In previous articles we have already talked about the importance that Machine Learning is acquiring in many sectors, both in the academic and corporate world. Today, thanks to this technology, great advances have been achieved, such as autonomous cars, automatic facial recognition or the detection of fraudulent web purchases.
However, in the healthcare sector there are a number of peculiarities that make the development of these technologies much slower and somewhat more complex. In this article we will review 7 peculiarities of Machine Learning in the healthcare sector.
One of the great challenges for the applicability of this type of algorithms in the healthcare sector is the risk involved in decision making. Machine Learning models can be used for both healthcare management and clinical aspects. Many of the challenges that are carried out in the second area, the clinical purpose, include problems such as early diagnosis, risk prediction in a patient, treatment recommendation, personalized therapies, etc., so that an error or misuse of the model could affect the patient's health. Sometimes, the reasons that lead the clinician to make a decision are beyond the scope of the model as they are not present in the data used to train it, so the knowledge and supervision of health professionals is necessary for the final decision.
Another of the great peculiarities of the healthcare field is the need to ensure the privacy of the data, since it is extremely sensitive data. These personal data, such as medical tests, treatments, diagnoses, etc., are specially protected by the Organic Law on Personal Data Protection and guarantee of digital rights. The processing of this type of data always requires prior consent and proper anonymization, it is also important to always know where they are stored and who can access them.
The healthcare sector is one of the sectors in which the use of technologies such as Machine Learning has the greatest potential and value, given that these models base their potential on the volume and richness of the data with which they are trained, and the amount of medical data generated every day is immense.
However, the availability of these data is one of the main limitations that prevent us from reaching this full potential. The reservations about making datasets available, even properly anonymized, in the medical field, given their particular sensitivity, is often a major barrier to the development of these technologies.
In addition, these data are often distributed in repositories of different and unconnected information, sometimes distributed between public and private entities. Medical information is not always stored in the same way and there are often many uninformed records and databases that sometimes have not received the proper data cleansing and quality assurance processes. A Machine Learning model is only as good as the data used to generate it, so this is a very significant problem.
On the other hand, the interpretability of the model acquires special importance when it is applied for clinical purposes. The 'black boxes' of algorithms, which we have discussed in previous articles, make Machine Learning models difficult to interpret. In this sector an error can put a patient's life at risk, as we mentioned in point 1, for this reason it is important to extract explanations that lead us to understand the decision of the model, in addition to allowing the clinician to realize inconsistencies that allow him not to take into account certain recommendations of an algorithm in certain situations. Therefore, in Horus ML we believe that this is one of the most important lines of research to follow for the implementation of Machine Learning in healthcare and you can read here how we try to solve this task, and also our comments on the difference between interpretability and causality.
Another of the most difficult challenges is the actual implementation of the model. Although research on Machine Learning models in healthcare is becoming more and more frequent, from diagnosis from X-rays to early estimation of certain tumors, when it comes down to it, these models require exhaustive validation by healthcare experts to allow us to trust their predictions.
In addition, the great variability of data sources, the differences in clinical registries, and the variance in casuistry and performance between different entities, add extra difficulty when implementing the same model within different hospitals.
The application of Machine Learning in areas of great social relevance should aim to be fair, and ensure that the use of these algorithms does not affect patient treatment due to biases in fairness (such as gender, place of origin, etc.). Fortunately, this ethical aspect of Machine Learning is undergoing an important boom in recent years, and there are already several techniques to control that these do not suffer from biases of this type, as we will discuss in more detail in a future article.
Despite this, the special sensitivity of data and decisions in the healthcare sector makes it necessary, or at least advisable, to have ethical committees to define the responsible use of artificial intelligence.
Not everything are extra difficulties or complications when we talk about the use of Machine Learning in healthcare with respect to other sectors, so we would like to finish with a positive factor and that is the main reason why from Horus ML we have decided to implement this technology in our products, and this is none other than its greater positive social impact.
The application of Machine Learning in this sector can have a great impact on the quality of healthcare worldwide, supporting clinicians and also healthcare managers in their day-to-day work. This can result in better efficiency in the use of available resources, as well as in direct improvements in diagnosis or treatment and, in general, in the quality of patient care.
On the other hand, we cannot forget that these technologies also serve to provide support in areas with fewer resources. The application of these models helps to improve diagnosis in areas where there is a lack of healthcare personnel, as well as optimizing the management of the limited resources available. In addition, the use of virtual assistants or healthcare monitoring are techniques that can be used remotely to increase the quality of patient care even in areas with less healthcare potential. In this way, Machine Learning algorithms can serve not only to improve healthcare services per se, but also to homogenize the quality of care as a whole.
