Oakden-Rayner, Luke, et al. “Precision Radiology: Predicting longevity using feature engineering and deep learning methods in a radiomics framework.” Scientific Reports, vol. 7, no. 1, Oct. 2017, doi:10.1038/s41598-017-01931-w.

 

In this experiment, researchers at the University ofAdelaide used artificial intelligence to determine if a computer could be taught to predict a person’s lifespan based on CT scans of chest organs. The computer was designed to tell whether or not the patient would survive 5 years based on the scan. The computer was given multiple scans of 48 different people to study. It was able to identify the mortality rate with a 69% accuracy. This number is comparable to human doctor’s determinations. The idea behind this experiment is that given enough images, the computer will be able to recognize patterns that a human would not, leading to a more accurate prediction.

In total, 15,987 unique data points were extracted from all the CT scans.The computer made its overall prediction based off of how each data point compared to an ideal data point. That is, the computer compared the test scan to a scan that showed perfect, healthy tissue. Statistical analysis of how well the computer matched these data points was used to determine the effectiveness of the study.

The researchers themselves identified several issues with their data set. One, the sample size needs to be increased. This was not possible due to the resources needed to gather large CT scans as samples. However, the 48 patients used in this study were the largest sample size used in any experiment like this to date. Another issue was that each individual scan must be segmented before the computer can analyze it: this takes a great deal of time to prepare, so this factor also limited the data set. The team of scientists saw this study as a proof of concept test, and hope it will encourage further research.