre, Department of Environmental Quality Promotion. The centre had received a 2019 budget to study devices for measuring VOCs from breath. The research team needed an algorithm and software to distinguish the patterns of 3 groups of volunteers: healthy people, cancer patients and patients with other diseases.

We did this work with the project's main contractor, an engineering company. Our part was the sensor reading circuit board, the prototype gas sensor box and the software for collecting data, building models and displaying results.

Requirements and constraints

The system had to collect breath data from volunteers in every group and let researchers build models from that data. The terms of reference and project requirements specified that the system must do the following.

• Receive breath blown by volunteers into a gas sensor box with 10 sensors, then read values from every sensor.
• Classify the VOC patterns of 3 groups of volunteers and display results as numbers and graphs.
• Hold data on the local machine, then send it to the online database one person at a time or as a complete batch.
• Let researchers build AI models from healthy group and patient group data, and import models for prediction.
• Collect data and view reports on computers, mobile phones and tablets.

The main constraint was the hardware. The box was made by an outside manufacturer, and once the actual box arrived, the holes for the on/off button, the cable ports and the mounting holes for the board and fan still had to be modified to fit. A further point is that this system is a research tool, not a medical device for diagnosing disease.

How does this system collect and classify breath data?

The system receives breath through the gas sensor box, reads sensor values with a circuit board we developed, holds data on the local machine with the Data Collector program, then sends it to an online database. The web application then uses this data to build models, predict results and display them as numbers and graphs.

We chose to hold data on the local machine first so that collecting samples from volunteers could continue even without internet access. Model building sits on the web so that researchers can create new models from healthy group and patient group data, or import selected models, through the web pages for creating and importing models.

1. 1

   Volunteers blow into the gas sensor box

   The box contains 10 sensors and an air chamber that uses tubes and air lines to disperse breath to the face of every sensor. The circuit board we developed reads values from the sensors and passes them to a computer.

2. 2

   Data Collector sorts and holds data on the local machine

   The Data Collector program sorts the measured values and stores them on the local machine first, then sends them to the online database one person at a time or as a complete batch.

3. 3

   Create or import an SVM AI model

   Researchers create an SVM AI model (Support Vector Machine, an algorithm for classifying data into groups) from healthy group and patient group data through the web page, or import a selected model, then use that model to predict results from new breath data.

4. 4

   View results as numbers, graphs and history

   The web application has pages for recorded values, graphs, predictions shown as probability values, analysis and measurement history. The terms of reference required it to be viewable on a computer, mobile phone or tablet.

5. 5

   Calibrate the sensors against standard gases

   We collected calibration data at the Environmental Research and Training Centre of the Department of Environmental Quality Promotion by releasing standard VOC gas mixtures at concentrations of 5, 10 and 25 ppb (parts per billion) into the device, then used linear regression to find equations converting raw sensor values into concentrations.

Deliverables

We delivered the prototype hardware, software, documentation and calibration work, as listed below.

• A sensor reading circuit board of our own design.
• A prototype gas sensor box with 10 sensors and an air chamber that uses tubes and air lines to disperse breath to the face of every sensor.
• The Data Collector program for sorting and holding data on the local machine before sending it to the online database.
• A web application (PHP, Python, JavaScript) on Ubuntu Linux using SQLite / PostgreSQL databases, with pages for recorded values, graphs, predictions, analysis, history, model creation and model import.
• User manual
• A report on sensor calibration results against standard VOC gas mixtures at 5, 10 and 25 ppb, with conversion equations from linear regression.

What phases did this project go through?

This project ran from Jul 2019 to Apr 2020, divided into work under the original requirements and additional work after the demonstration. The sequence of work was as follows.

1. Jul–Oct 2019: Started work to the requirements and completed the sensor reading circuit board.
2. Nov 2019: Demonstrated the device and program and handed over the user manual. The client asked for an air chamber to be added and the screens to be adapted for smartphones.
3. Jan–Feb 2020: Received the box from the outside manufacturer, checked the mounting points and cable ports that needed modification, and built the air chamber.
4. Feb–Apr 2020: Collected calibration data, found the conversion equations using linear regression and prepared the results report.

Results confirmed at project close

The confirmed results of this case are the demonstration and the calibration work, both supported by documents. In November 2019 we demonstrated the device and program to representatives of the main contractor and the Department of Environmental Quality Promotion, and handed over the user manual. In 2020 we calibrated the sensors against standard VOC gas mixtures at 5, 10 and 25 ppb and obtained equations converting sensor values into concentrations using linear regression.

Our annual records state that when the project closed in 2020 the software was complete, while the circuit could not yet read all the sensors. This case has no model accuracy figures from real-world use, so we report only what is confirmed by documents.

What we learned

This work offers 3 principles for gas sensor work that uses AI to classify patterns.

• The sample path is part of the measurement system. Breath must reach the face of every sensor. In this work, the air chamber that uses tubes and air lines to disperse breath was added after the demonstration. Sensor box design must therefore include the airflow path, not just sensor mounting positions.
• Separate calibration from classification. Calibration tells you what concentration a raw sensor value corresponds to, while the AI model does the pattern classification. When you plan the two tasks separately, you will know whether an error comes from the sensors or from the model.
• Check every signal channel before collecting real data. When the project closed, the software was complete but the circuit could not yet read all the sensors. Work with multiple sensors should therefore include a step to confirm that every sensor is sending values before data collection from volunteers begins.

How can you apply this approach to your own research?

If you are developing a prototype gas-measuring device and need software for data collection and model building alongside it, you can use the structure of this case: a sensor enclosure, a program that holds data locally, an online database, and a website for building models and displaying results. What needs adapting is the sensor type, the reading circuit and the concentration range used for calibration.

• Define the scope clearly from the outset, covering the software, the circuitry and the enclosure, because work that combines hardware and software often picks up additional work along the way.
• Allow time to wait for the enclosure from the external manufacturer, and time to modify the board mounting holes and cable ports once the actual enclosure arrives.
• Have sample data from each volunteer group ready, because the AI model needs real data to be built and tested.