AI system may help diagnose Zika

Scientists have developed an artificial intelligence system that can accurately diagnose Zika virus and several other viral, bacterial and even genetic diseases from the patient's blood.

The platform developed by scientists at the University of Campinas (UNICAMP) in Brazil, can identify tens of thousands of molecules present in blood serum, with an artificial intelligence algorithm.

"We used infection by Zika virus as a model to develop the platform and showed that in this case, diagnostic accuracy exceeded 95 per cent. One of the main advantages is that the method doesn't lose sensitivity even if the virus mutates," said Rodrigo Ramos Catharino, principal investigator at UNICAMP.

Another strength of the platform, he added, is the capacity to identify positive cases of Zika even in blood serum analysed 30 days after the start of infection, when the acute phase of the disease is over.

"None of the currently available diagnostic kits has the sensitivity to detect infection by Zika after the end of the acute phase. The method we developed could be useful to analyse transfusion blood bags, for example," Catharino said.

Development and validation of the platform involved analysis of blood samples from 203 patients treated at UNICAMP's general and teaching hospital.

Of these, 82 were diagnosed with Zika by the method currently considered the gold standard in this field: real-time polymerase chain reaction (RT-PCR), which detects viral RNA in body fluids during the acute phase of the infection.

The other 121 patients were the control group. About half had the same symptoms as the group that tested positive for Zika, such as fever, joint pain, conjunctivitis and rash, but had negative RT-PCR results for Zika.

The rest had no symptoms and also tested negative or were diagnosed with dengue.

All collected samples were analysed in a mass spectrometer, a device that acts as a kind of molecular weighing scale, sorting molecules according to their mass.

"We identified some 10,000 different molecules in the patients' serum, including lipids, peptides, and fragments of DNA and RNA. Among these metabolites, there were particles produced both by Zika and by the patient's immune system in response to the infection," Catharino said.

All the data obtained in the spectrometry analysis of both the group that tested positive for Zika and the control group were then fed into a computer programme running a random-forest machine learning algorithm.

This type of artificial intelligence tool is capable of analysing a large amount of data by specific statistical methods in search of patterns that can be used as a basis for classification, prediction, decision making, modelling and so on.

"The algorithm separates samples randomly, determines which one will be the training group and the blind group, and then carries out testing and validation," Catharino said.

"At the end, it tells us whether with that number of samples it was possible to obtain a set of metabolic markers capable of identifying patients infected by Zika," he said.

Each new set of patient data fed into the programme enhances its learning capacity and makes it more sensitive, he went on. In the case of Zika, the study established a panel of 42 biomarkers as a specific key to identifying the virus.

Twelve of these were found by the algorithm to be highly prevalent in the blood of patients who tested positive for the disease.