Thanks for conveying your interest!

You will get notified on the email you provided.
twitter Connect with us on linkedin

Humans and AI working in tandem: a healthcare application

Jure Brence, December 2021

Healthcare is an area where artificial intelligence can have an enormous impact and a direct effect on people’s lives. AI can help with preventing avoidable errors and improving efficiency, enabling better care through novel methods, reducing unwarranted variation in clinical practice, extracting new insights from data and assisting patients in managing their health.

But can we expect artificial intelligence to replace doctors and automate healthcare entirely? Probably not, at least not anytime soon.

The adoption of AI in clinical settings has understandably been slow, in no small part due to the high standards of medical practice and procedures, the low margin for mistakes and the growing concerns over patient privacy. Implementations of AI in healthcare have to deal with a lack of high quality, representative data, the difficulties of generalizing across technical differences of equipment and local practices, as well as issues with models performing poorly when transferred from controlled lab settings to real world settings.

These are all well-understood facts and challenges, so I was quite surprised when a cardiologist excitedly told me about the AI they are already routinely using in their work. They’ve recently acquired a new machine for echocardiography, featuring AI that automates a number of cumbersome steps and brings considerable time savings to their work process. Inspired by the endorsement, I decided to investigate this example of a limited, but high-impact, real-world application of AI in healthcare.

Echocardiography is a type of medical imaging that uses ultrasound to obtain a wealth of useful information about the heart. It is one of the most widely used diagnostic tests in cardiology.

The first AI application involves automated measuring. In order to assess and record details on the shape and size of the heart, doctors measure quantities such as wall thickness and internal chamber diameters. These measurements follow medical guidelines and require the physician to first find the best position for the ultrasound probe, then locate specific structures on the 2D image and indicate them to the machine so that the required quantities can be computed. A single exam can involve more than a dozen such measurements, some of which have to be repeated several times to achieve the desired quality.

The approach chosen to address this task was to train a neural network algorithm to automatically recognize the landmarks. The model relies on convolutional layers, taking as input the grayscale image data, but with added x and y channels that provide additional spatial context. After the AI identifies the features and performs the measurements, the user must only review and validate them, and retains the option to perform them manually if needed.

The second application saves time by recommending measurements to the doctor. Doppler spectral imaging makes use of the Doppler effect to determine the direction and velocity of blood flow. When it comes to measurements on spectral images, conventional data analysis methods have already been used to partially automate the work. However, after the cardiologist sets the parameters and performs the scan, they still have to navigate an expansive menu to select the appropriate measurement algorithm. Since there are typically many such measurements to perform, the time lost in the selection process adds up quickly.

This process is streamlined by a neural network classifier, trained to predict which measurement the doctor has in mind by considering the scanning parameters they’ve set, such as the imaging mode and the baseline level, as well as the Doppler spectral image itself. If the prediction is correct, the physician accepts the recommendation with a single click, and if it isn’t, they delve into the menus.

Further speedups in echocardiography examinations are achieved through automatic view recognition. As the doctor performs scans and acquires images, a convolutional neural network recognizes which standard 2D scan plane they represent and attaches labes to the files. These are used later to streamline workflows, for example by selecting an image, suitable for particular measurements, or by automatically selecting trios of »apical« views with compatible heart rate and frame rate in order to perform strain analysis.

Overall, the implementations of AI in this particular echocardiography machine help reduce operator fatigue and ensure consistent measurements. Although sparing doctors »a few clicks« might seem a trivial result for state-of-the-art AI, it must be stressed that in the overburdened healthcare of today, saving time can mean shortening queues and potentially saving lives. Furthermore, by cutting down the time doctors spend interacting with computers, more of their time is freed to interact with the patient. In this way, AI helps bring the real human contact back into healthcare

I believe this example nicely illustrates a path towards adoption of AI in sectors of high skill, high responsibility and high pressure. Not by replacing experts, but by working in tandem with them, automating the tedious and time-consuming parts of their work, and enhancing their decision-making through data-driven insights.

Further reading

  1. The Role of Artificial Intelligence in Streamlining Echocardiography Quantification, Kristin McLeod. GE Healthcare, white paper.

  2. Real-World Application Of AI In Healthcare Is Still A Challenge, Shraddha Goled.