Agenda

Meeting Minutes

Ho, Chih-Hsing , Associate Research Fellow, Institute of European and American Studies, Academia Sinica

Dr. Ho’s presentation primarily focused on the personal privacy concerns in the field of AI in healthcare and caregiving, as well as the regulatory challenges posed by AI. Currently, AI has been extensively applied to digital pathology data, and the FDA has already established relevant regulations. However, there are still many applications awaiting development, with potential connections to emerging technologies such as big data and precision medicine. Our aspiration is to ensure that as AI enters the domain, it contributes benefits rather than harm, particularly in terms of data usage and accountability. Dr. Ho summarized several key issues, including health data integration, data reusability, consent models, de-identification, commercial exploitation, bias and interpretability, as well as trustworthy AI.

The basis for collecting, processing, and utilizing health data in our country is the Personal Data Protection Act. The primary principle outlined in this law is that data collection should have a specific purpose. Medical, genetic, and health examination data are considered sensitive personal information, which generally should not be collected, processed, or utilized. However, the Article 6 of the Personal Data Protection Act includes exceptions known as subclauses, which provide regulatory foundations for cases such as explicit legal requirements or written consent from the individuals involved.

Usually, commercial companies need to engage in industry-academia collaboration with educational institutions to obtain secondary use of data like National Health Insurance (NHI) data. The terms specify that only public authorities and academic research institutions can use health data for statistical or academic research purposes, and this can only be done when the data is de-identified to the point where identification is not possible.

Dr. Ho mentioned that the most challenging aspect is de-identification. Taking the example of pseudonymization and anonymization practices in the European Union (EU), the EU views pseudonymization as not a sufficient condition for lawful data processing, as there remains a possibility of re-identification. On the other hand, anonymized data is completely unlinkable to the original data. However, a common scenario in Taiwan is treating pseudonymized data as anonymized data for usage purposes.

Dr. Ho then proceeded to analyze Constitutional Interpretation No. 13, Judgment No. 111 (NHI Database Case). The judgment pointed out that the lack of explicit provisions in Articles 79 and 80 of the National Health Insurance Act regarding the use of data beyond the scope of its original purpose could potentially be unconstitutional. However, the final ruling did not immediately declare a prohibition on data release. Instead, it granted a three-year period for legislation or amendment to clearly define legal authorization for such matters. If corrective actions were not taken within this period, a mechanism would need to be established for individuals to request withdrawal from data use. This was aimed at addressing the issue of ambiguity in the current regulations concerning the use of data beyond its original purpose.

The trend observed from this judgment is that when the wishes of the parties involved, especially in the medical field, are not taken into consideration, there can be issues regarding the legality of regulations. Unfortunately, at present, there is only separate legislation specifically addressing the NHI database, and it has not been able to address the broader range of healthcare data in our country.

Taiwan has put a lot of effort into addressing the secondary use of medical data, drawing inspiration from the practices of the European Union. One approach is to limit purposes to the public interest, while another involves implementing secure de-identification measures. However, what Taiwan currently lacks is legal authorization for the secondary use of data. The current reform direction primarily relies on regulations related to the National Health Insurance (NHI) database and the establishment of separate laws, such as the Human Biobank Management Act, to govern data’s secondary use.

Dr. Ho discussed the issue of biases in AI algorithmic decision-making. She mentioned that biases can arise from using one-sided or partial data during algorithm development or from data collection and classification influenced by human subjectivity. Biases can also emerge due to a lack of proper oversight in the design process, leading algorithms to reflect and replicate human biases, resulting in algorithmic biases that contribute to health inequalities.

She provided an example involving a dermatology AI-assisted system utilizing Convolutional Neural Networks (CNN). When training the system using skin lesion images, with over 90% of samples coming from white patients, the diagnostic accuracy dropped significantly. Additionally, when algorithms equate healthcare spending with health needs, they might wrongly conclude that Black patients have lower medical expenditures, leading to the erroneous recommendation that they require fewer healthcare resources. The reality is that healthcare spending is lower among minority populations because they often lack the resources to access medical facilities, resulting in insufficient medical records. This can lead to what she termed “invisible patients” in clinical algorithm training databases, resulting in erroneous allocation of healthcare resources and policy decisions. The solution involves including minority populations such as immigrants, children, and the elderly as extensively as possible in the training data to avoid overlooking significant demographics that algorithms may fail to diagnose or treat effectively.

