Detecting a heart attack thanks to artificial intelligence
In an international study, researchers at the USZ were able to show that artificial intelligence can outperform experienced cardiologists in the analysis of cardiac ultrasound data. Nevertheless, the road to clinical use is still long.
Takotsubo cardiomyopathy is an acute pump dysfunction of the heart that affects a majority of women and occurs mainly after emotional or physical stress events. The disease resembles a heart attack in the acute phase. Although the distinction is central for further adequate treatment, clear criteria based on a cardiac ultrasound examination are still lacking.
Does artificial intelligence recognize the difference?
In this collaborative project with ETH Zurich, the researchers investigated whether machine learning could help distinguish between the two cardiovascular diseases. As a basis for their study, they used data from the international Takotsubo Registry on the one hand and the Zurich Acute Coronary Heart Disease Registry on the other. In total, the cardiac ultrasound examinations of 224 patients with acute myocardial infarction and 224 patients with Takotsubo syndrome were included.
In a first step, a deep learning model was developed. Data from a total of 228 patients were used for training. The goal in such procedures is for the "artificial intelligence" to recognize patterns in the unstructured raw data and for these patterns to become continuously more precise as the volume of data sets increases. In this way, AI may be able to assign images or make distinctions that escape human attention.
AI was superior to cardiologists
In the next step, the algorithm developed in this way was used to analyze the other 200 data sets. To compare accuracy and precision, four experienced cardiologists evaluated the same 200 data sets. The evaluation of the results showed that the fully automated analysis using artificial intelligence was superior to the cardiologists.
However, further studies must follow before it can be used in everyday clinical practice. Not least because in this case the underlying data were limited to two disease patterns and a limited number of data sets. "Nevertheless, we were able to show the potential of AI with this study," explains Christian Templin, a cardiologist at USZ and the study's lead author. "If larger datasets are available in the future, predictions using Deep Learning could be significantly improved and provide further insights into the dynamics of normal and pathological cardiac function." With ever-increasing amounts of data in medical diagnostics, the need for efficient processing and analysis is also growing. The use of AI is only just beginning.
