Group1: Ethical III - Machines in Diagnosis

Artificial intelligence in medicine

Several companies have launched themselves into research on artificial intelligence in medicine. Be it IBM's Watson Health, Google's DeepMind Health or similiar, the trend towards the automated data exploitation via "artificial intelligence" or "cognitive" computing, as IBM prefers to term it, is clear [1][2].

Watson Health

IBM Watson Health is a framework of tools developed to streamline medical research and treatment. This includes research in genomics, drug discovery, oncology and the development of several convenience tools for medical staff, such as a "Care Manager". IBM Watson Oncology for instance  is designed to aid doctors in their medical decision making by providing evidence based suggestions for a treatment. The patients information, symptoms and preferences can be entered via a convenient user interface and treatments are suggested to the doctor together with a confidence score. The treatment can then be reviewed by the professional, based on the retrieved evidence. The technique is based on text-mining tool that has been trained by experts [3][4].

[4]

Critics of IBM Watson oppose the hype around it and claim that the system is not much more than a text mining tool which cannot, in contrast to IBM's claims, "understand, reason and learn" or "outthink humans" [5].

DeepMind Health

DeepMind Health is a London based company focused on research of artificial intelligence (AI) research. Since 2014 the company is part of Google and now belongs to the Alphabet group. One of the products, that the company develops is DeepMind Health a technology aimed to provide better care. The system has different layers, which focus on different aspects of healthcare.

DeepMind provides an application "Stream" that is supposed to improve the standard of care in hospitals by rationalizing the stream of information between the different people involved, be it patients, nurses, lab technicians and doctors. Like this, the limited time that the caregivers have at their disposition can be spend efficiently and the response time in critical times is decreased, resulting in potentially outcome for severe cases. [6][7]

[8]

Flow of information and response mechanism in a usual hospital setting and in setting where Streams is used, as advertised by DeepMind Health.

Another area of activity of DeepMind Health is to exploit the data that is continuously fed to the system, to retrieve unknown pattens via machine learning/deep learning and ultimately improve the patients outcome on the long run, by providing better test, fairer access to good care, and gain more knowledge in general.

Ethical or evil?

While the intention to develop an app to streamline the patients care seems noble at first, there are obviously some concerns, especially with the data privacy of the patients, since it is their data which is being used for machine learning. NHS, U.K.'s National Health Service has signed a contract which involved sharing the data of 1.6 million patients. However, there where some mishaps in the process, such as the failed registration of the app as a medical product to the authorities. Furthermore, the extend to which data was shared with DeepMind and to which it was used, was considered "unlawful". Critics also argue, that the agreement is an easy way for DeepMind to acquire data, they would not be able to procure otherwise and call the attention to possible misuse. On the other hands many sources also acknowledge the possibilities, that data driven medicine has to offer [9][10][11].

Medical records in the cloud?

The integration of all medical data in a shared cloud has several advantages, from patients, medical practitioners, but also from researchers views. The medical history of a patient can be shared between clinical centers, thus helping doctors do their job better, because they have comprehensive information on a person's health status, from genomic information, to behavioural patterns. When using cloud based solutions researchers have access to unprecedented amounts of data that they can use for analysis and development of new medical solutions.

However, the main concern when bringing medical data to the cloud, is whether the security and privacy of the data can be assured. Questions have also been raised concerning the transparency of data handling by the (third party) cloud provider and the lack of evidence and experience in this new technology. Several crucial requirements for cloud based data storage solutions must be implemented, such as sufficient legislation, external control mechanism that ensure that the cloud service provider implements the security and privacy protecting measures. Furthermore, the trust of the patients must be acquired by being as transparent as possible.

Ultimately, the question whether it is acceptable to store all medical records in the cloud, comes back to whether we are able to protect all the sensitive information from unauthorized or unwanted access by third parties [12][13][14].

Machine Learning in medical diagnosis

In the last 20 years, the advances made in computer technologies and machine learning made it possible to analyse big data sets and incorporate new techniques into the medical diagnostic process. Machine learning and deep learning are part of the big trend of artificial intelligence and aim to diagnose patients based on the provided data and the underlying data set that was used to train the deep learning algorithm. The main goal of such an application is the acceleration of diagnosis and the elevation of accuracy, but also to train medical students or facilitate the diagnosis by non-specialists. [15]

Requirements

A machine learning application has to fulfil certain requirements in order to apply it in a medical context. In the following are listed the more important requirements: [15]

• Good performance: algorithm has to be able to extract significant information from the underlying data
• Dealing with missing data: algorithm must compensate for missing data and still reach good conclusion
• Transparency of diagnostic knowledge: algorithm has to be transparent in its explanations and knowledge in order to offer new point of views to the doctor
• Explanation ability: reached diagnosis has to be explained and the decision process must be understandable
• Small number of test: training data should be as small as possible and require a small number of tests in order to keep the costs as low as possible 

A future possibility to improve the reliability and the comprehensibility of diagnosis reached by applying machine learning is to not only use one single classifier, but to apply different algorithms to the same problem and compare the results. [16]

Acceptance in practice

Many physicians have concerns with systems that use deep learning to diagnose patients based on input parameters. Firstly, they are not comfortable with the inflexibility of knowledge representation in those systems. In many algorithms the input parameters are fixed and do not take into account subjective parameters on which doctors rely when diagnosing diseases. Informal or fuzzy notions a doctor has when examining a patient can’t be used in deep learning diagnostic systems.

