Artificial Intelligence in Health Care: Current Applications and Issues

Medical image analysis

In addition to radiology, the use of machine learning algorithms in medical image analysis has been expanded widely to most medical departments that use images for fields such as pathology, dermatology, cardiology, gastroenterology, and ophthalmology. In detail, machine learning algorithms use computed tomography (CT), magnetic resonance imaging, ultrasound, pathology image, fundus image, and endoscope data to diagnose or classify the severity of the disease.8, 16, 17, 18, 19, 20, 21, 22 When the machine learning algorithm was applied to the real-time colonoscopy, the accuracy of the diagnosis was 94% and the negative predictive value was 96% in analyzing 466 tiny polyps.23, 24 Among the many deep learning algorithms, the convolutional neural network algorithm, which retains high performance in image pattern analysis, has proven to be beneficial in analyzing medical images with complex patterns.4, 25

In the medical industry, Siemens Healthineers has developed AI-based AI-Rad Companion Chest CT software to assist chest CT diagnosis26, and GE Healthcare is also working on the development of AI-based medical image analysis technology. Further, Philips Healthcare has developed IntelliSpace Discovery, an open platform for AI development and deployment,27 and is working to commercialize its IntelliSite Pathology Solution in the digital pathology diagnosis field.27 Arterys has received FDA clearance on its Cardio AI, Liver AI, and Lung AI software to establish Medical Imaging Cloud AI platform.28 Apart from the abovementioned examples, various companies, such as Zebra Medical Vison and Aidoc, are working to commercialize AI-based medical image analysis tools. In Korea, a number of startups such as Vuno, Lunit, JLK Inspection, and Deepnoid are in the process of commercializing AI-based medical image analysis systems by receiving approval from the Ministry of Food and Drug Safety.

Smart IoT devices, signal, and in-vitro diagnostic analysis

Various technology giants such as IBM, Google, Apple, and Samsung are competing to develop and commercialize devices and services that can assist in improving the user health by acquiring health information from daily life using a combination of Internet of Things (IoT) technologies and wearable devices. In 2017, Apple installed a deep learning algorithm that detects atrial fibrillation on their smartwatch and received FDA approval.23 Using the photoplethysmography and accelerometer sensors, it learns the user's usual heart rates at rest and at activity, and sends a warning sign if there is a significant difference from the expected values. The deep learning algorithm also shows excellent accuracy in analyzing electrocardiogram (ECG). In a recent study, 91,232 single-lead ECGs were analyzed using a deep learning algorithm and showed high diagnostic performance similar to those of cardiologists.29 Applying the ECG pattern analysis algorithm to smart devices would be especially useful for patients with cardiovascular disease or chronic kidney disease with high blood potassium levels.

Global companies are developing other innovative medical devices as well, including non-invasive glucose meters.30, 31 Such bio-signal monitoring devices can be used for disease management and treatment in different hospital environments, such as intensive care unit, operating room, emergency room, and recovery room, in addition to being used for predicting and diagnosing diseases based on the accumulated real-time information. One prime example of this is the collaboration between Medtronic and IBM in developing Sugar.IQ by combining Medtronic's continuous glucose monitoring and IBM's AI Watson.32 The Sugar.IQ has been proven to directly help diabetic patients by reducing the number of hypoglycemia and hyperglycemia symptoms.33 In addition, PhysiQ's pinpointIQ device, which recently received the FDA approval, helps in preparing for sudden and fatal situations by continuously measuring the patient conditions from the wearable biosensors and by detecting subtle changes in the conditions.34 Philips has also commercialized its Connected Care Solutions, which is an IT solution that enables the medical staff to monitor the patient conditions from anywhere including the emergency room, recovery room, and nurses station, through tablet PCs and smartphones.35 Meanwhile, findings from studies that reported improvement in diagnosis sensitivity and specificity by applying machine learning to the multi-biomarker analysis have been used in real-world cases to classify the cancer-related biomarkers; moreover, machine learning has been applied in cases of in-vitro diagnosis of diseases such as tuberculosis and dementia.36, 37, 38, 39

