Russia Develops Neural Network System to Diagnose Erythroderma
The system achieves an accuracy rate of about 77%.
Researchers at Sechenov University and the Higher School of Economics have developed a neural network system to diagnose erythroderma, a life-threatening condition in which inflammation affects up to 80% of the skin.
The system makes a diagnosis in two stages. First, a computer vision model analyzes images and produces the most likely diagnosis. A language model then compares that result with the patient's clinical presentation and refines the assessment. The developers note that erythroderma is extremely rare, so only a limited amount of data is currently available for training. On a small test dataset, the system achieved 77% accuracy after the two-stage analysis.
To obtain a preliminary diagnosis, a physician uploads images of the affected skin areas to a dedicated platform and completes a questionnaire about the patient's symptoms. The system analyzes the submitted data, identifies the most likely cause of the condition, and generates treatment recommendations.
The university said the algorithm was trained over approximately six months using anonymized skin images from 240 patients.
