Can algorithms anticipate heart failure? Turns out they can...

One of the key hallmarks of potential heart failure is the buildup of excess fluid in the lungs - a condition called “pulmonary edema”. The level of excess fluid in the lungs correlates to the risk the patient is in, and thus the course of action the doctor must take. However, gauging the amount of fluid in the lungs is tricky and can often lead to incorrect treatments. As such, researchers at the Massachusetts Institute of Technology’s CSAIL lab (Computer Science and Artificial Intelligence) have developed an algorithm that analyses an X-Ray image to gauge how risky the pulmonary edema is, quantifying it on a 0-3 scale where 0 is healthy and 3 is very risky.

Algorithms like these are becoming increasingly prevalent in modern medicine through their ability to spot subtle but crucial changes in internal conditions that doctors cannot see. In training the algorithm, the researchers collated over 300,000 X-Ray images and their corresponding radiologist reports, finding success even by just feeding the algorithm with the written reports that didn’t have any sort of ‘severity index’ on them.

Our model can turn both images and text into compact numerical abstractions from which an interpretation can be derived. We trained it to minimize the difference between the representations of the X-ray images and the text of the radiology reports, using the reports to improve the image interpretation.
— Geeticka Chauhan - PhD student at MIT who was co-lead author of the paper

The researchers didn’t stop there, though. In fact, the two most significant strides made through this research project came after the algorithm had been developed:

  1. The team took existing X-Ray images from public databases and used the algorithm to develop their own ‘severity indices’ - an idea which could potentially be implemented in the future as a standard baseline for future ML algorithm testing in the field.

  2. The algorithm can also explain itself - pointing to specific areas of X-Ray images or specific parts of radiologist reports that correspond to the model’s prediction, again seeking to improve the correlation between the X-Ray image and the corresponding radiologist report.

The use of Artificial Intelligence in medicine will surely expand in the future. However, in the here and now, this technology can be applied to attribute a risk level to other conditions that are commonly associated with pulmonary edemas, such as sepsis or kidney failure. Thus, the scope of machine learning algorithms extends beyond just diagnosing a disease. Rather, machine learning algorithms could almost standardise observational reporting across the entire medicinal field both increasing consistency and reducing potential for error. It shall surely be exciting to see how more such applications of AI emerge in medicine.

Risk Level diagram from MIT’s Press Release.

Thanks for reading!

Ayushman Nath