Nature Scientific Reports publishes novel decentralized federated AI learning algorithm that connects globally distributed data silos containing private and poor-quality data.

Nature Scientific Reports journal has published results of a new decentralized federated AI learning algorithm that can train AI on data that is distributed globally, without having to move private data to a central location. The breakthrough algorithm was able to achieve greater AI accuracy than traditional centralized AI training in real-world scenarios where data is poor-quality and contains errors, particularly when private data cannot be manually verified.

The algorithm, developed by AI company Presagen, presents a practical solution for connecting data silos and training AI in industries where data are sensitive and cannot be shared (moved), such as healthcare, defense, and finance. Research interest in distributed AI learning algorithms, called federated learning, has increased in recent years. However, many implementations have struggled with scalability, and typically require data to be transferred to a central location, potentially breaching data privacy.

Presagen’s Decentralised AI Training Algorithm (DAITA) approach ensures data privacy, is robust and is scalable as well as cost-effective for large real-world problems. Rather than moving data to the AI in a central location, the algorithm moves the AI to the location of the data, which can be distributed globally. Only the general abstract learnings from the AI trained on the data sources are shared, and never the individual datasets themselves.

Presagen Chief Scientist Dr Jonathan Hall explained “Using DAITA, we can optimise how the AI travels around the world – minimizing the cost of transfer, whilst simultaneously maximizing the performance of the final AI, all without looking at sensitive data.”

In collaboration with the Ovation Fertility network of IVF laboratories in the USA, the algorithm was applied to a healthcare problem of assessing the viability of embryos to assist embryologists in identifying embryos that are likely to lead to a pregnancy for IVF patients.

Ovation Fertility’s VP of Scientific Advancement, Dr Matthew (Tex) VerMilyea said “Embryo viability data has inherent errors. Patient embryos that are viable do not always lead to a pregnancy due to other factors related to the patient. Many problems in healthcare have poor-quality data, due to uncertainty or subjectivity.”

Presagen CEO Dr Michelle Perugini said “To train AI in healthcare that is unbiased and commercially scalable, you need to train AI on globally diverse data from clinics around the world, that represent different clinical settings and patient demographics. However, the challenge is that data privacy laws prevent sharing and centralizing medical data to train AI. Our decentralized AI learning algorithm addresses data privacy, data quality, and data diversity issues of connecting global data silos to AI for the benefit of patients around the world.”

Presagen has two commercial AI healthcare products in market globally, under its Life Whisperer brand for the fertility sector. Life Whisperer assess embryos for their viability and genetic integrity. Life Whisperer is currently being distributed globally by FUJIFILM Irvine Scientific.