Furthermore, doctors generally extend the diagnose if they have a doubt about their current diagnostic result, instead of using a system based on machine learning. In many cases, doctors don’t have the time to work with such a system during their regular work, since in many developed countries doctors have a high work load. Many other reasons have been collected to explain the slow integration of such assisting systems into the medical healing process. [16]

State of the art algorithms

The Assistant-R algorithm is a revised version of the Assistant algorithm and it uses a top-down induction of decision trees. The main difference of the Assistant-R to the Assistant algorithm is that it uses a heuristic measure of the attribute quality.

Lookahead feature construction (LFC) is also a top-down induction of binary decision trees in order to detect dependencies between attributes.

The k-nearest neighbours (k-NN) algorithm takes into account the k-nearest neighbours of the point that is being classified. This point is assigned to the class that is most frequent in the k-neighbours that are taken into account. [15]

Concrete cancer research

A new approach to diagnosing cancer is by comparing the genetic composition of regular tissue and cancerous tissue. This diagnostic procedure is very difficult, since it is hard to determine the difference between the two above mentioned tissues. This classification problem can efficiently be solved by the use of recent developments in the field of machine learning. The right choice of a proficient algorithm is based on the capacities to solve high dimensional classification problems with a low amount of available data. Many old approaches try to reduce the dimensionality of the problem without affecting the accuracy of the system. This can lead to overfitting due to the small available data sets. In general, such diagnostic algorithm are badly scalable in order to fit other cancer types.

The presented new procedure is not only using data from the cancer that is to be identified, but from other cancer types that stem from the same microarray. This improves the number of available data sets and and allows for unsupervised feature learning to be applied on this data.

The first step in this diagnostic process is to reduce dimensionality by applying PCA. The second step consist of using an autoencoder to classify the reduced data and in conclusion build and train the algorithm.

This approach has shown potential by outperforming traditional diagnostic methods, especially when there are only small datasets with a high dimensionality available. [17]

Legal Issues

AI and Machine Learning are becoming very spread and used in the current technological state. Starting from self-driving Tesla cars, to home assistants and industry. Another emerging area in which AI and Machine Learning are being incorporated is health care diagnosis and decision-making. However, this does not get incorporated as easy as in other areas, because of the human life factor involved. The introduction of AI and Machine Learning in diagnosis, has a very significant potential, because of the processing power, it may be able to create relationships between patient’s data that humans are probably not able to comprehend. However, this raises a lot of legal and ethical implications, as doctors may peruse a wrong proper treatment suggested by the algorithm. Currently, in the United States, when improper treatments are given to patients, doctors are the ones directly charged with negligence or recklessness. But, who is responsible in case the therapy or decision is made by a computer, what are the legal and ethical implications [18][19].

The answer for the questions above depends, considering the advancement and accuracy of the algorithms. Nowadays, these algorithms have not reached an accuracy rate higher than the average doctor. As the article discusses, when the time comes where the algorithms outperform doctors, proceeding with the decision of the algorithm will be always statically better compared to a doctor’s decision. Even if it turns out that the treatment is harmful to the patient, the article defends the doctor and the algorithm in such cases. Another area in which medicine would profit out of decision-making algorithms is defensive medicine. As it has been described in the article, many doctors claimed that they have prescribed more tests than needed for the patient, in order to be protected from any lawsuit for negligence. The costs of these unnecessary tests from patients turned out to be in the same value as the whole United Kingdom spending’s annually on health care ($210 billion). As such, the benefits of making use of advanced algorithms from AI and Machine learning, would significantly lower the cost of these test and possibly predict accurately for each patient [18][20].

Final decision or can doctors be replaced?

It is undeniable the fact that technology advancement has affected every sphere of our lives. There are significant benefits from this advancement overall, including the medical healthcare sector. Considering the industry and especially the healthcare sector, expert radiologists are facing the most threats from the possibility of replacement [21]. According to the source, there are highly trained, mathematically bulletproof systems (robots), which can do pattern recognition and identify in even very delicate cases signs of diseases. As for medical healthcare sector, robots are becoming a very important aspect to consider. Because of the computational power, and the capacity to go through data, even professional doctors have admitted that the results are astonishingly better and in greater detail than a human mind can consider. This is because there is a need to guide and have a human touch in the data that is transmitted to the robot, similarly to Doctor Herbert Chase saying "we need machines to make us smarter" [21][22].

The game-changing strategy would be a robot that would eventually perform a surgery on its own. According to [21], it is reported that "DaVinci" surgical system already allows doctors to perform online surgeries, even in further away distances. This can be considered as a more serious "threat" of robots replacing doctors. We can consider ourselves in the process of technology evolution. However, when it comes to a direct decision whether robots can or cannot replace doctors, it is almost impossible not to have doubts. Even though the advancement of artificial intelligence is outstanding, there will be always a need for a mediator between the robot and the patient, which is going to be a doctor. There is a significant work being conducted on the IBM's supercomputer "Watson" [23][24]. As for now speaking, Watson is being upgraded in order to reduce the number of deaths caused by medical errors. This robot can evaluate extensive amount of data in superfast ways and suggest a detailed medical report of illnesses which are inherited from family, have a cause in the specific patient or even are similar to the literature. After evaluations of the data of the patient, his/her families and the literature, for each of the hypothesis and medical reports, "Watson" supports it with a percentage of confidence. At the end, it outputs the answers which are again evaluated by a doctor [23][25].

IBM Watson in Healthcare [26]

Robots are already doing a lot of work in the medical healthcare; however, there is always the need of a human touch on feeding the data to the robot. As the example of "DaVinci" Surgical Systems, or even "Watson", there are thousands of robots conducting medical processes, which not far time ago, humans used to do them. Probably is better to ask the same question differently, not "can robots replace doctors?" but "how can robots replace doctors?", and which one is more accurate, probably we will never find out!