Scientific European (SCIEU)
Recent Articles
Monthly Popular Science Magazine
 87 views

Artificial Intelligence Systems: enabling fast
and efficient medical diagnosis?

Recent studies have shown the capability of artificial intelligence systems in
medically diagnosing important diseases


 

Artificial Intelligence (AI) systems have been around for quite some time and are now getting smarter and better with time. AI has applications is multitude areas and is now an integral of most fields. It is now strongly believed that AI can be an essential and useful component of medical science and research as it has immense potential to impact the healthcare industry.

Artificial Intelligence in medical diagnosis

Time is the most valuable resource in healthcare and early appropriate diagnosis is very important for the final outcome of a disease. The current healthcare is often a lengthy and a time and resource consuming process, delaying effective diagnosis and in turn delaying the correct treatment. AI can help to fill the gap between availability and time management by doctors by incorporating speed and accuracy in the diagnosis of patients. It could help to overcome limitations of resources and healthcare professionals specially in low- and middle-income countries. AI is a process of learning and thinking just like humans through a concept called deep-learning. Deep learning utilizes broad sets of sample data to create decision trees by itself. With this deep learning, an AI system can actually think just like humans, if not better, and therefore AI could be deemed fit to carry out medical tasks. When diagnosing patients, AI systems keep looking for patterns among patients with same illnesses. Over time, these patterns can lay the foundation for predicting diseases before they are manifested.

In a recent study1 published in Cell, researchers have used artificial intelligence and machine learning techniques to develop a new computational tool to screen patients with common but blinding retinal diseases, potentially speeding diagnoses and treatment. The researchers used an AI-based neural network to review more than 200,000 eye scans conducted with a non-invasive technology that bounces light off the retina to create 2D and 3D representations of tissue. They then employed a technique called ‘transfer learning’ in which knowledge gained in solving one problem is stored by a computer and applied to different but related problems. For example, an AI neural network optimized to recognize the discrete anatomical structures of the eye, such as the retina, cornea or optic nerve, can more quickly and efficiently identify and evaluate them when it is examining images of a whole eye. Such a process effectively allows the AI system to gradually learn with a much smaller dataset than traditional methods which require big datasets and is expensive and time-consuming.

This study focused on two common causes of irreversible blindness which are treatable when detected early. Machine-derived diagnoses were compared with diagnoses from five eye doctors or ophthalmologists who reviewed the same scans. In addition to making a medical diagnosis, the AI platform also generated a referral and treatment recommendation which not been done in any previous study. Thus, this trained AI system acted just like a well-trained ophthalmologist and could generate a decision within 30 seconds on whether or not the patient should be referred for treatment, with more than 95 percent accuracy. Further, they also tested their AI tool in diagnosing childhood pneumonia, a leading cause of death worldwide in children under the age of 5, based on machine analyses of chest X-rays. Interestingly, the computer program was able to differentiate between viral and bacterial pneumonia with more than 90 percent accuracy. This is crucial because though viral pneumonia is naturally rid by the body after its course, bacterial pneumonia on the other hand tends to be a more serious health threat and requires immediate treatment with antibiotics.

In another major leap2 in artificial intelligence systems for medical diagnosis, scientists have found that photographs taken of a retina (back of the eye) of an individual can be analysed by machine-learning algorithms or software to predict cardiovascular heart risk by identifying signals which indicate towards heart disease. The status of blood vessels in the eye which is captured in the photographs was shown to accurately predict age, gender, ethnicity, blood pressure, any prior heart attacks and smoking habits and all these factors collectively predict heart-related diseases in an individual.

The eye as an information block

The idea of looking at the photographs of the eye to diagnose health has been around for some time. It’s well established that the rear interior wall of the human eyes has a lot of blood vessels which reflect the overall health of the body. Thus, by studying and analysing the appearance of these blood vessels with a camera and a microscope, a lot of information about an individual’s blood pressure, age, smoker or non-smoker etc can be predicted and these are all important indicators of health of an individual’s heart. Cardiovascular disease (CVD) is the number one cause of death globally and more people die of CVDs compared to any other disease or condition. This is more prevalent in low- and middle-income countries and is a huge burden on economy and mankind. The cardiovascular risk depends on a multitude of factors like genes, age, ethnicity, sex, in combination with exercise and diet. Thankfully, most cardiovascular diseases can be prevented by addressing behavioural risks like use of tobacco, obesity, physical inactivity and unhealthy diet and making significant lifestyle changes to address the possible risks.

Health diagnosis using retinal images

This study conducted by researchers at Google and its own health technology company Verily Life Sciences, showed that an Artificial Intelligence algorithm was used on a large dataset of retinal photographs of around 280,000 patients and this algorithm was able to successfully predict heart risk factors in two completely independent datasets of around 12000 and 1000 patients with reasonably good accuracy. The algorithm used entire photograph of the retina to basically quantify the association between the image and the risk of heart attack. So, this algorithm could predict a cardiovascular event 70 percent of the time in a patient and in fact a smoker and a non-smoker were also distinguishable in this test 71 percent of the time. Further, the algorithm could also predict high blood pressures indicating heart condition and also predict the systolic blood pressure -the pressure in the vessels when the heart beats- within a range of most patients with or without high blood pressure. The accuracy of this prediction, according to authors is very similar to a cardiovascular check in the laboratory, where blood is drawn from the patient to measure cholesterol levels looking in parallel with the patient’s history.  The algorithm in this study, published in Nature Biomedical Engineering, could in most likelihood also predict occurrence of a major cardiovascular event -e.g. a heart attack in the future.

An extremely interesting and crucial aspect of these studies was that in both these studies the computer can also tell where it is looking in an image to arrive at a diagnosis, allowing us to understand the prediction process. Example, the study by Google exactly showed “which parts of the retina” contributed to the prediction algorithm, in other words how the algorithm was making the prediction. This understanding is important not only to fully understand the machine learning method in this particular case, but also for generating confidence and faith in this entire methodology by making it transparent.

Challenges

It must be pointed out that such medical images come with its challenges since observing and then quantifying associations based upon such images is not straightforward mainly because of several features, colours, values, shapes etc in these images. This study uses deep learning to draw out the connections, associations and relationships between changes in the human anatomy (internal morphology of the body) and disease in the same way as a healthcare professional would do when he or she is correlating patients symptoms with a disease. Also, these algorithms obviously require more testing before they can be used in a clinical setting. For instance, it’s still unclear how the various factors add up and, in some individuals, it may be anyway necessary to perform sophisticated tests, like coronary calcium CT scans to identify the risks.

Despite discussions and challenges, AI has huge potential to revolutionize disease diagnosis and management by doing analyses and classifications involving immense amounts of data that are difficult for human experts and it provides fast, cost-effective, non-invasive alternative image-based diagnostic tools.  The most important factors for success of AI systems would be more computational power and more experience of the people.  With each developing AI for medical diagnosis, healthcare delivery is improving significantly for potentially everyone. A probable future is visible in which new medical insights and diagnosis would be achievable with AI without human direction or oversight.

Source:

  1. Kermany, Daniel S. et al., 2018, ‘Identifying Medical Diagnoses and Treatable Diseases by Image-Based Deep Learning’, Cell, vol. 172, no. 5, pp.1122 – 1131. DOI: 1016/j.cell.2018.02.010
  2. Poplin, Ryan et al., 2018, ‘Prediction of cardiovascular risk factors from retinal fundus photographs via deep learning’, Nature Biomedical Engineering, vol. 2, pp.158–164. DOI:1038/s41551-018-0195-0

 

Download the full issue containing this article here

Subscribe to Scientific European here