In Feb 2023 Dr Jonathan Hall presented at the Indian Society For Assisted Reproduction (ISAR) the presentation “Data you never see: Federated Learning for Oocyte Assessment”. A video of the presentation, the presentation slides and transcript are available below.
Hi everyone, Thanks for joining me. I’m Jonathan from Presagen.
Today, I will be talking about how a new AI method made it possible to create a new medical innovation: human egg assessment in fertility, focusing on IVF.
Not all AI technology can be built using existing datasets.
Sometimes, new valuable AI technology requires a completely new dataset that does not currently exist. Other times, under-serviced areas such as women’s health lack data availability.
In this case, this new data needs to be generated and accessed by the AI.
But this has inherent challenges for healthcare.
We will explore why, and how we have solved this challenging problem, using our new AI product that’s in development as an example, which is oocyte assessment in IVF.
The IVF process is one of the last options for couples struggling with infertility.
It is a type of assisted reproductive technology.
In IVF, an egg is fertilized, and develops into an embryo.
The embryos are assessed in terms of their quality, and then transferred back to the patient, with the hope of a pregnancy or a live birth.
Artificial Intelligence is making a big difference in IVF.
There are many ways AI is helping.
AI can be used to look at human egg quality, sperm quality, and track developmental milestones.
AI can also look at the genetic integrity of a developed embryo, and predict the likelihood it will implant.
In this talk, we are going to focus on the human egg, or oocyte, and how AI can be used to assess their quality using AI-based computer vision.
To create an AI that can assess human eggs, we need data so that the AI can learn what a good quality egg looks like.
This is not so easy, because most hospitals and clinics don’t have a large database of images of eggs.
Clinics would like to collaborate to help create an AI, but they have stringent data privacy and regulations that prevent sharing data.
Instead, some clinics opt to create their own AI that works only for their clinical setting and patients.
This fails, because most clinics don’t have a large enough or diverse enough dataset to train a robust and scalable AI.
Larger hospitals can create an AI, but they typically only work inside that one hospital, meaning the AI is not available or scalable to the whole world.
So clinics struggle to innovate in that, they can’t collect data in one place, and they can’t access each other’s data.
This problem is called ‘Siloed Data’, where the data are locked up, distributed around the world.
It is common in the area of medicine and healthcare.
But using a new AI technology that we developed at Presagen, ‘Decentralised Federated Learning’, we turned the problem on its head!
“What if you could access distributed data without moving it or seeing it?”
Being able to access multiple clinics and hospitals data can help AI to generalize across distributions by accessing *diverse data*.
But we only want the AI Learnings from the data. We don’t want to know private patient information.
This new AI training methodology is a key enabler, by having the AI go to the data, and back out again, without removing anything.
If ‘AI goes to the data’, in each local region, instead of the data being moved, then AI can learn from a distributed and diverse dataset from around the world.
This allows clinics, hospitals, and AI developers to collaborate, all while respecting safety and privacy.
These results were published recently in Nature Scientific Reports, which showed that a Federated AI can even exceed the performance of a traditionally-trained AI, especially on difficult medical datasets.
Using this technology, where are we so far, with Egg (or Oocyte) Assessment?
How can clinics get involved?
I would like to introduce Presagen AI Open Projects.
These are projects where clinics and hospitals can safely collaborate, to help create targeted healthcare products using AI, to solve real-world problems.
The benefits are that the AI is robust, reliable, and more accessible than ever before.
As of February 2023, the Egg Assessment AI development is near to completion.
We have collaborated with a large network of clinics than span across India, The USA, Japan, South East Asia, and the Carribbean, to federate a diverse dataset.
Each dataset is kept in-region, and automatically makes use of Federated Learning to train AI and protect privacy.
The AI is then double-blind tested, which means that it needs to work in hospitals it has never been trained on. This is a necessary step before it is ready for clinical use around the world.
This process ensures the AI can generalize to new demographics, clinical settings and patient groups.
If you are a clinic, to be involved, we have a Research & Development Protocol and various forms and templates needed to get started quickly. Clinics that contribute to development, or to data used for testing the AI, can earn royalties in return for the collaborative assistance given.
How does Presagen AI Open Projects work?
We have a Partnering Program, where a clinic or hospital may join our global cloud platform, receive the necessary protocols to collaborate, and join with clinics around the world to help build scalable products.
The clinics can safely prepare their data in a Clinical Data Portal that helps structure the data and prevents private data from being seen. All data remains in region.
Then the Federated AI will run on the distributed data. After testing and validation, we work with clinics to help conduct trials and we obtain the regulatory approvals needed for global markets.
Finally, clinics and hospitals can receive these exciting new products using a cloud-based delivery platform, with a simple drag-and-drop interface.
Now I would like to describe early results for Egg Assessment.
While we continue to add new AI Open Projects, within and beyond IVF, the clinical relevance of Egg or Oocyte selection is as follows.
What if we could help select eggs that result in good quality embryos, and help patients understand the likelihood that the egg will develop into a good quality embryo, prior to fertilization?
What if we could help make decisions about whether additional rounds of egg collection are needed? This could help patients to plan their fertility arrangements.
Using computer vision, the AI analyzed eggs that had been prepared in a special way, prior to fertilization.
The AI had access to knowledge as to whether the egg actually developed into a blastocyst stage embryo, and learned the features of the egg that make it most likely to develop.
The AI was actually a full system designed to be robust at object detection, finding different key regions of the egg called segmentation, and analyzing all these features to give the best assessment of likelihood to develop.
A separate AI system was designed to ensure that male infertility factors could be identified so that we can be sure it is the egg’s quality that is being measured.
In a collaboration with California Fertility Partners and Ovation Fertility, results were presented at the American Society for Reproductive Medicine annual conference in 2022.
It was found that the AI was able to identify 31.6% of known cases of male infertility, a significant proportion, and other data that could confound the results.
When identifying these factors, our AI improved almost 10%, and leading to a preliminary predictive power of 83.7% of identifying which eggs would develop into embryos.
In addition, the features of the eggs 1-day after fertilization, with the ability to create an AI that can help predict different developmental milestones of the embryo were collected.
The AI is able to provide a score from 0 to 10 that indicates the likelihood the oocyte will develop.
These promising results will be tested on data that the AI has not been trained on, to ensure it is robust and scalable.
To summarise, a collaborative network of clinics is vital for collecting diverse data.
We can create AI that works across demographics and clinical settings, all while respecting data privacy.
We currently have collaborators world-wide from 48 countries, contributing on a range of projects. Please reach out for more info at email@example.com.
We found that AI can help assess human eggs, or oocytes, and demonstrated predictive power for whether they would develop into good quality embryos.
The final AI is due to be tested in the coming months, to ensure it works in clinics it has never been trained on.
Life Whisperer Oocytes will work in tandem with our other AI technologies, Life Whisperer Viability and Life Whisperer Genetics, to help improve embryo quality and pregnancy outcomes.
I would like to thank you for joining this webinar.
This presentation is available on our websites at Presagen.com and LifeWhisperer.com