From discovery to diagnosis, how AI could support medical research

Artificial intelligence (AI) is a hot topic in the research world, and for good reason. The promises made by AI optimists are exciting, from AI drug discovery to robot doctors as good as any human physician. AI pessimists are quick to point out that we have a long, maybe impossibly long, way to go before we see any real substantive benefits in the medical research field.

Whether you are an optimist, a pessimist, or just like watching how things play out,  keeping up with the latest AI news can be entertaining and informative. Here’s a roundup of some recent advances to catch you up.

Side note: you have probably heard the terms AI, machine learning, and deep learning, and sometimes they are used interchangeably. They do have specific meanings for computer scientists, but in the simplest terms AI is the broadest category; machine learning is a narrower type of AI, and deep learning is a narrower type of machine learning. One kind of deep learning algorithm is also known as artificial neural networks. Read this primer if you’d like to know more.

Relative Prognostic Importance and Optimal Levels of Risk Factors for Mortality and Cardiovascular Outcomes in Type 1 Diabetes

(Circulation, A. Rawshani et. al., 2/25/2019)

In this observational cohort study, researchers used machine learning and traditional analytic tools to evaluate cardiovascular and mortality outcomes for people with type 1 diabetes from the Swedish National Diabetes Register from 1998 to 2014. Researchers ultimately narrowed the data down to the five best predictors for mortality and cardiovascular events: glycated hemoglobin, albuminuria, duration of diabetes, systolic blood pressure and low-density lipoprotein cholesterol.

For this study, machine learning was used in conjunction with traditional tools to analyze the data. This is one of the most modest and realistic uses for AI in research – ultimately, it is just a more sophisticated statistical analysis tool. However, the benefit comes from the speed at which these studies could be produced in the future. 

Net-FLICS: fast quantitative wide-field fluorescence lifetime imaging with compressed sensing – a deep learning approach

(Light: Science and Applications, Ruoyang Yao et. al., 3/6/2019)

Rensselaer Polytechnic Institute researchers used a convolutional neural network (a set of deep learning algorithms) to improve the reconstruction of images of organs and tumors in live subjects. Their program, which they called Net-FLICS (fluorescence lifetime imaging with compressed sensing), processes the compressed sensed (CS) measurements from macroscopic fluorescence lifetime imaging (MFLI) at lightning speed, and could be used to produce real-time images.

Using an imaging system trained though deep learning, doctors could potentially see fluoresced tumors in patients with high accuracy. This could lead to precisely targeted therapies, or even surgical tumor removals with much smaller margins of healthy tissue needed.

Much more research is needed before this program can be used on patients. However, the greatest strength of machine learning lies in the second part of the name: learning. The more data these programs assimilate, the better they get at their jobs. By the time algorithms like these can be used by doctors, they will have absorbed an incredible amount of data, making them much more accurate.

Automated Triaging of Adult Chest Radiographs with Deep Artificial Neural Networks

(Radiology, Mauro Annarumma et. al., 1/22/19)

Here is another use of artificial intelligence in imaging, this time as a sort of assistant. In this paper, researchers trained a natural language processing (NLP) system by importing 470,388 anonymous adult chest radiographs, including metadata like the level of critical damage to the chest shown in the images. A computer vision program enhanced with artificial intelligence used this data to learn what types of radiographs show the most dangerous injuries, rating the images from critical to normal.

Once the machine was trained, it was fed a further 15,887 radiographs and asked to categorize them. The machine displayed good accuracy for categorization very quickly, which means that it could be relied on to determine whether a scan should get immediate attention from a radiologist, or whether it could wait.

This study represents a seemingly modest use of artificial intelligence that could have a large impact. With too many patients competing for too few doctors, it is important that the most critical patients get seen the soonest. A few days could mean all the difference, so a program that cuts the wait for the sickest patients from 11 days to under three days is a huge improvement.

Each of these examples shows the promise of AI, but we still have a long way to go before most doctors will feel comfortable relying on computers to do their thinking. And that is a good thing. These programs need to prove themselves before we are willing to trust their judgement. In the meantime, computers will continue to learn and improve until that inevitable day when they take their place as the newest weapon in a doctor’s arsenal.

Mary recommends: AI in sci-fi

If this primer has you left you craving a good AI morality parable, you can’t do better than the underrated CBS series Person of Interest. While there are many, many, many movies and television shows about AI running amok, POI took a more nuanced look at what big data could mean in the real world. While the science section of this science fiction show was shaky, the writers did accurately predict many of the moral quandaries we now face in the age of social media data mining.

Trending topics is an occasional series focused on some of the hottest topics in drug discovery.