AI using electronic medical records (EMRs)

Currently, various attempts are actively being made to develop AI systems using EMRs. In the U.S., EMR companies including Allscripts, Athenahealth, Cerner, eClinicalWorks, and Epic are conducting research on optimizing hospital treatment processes using AI technology.40 In Korea, EvidNet is working to develop multi-hospital clinical big data analysis technology based on observational health data sciences and informatics and common data model (CDM).41, 42 Further, IBM has developed Watson for Oncology, which provides optimal personalized treatments for cancer patients, and its clinical trials are underway.43 As for the Korean companies, Selvas AI and Linewalks are conducting similar studies.

Issues of utilizing health care data

In most cases, health care data typically include personal identification information such as personal code, number, text, voice, sound, and image. To create a data-driven AI medical device, a large amount of these data carrying sensitive personal information are required, but obtaining such sensitive information may lead to legal issues regarding personal privacy.44, 45, 46 The widespread use of wireless AI devices in health care requires a new technologies such as IoT and cloud computing, which enables to deal with the processing and storage limitations in such devices.47 However, cloud-assisted AI devices can give rise to serious security concerns to the private health care data. To solve this issue, technical research and attempts to change laws and regulations are in progress. In terms of technical research, various encryption technologies and de-identification or anonymization technologies of removing identity information are being developed. A popular example is the CDM-based distributed research network, which has recently received considerable attention in Korea.48 Furthermore, a number of privacy-preserving data mining technologies such as the federated learning and homomorphic encryption have been developed in the U.S.49, 50

Many countries worldwide are currently forming institutional and legal systems to address the conflicting interests between the use of health care information and the protection of personal information. In the U.S., the Health Insurance Portability and Accountability Act (HIPAA) established in 1996 has given individuals the data rights for medical information copies, and the Blue Button system has been established to allow individuals to diversify the use of data through viewing their own personal health records online. According to the HIPAA guidelines, anonymization on 18 identifiers of protected health information has been established to efficiently facilitate the use of health data.51 In addition, through the enactment of the Health Information Technology for Economic and Clinical Health Act (HITECH Act) in 2009, an electronic health record has been developed and promoted to increase the interoperability of medical information between hospitals.52 Furthermore, the Centers for Medicare & Medicaid has launched the MyHealthEData and Blue Button 2.0 services in 2018 to enable patients to access and control their medical records and other health data. In Europe, through the General Data Protection Regulation established in 2016, the basic individual rights for personal information have been reinforced by mandating the EU members to protect personal information in accordance with the six data protection principles.53 In Korea, efforts have been made in recent years to promote research on big data by providing rights to collect and use health care data based on the revision of the Bioethics and Safety Act. However, despite these efforts from many countries, no country has been able to systematically resolve the privacy issues regarding health care data.

Regulatory affairs and policies for new devices

Most AI-based medical devices exist in the form of software, and they are generally new devices different from the traditional devices in terms of regulatory affairs. Hence, new policies need to be established to approve and regulate such devices. The International Medical Device Regulators Forum has categorized these AI software intended to be used for medical purposes as “Software as a Medical Device (SaMD)”.54 In the U.S., the Digital Health Unit was established by the FDA's Center for Devices and Radiological Health in May 2017 to promote the expertise in digital health care device approvals and regulations, and the FDA has announced the guidelines for SaMD in December 2017.55 The FDA acknowledges that the current regulations for traditional medical devices are not suitable for SaMDs that are faster in development and modification.56 FDA has recently developed a Software Precertification Program, which enables faster marketing of SaMD through a developer-centered certification pathway unlike the existing pathway centered on individual products. Manufacturers who achieved ‘organizational excellence’ in this pathway can obtain exemption from premarket review for low-risk products. Japan, in accordance with the AI medical development plans announced in 2018, is planning to create comprehensive rules governing the use of AI in medical devices to minimize the existing AI medical device–related disputes and prevent the resulting R&D hinderances.57 Lastly in Korea, the “Approval and Review Guidelines for Big Data and AI-based Medical Devices” and “Review Guidelines for Clinical Effectiveness of AI-based Medical Devices” were announced in 2017, making them some of the first AI-related approval guidelines in the world.58 However, the standardized review index for the safety and effectiveness of the AI-based medical devices is still lacking worldwide.4

