Artificial intelligence predicts life expectancy based on CT scan

Artificial intelligence predicts life expectancy based on CT scan

Scientists from the University of Adelaide in Australia has taught artificial intelligence to predict life expectancy of patients with serious diseases. Scientists say that the accuracy of their programs is 69 percent, which is comparable with the predictions of the doctors.

The paper was published in the journal Scientific Reports.

Usually to examine the condition of the internal organs of patients imaging is used allowing to obtain cross-sectional image of the object. In order to determine the patient’s condition and to forecast about its future change, requires high qualification of the observation of a specialist. In order to accelerate, automate, and Refine the diagnosis of diseases, scientists are developing a computer system with machine learning.

In their work, the researchers decided not to teach the computer to search for a specific disease, and to develop a system that can give an overall assessment of the health of patients and the forecast of further life expectancy.

For this purpose they used data of computer tomography of the chest patients with age 60 years or more, observed within the previous few years.

Thus, the researchers had data on the initial condition of the patients and the further course of the disease. It was formed two sets of data. The first was assembled from pictures of the 24 patients who died in 2014, who within five years prior to death, were examined using computed tomography. When forming groups for selected patients without apparent signs of acute illness, metal objects in the chest and diagnosed with active cancer. The second group was formed from a similar group of 24 surviving patients.

The program, developed by the researchers based on convolutional neural networks. This type of neural networks is often used for image processing because it allows to allocate in them the features of different scale. In this case, the neural network is trained to distinguish images of patients so-called biomarkers — sets of diagnostic parameters that distinguish different disorders on the background of healthy tissue. The system searched in the organs symptoms of various diseases and syndromes such as emphysema or heart failure.

As a result, observing the changes in images of the same patients, the neural network was able to identify their five-year survival rate of patients with an accuracy of 69 percent.

Scientists say that this figure is generally similar to the accuracy of doctors ‘ estimates. It should be noted that the used sample is insufficient for an unambiguous statement about the accuracy of the presented method. Usually training of neural networks is used several orders of magnitude larger volume of source data. Probably, when the number of examined patients it can outperform physicians in accuracy of diagnosis and assessment of the General state of health.

This is not the first case of disease diagnosis by artificial intelligence. In 2016, the computer using machine learning taught to separate melanoma from benign pigmented lesions with an accuracy of 98 percent, and in early 2017 scientists have presented a system to predict the risk of death in patients with cardiovascular disease with an accuracy of 73 percent.

Gregory Copies