“The model sees what people may not notice”: US scientists create a neural network that predicts an imminent death

The authors argue that the system works more accurately than other existing methods.

Pennsylvania Healthcare Operator Geisinger taught AI to predict an increased risk of patient death. The system has enough ECG to say who will survive and who will die within the next year. This was reported by the New Scientist website, citing researchers who created the algorithm.

According to the authors, the system turned out to be more accurate than any other ECG-based forecasting methods. The Geisinger model detected heart problems even in patients who had previously successfully passed cardiologists.

The effectiveness of the algorithm was measured using the AUC indicator. This ratio measures how well the model sees the difference between two groups of people: those who die and those who survive over the next year.

The Geisinger system reached a level of 0.85 AUC with an ideal unit rate and no difference of 0.5. Moreover, the results of the model used by doctors in medicine range from 0.65 to 0.8.

We found that the model sees things that people apparently cannot notice or we simply ignore them and consider them normal. In theory, AI can teach us things that we probably misinterpreted for decades.

Brandon fornwalt

Principal Investigator Geisinger

Researchers trained the algorithm on 1.77 million ECG recordings from 400 thousand patients with voltage measurements at different time intervals. So the system has learned to see patterns that may indicate future heart problems, including heart attacks and atrial fibrillation.

At the same time, the researchers themselves do not know exactly which patterns allowed the algorithm to achieve such accuracy. As noted by New Scientist, because of this, some doctors do not trust predictive systems and criticize the possibility of their use in medicine.

This is not the first attempt to create a death prediction algorithm. In 2018, researchers at Google introduced a predictive model based on electronic medical records, which predicts the duration of patient care, discharge date and time of death in two US hospitals.

AI-based models are also often used to diagnose heart failure and lung cancer. In some cases, the algorithms deal with the analysis more precisely than people.

Back to top button