Achieving trustworthy algorithmic decision-making involves several aspects. From a regulatory perspective, it’s important to establish sufficient transparency in AI governance to serve as a basis for oversight. In turn, transparency can also foster trust between medical professionals and patients. But how should e address opacity? Opacity can be categorized into three levels:

1. Disclosure:
   At the first level, it’s about knowing whether the recommendations are generated by AI. Disclosing AI involvement helps individuals understand the role of automation in decision-making.
2. Explanation:
   The second level involves understanding why the AI system provided a particular recommendation. This includes considering the parameters or datasets that influenced the AI’s decision. Providing explanations can enhance the perceived legitimacy of AI decisions.
3. Comprehensibility:
   The most challenging level is understanding how the AI arrived at its recommendation. This requires comprehending the intricate workings of complex algorithms, which may involve technical complexities that are difficult for non-experts to grasp.

In conclusion, Dr. Ho emphasized that all these issues can be aligned with the 7 key principles of trustworthy AI established by the European Union. These principles include enhancing safety, improving data governance, respecting privacy, ensuring AI applications are more transparent, fostering inclusivity, enhancing accountability, and keeping a human-centric approach throughout the process. It’s important to continually review and assess whether the development and deployment of AI are centered around human well-being and ethical considerations. By adhering to these principles, the path toward building trustworthy and responsible AI systems in the medical field becomes clearer.

Fan, Chun-I , Professor, Department of Computer Science and Engineering, National Sun Yat-sen University

Professor Fan’s presentation mainly covered the threats and challenges of cybersecurity and privacy in the field of smart healthcare, as well as the technological tools that can be utilized. In the section on cybersecurity threats and challenges, Professor Fan first reviewed major global cyberattacks that occurred during the pandemic. The shift to remote work due to the pandemic led to increased cybersecurity risks, and vulnerabilities in home network environments created entry points for attackers.

In the post-pandemic era, geopolitical factors and conflicts have given rise to different cybersecurity issues. Sectors including finance, government, enterprises, and hospitals have experienced breaches due to political events. Currently, cyberattacks primarily target critical infrastructure, and the healthcare sector has become one of the key targets. On average, healthcare organizations take around 4 days to recover from cyberattacks.

IOT devices utilized in medical institutions are indeed one of the prime targets for hacker attacks. These devices often hold substantial amounts of sensitive personal information. Weaknesses in their original designs or a lack of security awareness among users can make them susceptible to hacking attempts. Professor Fan referred to Palo Alto’s analysis of over 200,000 IoT infusion pumps from 7 medical device manufacturers used in healthcare facilities. The study found that 75% of these devices had security vulnerabilities. This means that hackers could tamper with drug dosages without verification, posing serious threats to patient safety.

In the recent period, Taiwan’s Ministry of Health and Welfare has officially opened electronic medical records to cloud storage and allowed hospitals to go completely paperless. It’s crucial to pay close attention to the security threats posed by cloud environments. These threats can include vulnerabilities in virtual platforms, issues related to data migration and backup management, and attacks that exploit vulnerabilities across virtual hosts. As the healthcare industry transitions towards digitization and cloud-based solutions, safeguarding against these security risks becomes paramount.

Professor Fan stressed that cybersecurity is a cross-disciplinary issue with no “magic bullet” solution. Each organization should understand its systems and needs before formulating response measures. To address both smart healthcare cybersecurity and privacy concerns, he advocated for cryptographic-driven technological tools, emphasizing a data protection perspective.

AI usage is prevalent in healthcare institutions, and Professor Fan used Google’s Federated Learning model as an example. He explained that multiple local models can collaboratively train a centralized model on the cloud without uploading local data. This approach ensures data privacy for each local model. He mentioned the example of Taiwan AI Labs collaborating with the four major medical centers in Kaohsiung to train AI tools, demonstrating practical applications.

He further highlighted the risk of hackers reverse-engineering local personal data through model parameters. He proposed strengthening the security of Federated Learning through techniques such as homomorphic encryption and differential privacy. In essence, Federated Learning occurs within an encrypted context, enhancing its security while maintaining data privacy.

Professor Fan believes that data is the core of information security, and all security issues stem from data. Many modern technologies allow data to be processed while fully encrypted. Examples include IBE (Identity-Based Encryption) encrypted email services and searchable ABE (Attribute-Based Encryption) encrypted file-sharing services. He also introduced his recent research on “Privacy-Preserving Healthcare Data Mining Warehousing System.”

Professor Fan’s ultimate recommendation is to build a foundation of data protection, ensuring that data is encrypted to the greatest extent possible (maintaining data in an encrypted state most of the time). By addressing the core issues of data risk, organizations can then accurately establish precise security measures, similar to the concept of precision medicine. Subsequently, customized security solutions can be tailored for organizations, ultimately achieving low-carbon and sustainable security at minimal cost.

Presentation Download

**The speakers doesn’t agree to share the presentation file**