Safety and liability issues

The report on AI published by the U.S. National Science and Technology Council emphasizes on the fairness and safety to prevent any discrimination or failure and to prevent AI from causing unintended consequences (Fig. 1).59, 60 For example, if the AI is developed to have a bias toward a specific population group, a mismanagement could occur in the prevention or diagnosis of a disease and thus discrimination could arise in which a specific population group is excluded from benefits. To address this, the federal government is promoting verification of the effectiveness and fairness of AI through the evidence-based assessment, and the federal government funding for research is mandated to be allocated based on the transparency, effectiveness, and fairness. Furthermore, the government is recommending the universities and secondary schools to include topics related to ethics, safety, and privacy in the AI or data science curriculum. The council has also highlighted the cyber security related to AI and emphasized on establishing responsible strategies and plans at a federal level, including R&D for the sustainable security system development and operation in response to cyberattacks.

Fig. 1
Research and development strategic plan of artificial intelligence.60
AI = artificial intelligence.

• Click for larger image
• Download as PowerPoint slide

The present health care system assumes that all responsibility lies in the hands of the medical staff in the event of a medical accident. The AI-based medical technology may affect the judgment of physicians in various areas and may sometimes cause negative impacts, resulting in medical accidents. In such cases, liability issues would arise, and in the current health care system, it is highly likely that the medical institution or physicians who ultimately introduced the AI-based medical technology would be responsible for the case. Hence, physicians need to learn how to better utilize and interpret AI algorithms and be aware of potential legal consequences associated with AI use in medical practice.61 In addition, efforts in academia and policymaking should be made to straighten the liability issues and to evaluate the medical accident risks based on the various characteristics of AI technologies.62 Specifically, new policies should be introduced for the establishment and operation of the AI monitoring centers in medical institutions and a national level safety monitoring center for monitoring the safety of the AI-based medical technologies. In addition, a system for measuring the liability and strengthening the awareness of patients and medical staff on medical accidents that may take place when applying the AI-based medical technologies should be established.

Balanced application with existing health care systems

Applying a newly developed AI technology to real-world health care service can lead to unexpected problems. Therefore, the introduction of AI devices should be in harmony with existing health care systems, and the performance of AI devices must be monitored periodically. It is also important that AI devices should be easy to use and familiar to medical staff and patients to avoid any misunderstandings and errors when making medical decisions.

As the AI technology in its nature mainly relies on data, performance changes may be found when the pre-applied AI technology is retrained with the desired field data. This change may not always result in a performance improvement; rather it may result in a performance degradation.63 In addition, the AI device performances may vary when the distribution or severity of patients in the institution changes depending on the differences in social, economic, and medical environments. Although the current laws and regulations in Korea do not allow the use of field data in improving the AI device performances, considering the nature of the technology, retraining using field data could be approved in the near future provided certain conditions are fulfilled. Accordingly, the performance of AI devices should be periodically checked even after the clinical application to prevent any unexpected performance degradation or malfunction.

Considering the importance and complexity of the modern health care field, the AI technology should be applied to health care as naturally and seamlessly as possible without causing excessive changes in the current medical practices. For this reason, it is necessary to implement the interaction and interface technologies that can enable the medical staff to apply AI technology to the medical field in a natural way even if they do not directly understand the technical aspects of the AI devices. These technologies can be implemented in the form of conversational AI, voice recognition, real-time recommendation, monitoring, and various visual overlay technologies. At the same time, careful considerations should be given to these user interface elements when developing and integrating AI technologies into health care.