Artificial intelligence (AI) used in medical devices
Artificial intelligence (AI) technologies are being used more and more boldly in areas that until now were reserved exclusively for humans. They are mainly used in decision-making, image analysis, speech, or natural languages analysis. These systems involve creating models of intelligent behaviour that can be used in computer programs and solve problems that cannot be solved using standard classical algorithms. This requires large amounts of data through which AI applications learn patterns and features that predict the output. This makes them different from classical algorithms, which have predefined rules and which AI creates on its own based on data provided to it. In medical devices AI causes operational speeding up of mechanisms and analysis of particular cases.
Artificial intelligence in medical diagnosis and prognosis
AI is developing rapidly and represents the future of medicine. It is already proving excellent at diagnosing various diseases. All thanks to years of research and the implementation of the ‘deep learning’ process in AI-equipped machines. Diagnostic tools that are supported by artificial intelligence improve the quality and efficiency of healthcare. They are used primarily in interviewing patients, analysing test results, monitoring the condition of the person being treated and performing many other activities by the doctor in order to make an appropriate diagnosis. Their correct functioning depends mainly on a large amount of diverse information, including patient-specific data.
Can prognosis be qualified as medical treatment?
The use of artificial intelligence in medicine is not only limited to diagnosing already existing diseases, but also in predicting their occurrence in people who have not yet had any lesions. Several AI systems are already being used to predict breast cancer occurrence. Startups such as Keheiron Medical with its “Mia Solution” system or Screen Point with its “Transpara” algorithm already provide broad solutions related to breast cancer prediction. Their solutions start at the stage of taking a photo, where algorithms automatically select the best image, through diagnostics, to recognising disturbing patterns that may indicate future lesions.
Algorithm as a medical device
Since we now know that algorithms can diagnose and predict the onset of breast cancer (in some cases even better than humans) can we classify them as medical devices under current EU law?
The legal instrument in the EU that regulates medical devices is the Regulation (EU) 2017/745 of the European Parliament and of the Council of 5 April 2017 on medical devices (MDR), amending Directive 2001/83/EC, Regulation (EC) No 178/2002 and Regulation (EC) No 1223/2009 and repealing Council Directives 90/385/EEC and 93/42/EEC.
Importantly in our case, the MDR defines “medical device” in article 2 of this Regulation as a tool, apparatus, device, software, implant, reagent, material or other article intended by the manufacturer to be used, either alone or in combination, in human beings for one or more of the following specific medical applications:
– diagnosis, prevention, monitoring, prediction, prognosis, treatment or alleviation of disease.
Strangely enough that this Article 2 of the Regulation is the only place in the entire Regulation where the word “prognosis” is used. It results from the fact that prognosis is a very delicate border of whether treatment starts here and whether prognosis can be indeed qualified as treatment.
The Regulation is the proper legal act to be used to determine the legal status of medical devices using software and their certification. The type of certification will depend primarily on whether the software is used as a stand-alone product or as software to control another device. In the first situation, the software will get its own certification class, in the second situation it will be the same as for the device.
The determination of the certification class will depend on a number of factors.
First of all, the regulation defines software as an active product, i.e. a product that depends for its functioning on an energy source other than the energy generated by the human body. This qualification makes the provisions of rule 10 of the resolution applicable to software, i.e. if we have active medical devices that are used for monitoring and diagnosis, they are by definition in Class IIa (medium risk device).
Rule 10 of the Regulation also states that if we have active products intended for ionising radiation and for diagnostic or therapeutic radiology – including devices intended for interventional radiology and devices which control, monitor or directly influence the performance of such devices – then they are in Class IIb (medium to high risk device).
The next relevant provision, Rule 11 of the Regulation, states that software that will provide information used in making decisions for diagnostic or therapeutic purposes is in Class IIa. The exception is software related to decisions that may cause death or irreversible deterioration of a patient’s health. Such software falls into class III (high risk device). If decisions may result in surgical intervention for the patient as a result, then such software is classified in Class IIb.
The ultimate software certification class, as can be seen, will depend on a number of factors and the purpose for which the device has been designed.
The legal challenges
Software for prognosis is a new method to combat illnesses consisting in the fact that modern medical device predicts the illness long time before the symptom of the illness determined by the medical device for diagnostic purposes. AI software and algorithms provide therefore new quality in medical prognosis. Due to AI there increases the probability of effective use of such device at the stage of prognosis, hence at the stage when the patient is still healthy (more or less 5 years before first symptoms of the illness occur). As can be seen, the science aims to create new prognostic functionalities for ready-made products that have been patented only for diagnostic purposes, namely for finding illness in a healthy human being and not for the purposes of predicting who among the patients will be ill in fact. Such change of functionality, leading towards predictions and going away from diagnostics may prove a great legal challenge. It relates not only to a legal use of such a product for medical procedures but also for liability for defects and procedural medical errors in using such device.
In a legal sense, the liability of artificial intelligence is most questionable. If we teach a machine something, it gets the ability to make decisions. If a faulty learning algorithm is created, the developer of the source code will be responsible for the damage caused by the AI. However, what if the algorithm is correct, but there is a hacking attack resulting in the “unreadability” of the algorithm.
A possible answer is the emerging concept of a third legal personality – i.e. besides natural persons and legal persons. In the European Parliament Resolution of 16 February 2017, the concept of “giving robots a special legal status in the long term” and applying electronic personhood when robots take autonomous decisions or their independent interactions with third parties” emerged. This implies the potential creation of a new legal category, distinct from natural persons, legal persons or crippled legal persons – ‘electronic persons’.