The scientists of the Institute for Problems in Mechanical Engineering of the Russian Academy of Sciences, together with the colleagues from the Institute of the Human Brain of the Russian Academy of Sciences and the Swiss Brain and Trauma Foundation, developed a method for detecting schizophrenia by hidden components of electrical signals in the brain.

A new approach based on the mathematical separation of brain waves into separate sources allowed the machine learning model to classify patient data with a sensitivity of 96.7% and a specificity of 97.7%. The results of the study were published in the European Physical Journal.

Schizophrenia is a severe mental illness that affects about 1% of the population throughout their lives, with almost half of patients facing lifelong disability. The key problem remains the late and inaccurate diagnosis, which to this day relies on clinical interviews and questionnaires, which allow for ambiguous interpretation. The delay in starting treatment worsens the prognosis directly, so the search for objective biological markers, or neuromarkers, is one of the priorities. For several decades now, studies of the electrical activity of the brain have been conducted using electroencephalography (EEG) and the evoked potentials (EP) method.

Previously the scientists have already tried to use the parameters of P300 waves, conditional negative deviation and other EP components to classify data, and also used machine learning methods, achieving accuracy of up to 90%.

However, the EEG signal recorded from the scalp surface is a mixture of activities from many deep brain sources, which masks important diagnostic phenomena.Previously, the IPMash RAS scientists had already achieved impressive results by applying the support vector method to traditional EP, obtaining a sensitivity of 91%.
“In this work, we took the next step, aiming to improve diagnostic accuracy by identifying hidden, «latent» signal sources. To do this, we used a unique signal separation method called «blind separation of sources », which allows us to reconstruct mathematically the original signals from different neural networks of the brain, even if they overlap greatly in time”, — said Nadezhda Shanarova, an intern researcher at the Institute for Problems in Mechanical Engineering of the Russian Academy of Sciences.

The data was collected during the visual test, a task which evaluates the ability to cognitive control, inhibition of impulsive actions and concentration of attention. These very functions are often disrupted in schizophrenia. The study involved 68 patients diagnosed with schizophrenia and 132 healthy volunteers.

Instead of feeding the machine learning model with raw or averaged EEG signals from 19 electrodes, the scientists first converted the data from each tested participant into 11 components. Then they compared statistically the signals of these components between the groups and determined the time intervals where the differences were most significant.

Many features were extracted from these sites, supplemented with behavioral data (for example, reaction time and number of errors), and the most informative ones were left using sequential selection. The result demonstrated the high effectiveness of the proposed approach. The model, trained on a combination of latent component features and behavioral data, showed sensitivity of 96.7% and specificity of 97.7%. This means that the system identified correctly almost all patients and showed the minimum number of false positive classifications among healthy people under test.

High accuracy was achieved not just by making the algorithm more complex, but by seeing hidden processes in individual functional brain networks. Thus, one of the isolated components was associated with the processes of inhibition of action and had previously shown a specific response to therapy in patients with ADHD (attention deficit hyperactivity disorder). Thus, the application of the «blind separation of sources» method to evoked potentials in combination with machine learning can improve significantly the accuracy of instrumental diagnosis of schizophrenia compared with the traditional method analysis. The discovered hidden components indicate the particular brain networks which are disrupted by the disease. This opens up opportunities for development of the targeted therapies using neural feedback or noninvasive brain stimulation.