Machine learning models have learned to quickly and accurately diagnose stroke
Researchers from Carnegie Mellon Universities, Florida International University and Santa Clara University have developed a new model for diagnosing stroke that uses machine learning principles in its work. Its main difference from existing analogues is that it works without the need for diagnostic imaging of the state of human internal organs. Therefore, it can be used even where there is only a minimal set of equipment.
Diagnosing a stroke is extremely difficult due to the specifics of the disease – for example, in 25% of cases there are simply no characteristic external symptoms. The death rate from a misdiagnosed stroke is 30 times higher than that from an undiagnosed heart attack. It is easy to confuse the symptoms of a stroke with dozens of other diseases and abnormalities, so the need for automatic diagnosis is very high.
The new model was trained on 143,000 medical records of patients in Florida hospitals, the sample included people of all races and ages. Testing has shown that it can detect a stroke, or determine that it has happened in the past, with an accuracy of 84%. For other models, this figure does not exceed 30%. However, the authors of the development insist that their model is not yet suitable for autonomous use, it should be used together with other methods.