← All stories

Can AI and knowledge graphs help diagnose motor neuron disease earlier?

2 Jul 2025 · 2 min read · by the KnowLab team
The short version

In records of 67 million people in England, the current MND guideline was only moderately good at spotting the disease, and combining it with AI-derived signals improved early prediction.

Listen to this story · 2 min

The question

Motor neuron disease (MND) progresses quickly and cannot yet be cured, so catching it early matters. A Red Flag list from the MND Association has guided GPs since 2014, but nobody had tested at national scale how well it works, or whether data-driven methods could add to it.

What we did

We used linked health records for 67 million people in England. We compared the Red Flag list with three other ways of picking out the signs of MND: a medical knowledge graph, GPT-4, and machine learning on the records themselves. Each was used to diagnose MND and to predict it one, three and five years before a coded diagnosis. We also looked at how people with MND fared during COVID-19.

What we found

About 12,240 people had a coded MND diagnosis, roughly 18 per 100,000 over six years. Used alone, the Red Flag list gave only poor to moderate discrimination for early prediction, and it missed many cases. Combining it with the knowledge graph and the real-world-data approach did better, with the biggest gain five years before diagnosis. The purely data-driven approach was the best single method and found 13 new phenotypes, from seven chapters of the ICD-10 disease classification, that predicted MND. In the first COVID-19 wave, people with MND had about three times the risk of dying. By the second wave, the excess risk in fully vaccinated people was not statistically significant.

Why it matters

Guidelines built on expert knowledge are a good start, but this work suggests they miss signs that routine records contain. Mixing expert knowledge with knowledge graphs and large-scale data may help GPs notice MND sooner. This is a preprint that has not yet been peer reviewed, and prediction in records is not the same as proving that these signs cause or reveal the disease.