August 11, 2021
Technology and the healthcare industry have not always been happy bedfellows. Given the privacy associated with patient data, security concerns are paramount - and with good reason.
A recent ransomware attack by Eastern European hacker group Ryuk targeted over 235 medical facilities in the US, costing medical institutions in Las Vegas, Oregon, and New York more than $100 million.
It’s no wonder then that the industry is cautious about adopting more innovative technologies such as artificial intelligence (AI) or cloud computing.
More tech can mean more access points and subsequently increased risk to the confidentiality of patient data. Yet, conversely, healthcare is an industry that would reap significant benefits from AI adoption.
Harnessing AI in healthcare is possible but if the true gains are to be realized, we simply cannot rely on the cloud.
Where the cloud can fall down
Training AI using cloud infrastructure has been a longstanding process for any industry wanting to harness this technology. While in some sectors, this standard can be an effective method of producing a functioning AI that improves processes, for healthcare there are inherent flaws with the technology.
Privacy issues are the greatest barrier to adopting AI in healthcare. Patient information is too personal to be shared outside the hospital it is stored in, meaning relying on the cloud to train the AI is fundamentally impossible.
Some developers have tried to find a workaround - scrubbing the data of any personal details before uploading it to the cloud, but this is not enough. Anonymized data may seem like a good solution, but it lacks the full picture that the AI needs to make informed decisions.
Keeping things local
To remove these concerns around security, data quality, and the effectiveness of AI training, there needs to be a rethink on how AI is approached in healthcare. Data needs to be kept in the building to be secure and effectively train the AI for managing the unique challenges present in healthcare.
Naturally, keeping the data in the hospital instead of the cloud will overcome any concerns around confidentiality. Instead of being in the cloud and vulnerable to an attack, the data will remain in the building and accessible only with approved personnel. While this doesn’t completely stop the risk of a cyber-attack from taking place, it reduces the number of ways the data can be targeted.
With the data sets available, the AI will be trained to identify normal and abnormal studies in a specific procedure - for example, a chest X-ray. This process would be iterative, meaning that the AI will become more refined with each case being analyzed. By distinguishing between normal and abnormal X-rays, the AI can then be trained to identify specific pathologies and classify them.
Training the AI inside the hospital will also allow it to factor in the background of a patient as an abnormal X-ray for a male in their 20s will mean something different compared to an abnormal X-ray for a female over 50.
Through enhancing these processes, the AI will be able to aid hospital staff in carrying out these procedures. For example, an AI trained in conjunction with an MRI scanner will establish algorithms to simplify the process of taking a scan.
Once trained, the AI can aid the user in such things as the optimal place to position the patient, improving the image quality, or guiding less experienced doctors on using the machine. By being trained through the previous scans used by the MRI, the AI will then be able to detect, segment, and quantify the resulting images making for faster diagnoses.
Taking this approach with AI will also have a cost-saving advantage. Using the hospital's own edge devices, software and networking will allow the AI to be closer to the data and remove the additional burden of buying new hardware or investing in a cloud service.
Going beyond a single place
In the short term, an AI framework may focus on analyzing the data from a single hospital, however, the lack of dataset's variety may limit the ability to generalize a model.
Developing a concept that can be received from a few hospitals may improve the trained models for specific procedures or tools used at that institution. Moreover, once it is fully trained, this AI can be rolled out to other healthcare providers as well.
While each hospital will face its own unique challenges based on the location or demographic of patients, many processes can be improved across every institution and share the knowledge between one to another. An AI trained at a single hospital in combination with other locations can establish best practices for all medical institutions.
For example, it’s sometimes a challenge for medical professionals to identify which patients with a similar ailment – such as a broken bone – need immediate attention. An AI trained in line with an X-Ray machine as an edge device can develop models on which patients need the most immediate care.
It’s important to note that AI has the potential to become another effective tool for medical practitioners. As with any tool, it won’t remove the need for first-hand experience and expertise. Rather, the AI models can reduce the time taken on certain administrative tasks, improve the quality of certain procedures such as scans, and give more time back to doctors for patient care and research.
The cloud may be the ‘go-to’ for many industries when training an AI and many have tried to force the healthcare sector to follow suit. This could be challenging.
Now, with the option to train the AI within a single institution by feeding on other hospitals, the healthcare sector finally has the opportunity to take full advantage of this technology, retain the privacy of patient information and offer another tool for the medical community.
Nadav Israel is Edgify’s chief technical officer. He previously spent time as an algorithm engineer researcher for Samsung.
You May Also Like