(Artificial) intelligence meets heart - How machine learning changes cardiology

Intelligent software solutions, this seems to be certain, will not replace cardiologists. They will, however, replace processes in cardiology that are ubiquitous today but suboptimal due to the limitations of the tools. And it will be the cardiologists who will benefit from these developments. We had a look in the development labs of innovative IT companies to better understand the processes that will be affected, the tasks software will take over and what machine learning has got to do with all this.

From an IT point of view the crucial question concerns the role machine learning* will play for enhancing the quality of processes and the quality of care. The answer to this question explains the characteristics of intelligent software: the way bits and bytes are trained to recognize patterns and make decisions. The logical next question is “How do we measure the results of AI?”

It’s about understanding, not about learning by heart

Very often validity of the data is considered the only quality criterion of neural networks for machine learning. Valid data by itself however does not make a machine learning system. Here artificial intelligence works just like natural intelligence: a not so clever student who simply learns the cases by heart without understanding the underlying principles will never be a good physician, no matter how ‘valid’ the data might be. Really intelligent software thus is based on networks that learn to understand the principles. Consequently, above and beyond the case data,  contingencies have to be culled from and trained for each clinical case in order to be able to identify trade-offs and correct errors.

Moreover: absolute truths are very rare in medicine. Three cardiologists will arrive at three different measuring results and interpret the very same ultrasound image in three different ways. The software has to be able to handle this kind of variances. Its crucial feature is the fact that unlike the human brain it does not produce statistical abnormalities. Thus the network has to be fed with many interpretations of a given exam in order to be able to arrive by itself at the majority opinion. And it has to be in constant competition with human intelligence. In software development, the clinical partners are not only case delivery services, they are permanent sparring partners.

Whoever entrusts the human heart to the bits and bytes of cardiac software should have a very close look at the underlying machine learning methodology – because this methodology separates average from very good software.

Less routine, more time for demanding patients

An important field of application for any well-trained software is conducting routine exams and preparing the results for human reading, for instance follow-up exams. When cardiac performance is regularly tested, as might be the case with cancer patients, good software can automatically perform all necessary measurements. It uses prior exams, chooses suitable images from the image pool and measures in the appropriate planes. Furthermore it compares the measurement results with the reference values from similar studies. For the subsequent human reading the software offers the physician a selection based on relevant, maybe even critical content and/or a simple standard finding. To finalize the exam the physician will be guided through a control process to confirm the findings and release them. In clinical routine such automation saves a lot of time and enhances the reporting quality.

Automation enables true mobility

Let’s be honest: mobile ultrasound machines are handy, but without the ability to measure and evaluate the data that were generated ‘on the go’, the potential of mobility is not fully realized. The automation capabilities of an intelligent software package close the gap between generating and evaluating data as they render obsolete the many workflow steps that have to be performed manually on a (small) touchscreen, be it the selection of the appropriate image series, zooming or setting the measurement points.

Another important aspect in this context is transferring the data to the correct storage location, either in the cloud or on a server. The storage space of mobile devices is often too small for long-term storage of complex exams. Not to mention the fact that measurements require enormous amounts of energy and computing power which would make both battery and processors go up in smoke. Thus ‘outsourcing’ data processing is a technical necessity.

However there are process arguments in favor of mobile use: in healthcare networks the results of measurements conducted on mobile devices are available across the facility in real-time. This allows quick decision making among physicians and across departments and ensures seamless diagnostic processes. Thus fast diagnoses become possible, even in decentralized structures. A hospital does not have a cardiology department or requires the knowhow of an external partner? No problem since this third party has easy access to all relevant images via a mobile device or the browser – any time and any place. With the images being presorted and measured! Another scenario: the emergency physician transmits the ultrasound images from the site of the accident directly to a central storage space and by the time the ambulance arrives at the hospital the cardiologist on call has already reviewed all relevant images, measurement results and findings.

Quality assurance and reproducibility

Automated processes, particularly in routine tasks, improve quality assurance since they make sure exams are always performed and documented according to a pre-defined pattern – which is not always the case with exams performed by humans. Lack of time or experience can frequently lead to very different results with regard to both diagnosis and documentation, which my compromise the quality of care. Automated processes iron out quality differences for example by pointing out missing image series or by consistently using identical measurement criteria. Thus intelligent software can ensure that within a given facility or network binding standards are established which facilitates compliance with certain treatment paths.
This is good news for the hospital management which relies on impeccable quality assurance and documentation in the communication with insurers or lawyers in case of litigation.

Even today a number of cardiac measurements can be automated and they already improve diagnostic quality, e.g. ventricular measurements or flow measurements. Future software generations will be able to use neuronal networks to arrive at decisions faster, in a more comprehensible, reproducible and scientifically sound manner. And they will relieve
cardiology staff from drudgework so they can focus on their core competency: caring for the patients.


*Currently TOMTEC is in the process of developing such AI-based solutions; to date they have received neither CE certification nor approval for patient use.