Since the dawn of the millennium, the amount of digital data that is produced has risen exponentially. This huge amounts of data and algorithms generated based on this data has helped to make leaps in technology in the last decade. This progress has been relatively slow to come to healthcare for a variety of reasons, some of which will be touched upon in this article. As electronic records and digital storage become the norm; large amounts of data is being generated and there are many opportunities to take advantage of. Once this ‘Big data’ is properly analysed, it holds the key to making new discoveries and insights. In medical research especially, these new discoveries are validated by the high sample sizes that are powering these findings.
Artificial intelligence is the development of algorithms that can learn from large amounts of data and then run in an automated manner. There are examples of this already in use with University College Hospital developing an A.I algorithm that highlights which patients are most likely to not attend appointments, and target these patients with extra reminders.
Current literature suggests that imaging is leading the way in implementing artificial intelligence in healthcare. This makes sense since there are large amounts of standardised data for an A.I algorithm to learn from. We have studies demonstrating that A.I can be as good as senior clinicians in diagnosing skin cancer, lung cancer and many others diagnosis that are made through imaging.
Cardiac imaging in particular could be transformed with artificial intelligence as there are large databases of images for an AI to be trained on. In fact an oxford company known as ultromics claims its AI algorithm is more accurate than clinicians in analysing stress echocardiograms. This company boasts a 90% accuracy rate compared to 80% with clinicians interpreting. Artificial intelligence could also help in automated analysis of the more mundane aspects of reporting in cardiac imaging. A good example is in cardiac MRI, where an AI algorithm can be trained to reliably perform the contouring of the left and right ventricles, allowing the clinician to focus on the diagnostic aspects. This would certainly increase efficiency and output of any department.
Artificial intelligence could be used to address some of the shortfalls in medicine as well. An example would be to train an algorithm to correctly identify pathology specimens once it is trained on a database of known pathology specimens. This could help in clearing the backlog and the delay that there is in cancer diagnosis due to a shortfall of trained specialists.
The use of artificial intelligence is giving rise to new fields in medicine such as radiomics and precision medicine. The idea behind radiomics is that there is information in images which is not visible to the naked eye. Data such as the texture, shape and intensity of a pixel hold signatures of the disease pathology that we cannot see. This data can be uncovered using AI. It is already being used in oncololgy whereby an algorithm has been mining data from medical imaging database and also trained in the associated patho-histological diagnosis. This algorithm can now accurately and reliably detect the type of cancer from a CT scan. The potential for more applications in other areas of medicine is endless and still to be realised at this early stage in the development of this field.
Furthermore, artificial intelligence can be the key for healthcare to evolve towards precision medicine. The main area where AI can help is when it comes to genetic analysis, which had been relatively slow due to the challenges of analysing such large amounts of data. However, with AI algorithms being trained, this area of medicine is expanding rapidly with new discoveries and association being made on a daily basis. This information will eventually filter into clinical practice and will dictate the most appropriate treatment options for a patient based on their genetic profile as well as the clinical.
Limitations and challenges
The NHS in general produces huge amounts of data which is ideal to train algorithms on and potentially has the infrastructure to implement the applications. However, an AI algorithm is only ever as good as the data that it is trained upon. Therefore, if there are errors or biases in the data, that will affect the efficacy of the AI. Most of the medical research is using databases which are predominantly consisting of Caucasian patients/ participants and therefore there will be limitations in applying the results to the wider population. A real-life example is the face recognition technology recently trialled by many companies including amazon, which has been criticised for not being able to effectively distinguish between people of darker skin tones. This is because the algorithm has predominantly been trained on those of Caucasian descent.
Another major challenge is cybersecurity. The cyberattack in 2017 proved just how vulnerable the NHS cybersecurity is. In order to truly create a large wealth of data to train A.I on, we would need to potentially share or pool the data between different NHS trusts and with industry organisations. However, if security concerns cannot be addressed this will be a major limitation.
In addition, we need to ensure that patient confidentiality is not compromised. In order to, develop AI applications healthcare organisation and industry would need to work together. This will involve sharing data with these organisations and there is a danger that confidential data may be inadvertently or illegally transferred. In fact an NHS trust has been criticised for sharing identifiable data of millions of patients’ in the process of developing an application with industry (google). This was not found to be strictly illegal, but raises new questions of what the legal framework is for this sort of collaboration.
Some will argue that the development of artificial intelligence will leave a large part of the workforce redundant and need to be retrained. A counter-argument would be that firstly, this technology is a long way from being implemented in this way in the near future. In addition, highly trained individuals would still be required to supervise, interpret and finalise the findings of an automated system. Therefore, the AI will essentially be enabling a much more efficient system, which will need to be considered when planning future work force requirements. In fact, the current issue is that we don’t have enough people who are trained healthcare professionals with working knowledge of how to analyse large data sets and development of algorithms that is limiting the applications of AI in healthcare.
Despite the limitations, the opportunities and the potential far outweigh the challenges involved. In order to take advantage of these opportunities we need to train our future medical workforce that can understand how to handle large datasets and be able to guide the development of AI in healthcare in order for it to be safe, secure and effective.
Remarxs is a powerful academic collaboration platform - to find out what we do, and join our network, click here