It means retraining a model where it is used, rather than sending the data away to a central lab. In this study the team labelled images collected on day one, retrained the model between the two days, and sent the new model back to the microscope. Retraining took about two hours.

how big was the study?

About 100 community members aged five and older near Azaguie in Cote d'Ivoire each gave a urine sample on two days in a row. The team processed 75 samples on day one and 68 on day two. That is a small field study, not a national trial.

is the tuned model good enough for real programmes?

Not yet, by the authors' own account. Against normal light microscopy the tuned model caught 97.8 percent of infections but called only 68.2 percent of clean samples clean. The authors write that this is close to meeting World Health Organization requirements and that more field tuning should close the gap.

can anyone else check these results?

Not easily. The paper says the data and algorithms may be made available upon request, and asks interested parties to contact the corresponding author. There is no public repository link for the data or the code, so an outside group cannot rerun the work today.

Retuning AI on site raised parasite detection to 98.1%

What happened

Think of a driver who learned in another country. The driving is fine. The local signs are not. One afternoon with a local guide fixes more than a year of extra practice would.

Software has the same gap. A model trained on images from one place lands in a new place and stumbles. The usual fix is to ship the new data back to a big lab and wait weeks.

A team in Cote d’Ivoire did not wait. They ran a clinic day, labelled the day’s own images on site, retrained the model that evening, and pushed it back to the microscope. Then they tested it the next morning.

The test

The paper is “Edge-tuning of artificial intelligence improves diagnostic performance for Schistosomiasis haematobium in a rural setting of Cote d’Ivoire”. The authors are Maria Diaz de Leon Derby, Jean T. Coulibaly, Elena Dacal and colleagues. The work involves Universite Felix Houphouet-Boigny and the Centre Suisse de Recherches Scientifiques in Abidjan, Cote d’Ivoire, along with groups at the University of California, Berkeley, Stanford University, Lawrence Berkeley National Laboratory and the Chan Zuckerberg Biohub in the USA, Spotlab and the Universidad Politecnica de Madrid in Spain, and the University of Toronto and Toronto General Hospital in Canada.

The disease is caused by a worm. People pick it up from fresh water. The eggs show up in urine, and you find them under a microscope. The paper says the disease affects over 200 million people.

The team worked near Azaguie, Cote d’Ivoire, in January 2024. About 100 community members aged five and older gave a urine sample on two days in a row. The samples went under a portable microscope called the NTDscope, which takes two kinds of image. On day one the team processed 75 samples. On day two they processed 68.

The tuning happened in between. Seven people spent four hours marking up images from 36 patient samples. They found 190 positive images and made 7,900 new marks. The model was then retrained on those marks plus the old training data. Retraining took about two hours. The new model went back to the microscope and ran on every day two sample.

Both the old model and the new one were scored on the same day two samples. That is what makes this a fair comparison.

The result

Two numbers matter for a test like this. One is the catch rate, or how many infected samples it finds. The other is the clean rate, or how many uninfected samples it correctly calls clean.

For the brightfield images, the catch rate went from 75.5 percent to 98.1 percent. The clean rate went from 46.7 percent to 100 percent.

Read that clean rate again. Before tuning, the model called more than half of the healthy samples infected. THE OLD MODEL WAS BARELY BETTER THAN A COIN TOSS ON CLEAN SAMPLES.

For the darkfield images, the catch rate went from 72.2 percent to 90.7 percent. The clean rate went from 64.3 percent to 100 percent.

The team also compared the tuned model against normal light microscopy, the everyday standard. There the tuned brightfield model caught 97.8 percent of infections, but its clean rate was 68.2 percent. The darkfield model caught 95.7 percent, with a clean rate of 77.3 percent.

What it means

Four hours of local labelling beat every clever thing that came before it. That is the headline for anyone shipping software into a new place.

It also shows the shape of the problem. A model that arrives from elsewhere is not broken. It is simply tuned for different water, different light and different sample handling. Local examples fix that faster than a bigger model would.

And there is a warning inside the good news. The old model looked fine on catch rate and failed badly on clean rate. If you only check one number, you will ship a tool that sends healthy people for treatment.

Business ideas from this paper

A one-day tuning visit, where you bring labellers and a laptop to a clinic, retune their deployed model on their own images, and hand back the update before you leave. Who buys it: health programmes and device makers taking a tool into a new country. A price to test: 3,000 dollars for one site visit. A one-week test: run one visit free, chart the before and after, and show it to two device makers.
A small labelling app for clinic staff, in the local language, that turns one shift of marking into a ready training file. Who buys it: research groups and diagnostic device makers who need field data. A price to test: 300 dollars a month for each site. A one-week test: give it to one lab for seven days and count the marks it produces.
A does it work here report, where you run an imported diagnostic model against 50 local samples and publish both its catch rate and its clean rate. Who buys it: ministries and NGOs about to buy a device. A price to test: 1,200 dollars for each report. A one-week test: do one report on a device you can borrow, then send it to three buyers and ask what they would pay.

How sure can you be?

The comparison is honest, and the study is small.

Look at the size. One area, around one town, in one month. About 100 people. Sixty-eight samples on the second day. That is enough to show a large effect. It is not enough to set a policy.

Look at who the model was tuned on. It learned from images of people in that same community, and it was then tested on people in that same community, on the next day. Nothing here shows what happens in the next district.

Look at the comparison with normal light microscopy. The catch rate held up at 97.8 percent. The clean rate did not, at 68.2 percent. About three clean samples in ten were still flagged. The authors say their performance against light microscopy is close to meeting the requirements of the World Health Organization, and that more field tuning should get them there. So it does not meet them yet.

Two more limits. This is a preprint, so no reviewer has checked it. And the paper says the data and algorithms may be made available on request, which means you cannot check the numbers yourself today.

We score this paper 4 out of 10. The team is strong and the field work is real. The venue, the openness and the small size all hold the score down.

What would settle it? The same tuning method run at several sites in different countries, tested on people whose samples were never used for tuning.

Do this today

If you run any model that gives a yes or no answer, write down both error types. Count how often it misses a real case, and count how often it raises a false alarm.

Source: Edge-tuning of artificial intelligence improves diagnostic performance for Schistosomiasis haematobium in a rural setting of Cote d'Ivoire, July 2025. doi:10.1101/2025.07.11.25331398 · www.medrxiv.org (preprint · not yet peer reviewed).

Just Out Tech explains new research in plain language. This article was drafted with AI assistance and checked by a human against the original source.