Abstract
Machine Learning is increasingly becoming utilised in various sectors from engineering to psychology, and new successful developments of machine learning indicate that these technologies could be beneficial in medical settings. However, the viability of these technologies is questioned, given the ethical and logistical difficulties in medicine.
Introduction
On the surface, machine learning (ML) is one of many branches of artificial intelligence (AI), whereby AI refers broadly to the development of machine capabilities. Arthur Samuel – an American pioneer in machine intelligence – defined machine learning in 1959 as:
‘a field of study that gives computers the ability to learn without being explicitly programmed.’ [1]
Here, the anthropomorphised term ‘learning’ in machine learning refers to the desire to create models that can learn like human beings, through experiences and evaluations, achieving objectives and creating outputs with minimal human assistance.
Structure of Machine Learning
In ML, unlike traditional computer programming technologies, there is no manual coding and once the framework is built for an ML model, it can learn patterns and rules independently, similarly to a human.
For example, to create a ML model to derive a differential diagnosis for abdominal pain, the series of decisions are not explicitly written into the computer. Instead, input-output data pairs (e.g., right lower quadrant pain is suggestive of appendicitis) are passed into the ML model, which learns the relationship between input and outputs. The feedback leads to a model that learns the important features automatically and generates the desired output. The result is the automated diagnosis of abdominal pain. The ability to adjust the function is the most notable aspect of ML models. The algorithm repeats evaluating and adapting its function and updates its rules autonomously until the required accuracy is met – this is how the automated learning occurs.
Previous Usage of Machine Learning
In the early 1970s, an ML system named MYCIN was developed to identify disease-causing bacteria and recommend antibiotics with dosages dependent on patients’ body weight. This was a significant breakthrough in medical ML, with higher accuracy and performance achieved than expected, however MYCIN was never used in practice.[2] Despite the development of successful prototype systems, most clinicians were reluctant to use these systems. Contributing factors included general distrust, concerns around accountability, and the great effort needed to keep the ML knowledge updated and current with the relevant science and clinical practices. Furthermore, the AI winter in 1970-1980 resulted in reduced funding and interest and subsequently fewer significant developments.[3]
Even after several innovations within other disciplines outside of medicine including the ‘first electronic person’ and the ‘first chatbot’, medicine was very slow to adopt AI. However, to establish the foundation for ML development, clinical information and medical records were first developed and digitalised.[4] The development of other sophisticated medical technologies, such as various imaging machinery and constant patient monitoring, has also led to an increase in the quantity of data from each patient. Even though medical recording systems are in place, healthcare systems struggle to integrate and analyse these datasets due to the increasing global population, resource shortages, and the sheer size and complexity of healthcare datasets. This contributes to an increase in medical diagnostic errors which are significant source of morbidities and mortalities, and unnecessary costs.[5]
Future Outlook
ML development has been markedly rapid in the recent years to aid medicine by analysing and classifying this clinical data, whilst outsourcing everyday tasks in medicine to technologies. ML technologies are driven by the expanding power of computer processing, the availability of large datasets and the financial input from private companies and governmental sources.[6] However, although ML is currently developing rapidly, there remains numerous unresolved challenges – some which have been identified in previous ML developments, as well as possible new challenges, such as data privacy and public bias.
It is important to identify the realistic potential of ML in medical diagnostics in modern clinical settings, as developments in the ML field have drastically progressed since the discovery of ML in the 1950s and challenges have not been thoroughly realised to date. Therefore, new studies should highlight the value of ML in diagnostics within the future and evaluate the true potential of ML benefits. With AI and ML becoming more available and capable, there is a need for further research into this topic, evaluating all aspects of ML in diagnostics in a broader viewpoint with acknowledgement of different stakeholders. The human and systematic effects are closely linked and identifying contributing factors will ensure that ML benefits can truly benefit patients and clinicians, whilst avoiding unnecessary costs and patient harm. The regulatory and ethical frameworks for ML must also be clarified so ML can reach clinical settings quickly and safely.
Swetha Babu, Youth Medical Journal 2022
References
[1] Samuel, A.L., (1959).
[2] Trivedi, M.C. (2014).
[3] Kaul, V., Enslin, S. and Gross, S.A. (2020).
[4] Kaul, V., Enslin, S. and Gross, S.A. (2020).
[5] Institute of Medicine (US) Committee on Data Standards for Patient Safety, et al. (2004).
[6] Barber, J. (2012).
Youth Medical Journal 