AI can spot parasites in stool samples to help diagnose infections

Artificial intelligence can spot parasitic worm eggs in human faecal samples – including those from parasite species missed when lab technologists use a microscope to study the same samples. The discovery suggests AI could help us better diagnose and treat parasitic worm infections across the globe.

The World Health Organization estimates that almost one-quarter of the world’s population – or 1.5 billion people – are infected by parasitic worms living in their intestinal systems. The infections can lead to malnutrition, anaemia or stunted cognitive development. But diagnosis and treatment is often inaccessible because there are a limited number of experts trained to spot the infections.

Johan Lundin at the Karolinska Institute in Sweden and his colleagues wondered whether AI could help. “The method is primarily about enabling wider access to diagnosis of parasitic worm infections,” he says.

The researchers trained and tested their AI system on about 1300 stool samples collected from school students in Kenya. The samples were prepared by a local healthcare laboratory and digitally scanned under a microscope. Those scans were then uploaded via mobile internet to the cloud for the AI analysis.

AI training focused on identifying eggs from three types of parasitic worms: the roundworm Ascaris lumbricoides, the human whipworm Trichuris trichiura and hookworms such as Ancylostoma duodenale or Necator americanus.

The team assessed the AI’s performance against that of a trained lab technologist who inspected the samples manually. The AI method accurately detected 76 to 96 per cent of infections spotted by the technologist, depending on the parasitic species.

“This study showed a relatively decent sensitivity and a high specificity for identifying these parasitic worm infections,” says Isaac Bogoch at the University of Toronto in Canada, who did not participate in the study. “It’s very good news.”

Importantly, the AI also proved capable of identifying infections even with relatively few parasitic worm eggs. In fact, it spotted 79 cases that the human expert had missed. But at the same time, the AI avoided falling into the trap of falsely identifying infections that weren’t actually present. In just 1 to 2 per cent of cases did the AI incorrectly identify an infection in a sample that ultimately proved to be free from infection.

At the moment, it takes 20 to 35 minutes to process each sample using AI. But this is largely because the study involved using slow networks to upload the data. The actual AI analysis takes just 5 minutes, and so the researchers suggest processing times could be reduced by 8 to 17 minutes with access to 5G networks for faster data uploads. What’s more, unpublished cost estimates suggest the AI detection will “certainly be cheaper than entirely manual methods”, says Nina Linder at Uppsala University in Sweden, a coauthor on the study.

Even so, it remains to be seen how well the AI method could speedily diagnose people in their home communities to inform clinical decisions, says Bogoch. And given that the technique involves uploading health data to the cloud, he points out that there are ethical considerations in terms of informed consent and data privacy.

“This is incredible technology but it’s got to be done within an ethical framework,” says Bogoch. “And I have no doubt that they’re doing that.”