Screenshot 2023-03-07 at 2.48.00 PM.png

Source: dealroom.co

Artificial Intelligence (AI) has propelled medicine and healthcare in ways that were inconceivable a decade or two ago. Subfields including deep learning and generative AI transformed many areas in the medical field from patient care to diagnostics and drug discovery. Health technology (healthtech), the fastest growing startup vertical across the Middle East and North Africa (MENA), has started to capture the potential of AI in various ways across the region. Startups such as Virasoft and Sohati occupy the clinical decision support with AI segment, which is the second most funded segment in MENA’s healthtech vertical (Dealroom, 2022). Government initiatives such as the Abu Dhabi Healthcare Information and Cyber Security Standard (ADHICS) started establishing regulations around health data (ADDH, 2019). And collaboration among researchers and innovators is also demonstrated through programs like the AI and Data for Health program at Birzeit University. The program aims to provide knowledge and talent development for stakeholders in the health sector to go on and build responsible AI and data solutions (CCE Birzeit University, 2023).

However, the vast potential of AI in healthtech comes with a unique set of opportunities and challenges. Many of these relate to the sensitive nature of clinical data that AI relies heavily on. This blog post is by no means exhaustive, but it highlights key areas that need to be considered for a safe and responsible health AI ecosystem, including data quality, bias and ethical issues.

 

Data Quality


data-quality.png

Source: Passioned Group 

The accuracy, reliability, and generalizability of AI systems are directly impacted by the quality of the data used to train them. A small negative impact on the accuracy and precision of an AI model can have dire consequences in the clinical context. Interpretability or the understandability of a machine learning (ML) model is also enhanced with high-quality data (Kolyshkina and Simoff, 2021). Data quality in the medical AI domain rely on factors including accuracy, validity, consistency and timeliness among others (Heavy AI, 2022). The reliability of the methods and equipment used to collect clinical data is also crucial to maintain accurate and complete datasets. Some data modalities including x-ray images are often coupled with annotations that can become the core of ML systems and require a distinct level of care and attention (Javaid, 2022). 

Standardization of data quality requirements during data collection, storage and processing is crucial for building multi-source clinical datasets and accelerating medical AI development. Lack of standardization and minimal interoperability between healthcare systems impede cooperation between stakeholders due to the variation and fragmentation in clinical data (Li, Yang and Lin, 2021).

Initiatives such as Dubai’s NABIDH and Saudi Arabia’s Unified Health Record Systems are encouraging examples of efforts in the MENA region to promote data consistency and interoperability (Kharbanda and Linnenbank, 2022). 

 

Data Bias

Biases and errors are significant challenges in clinical AI systems. Bias in source data could disrupt the performance of medical ML models and even render them useless in some cases. Studies have shown the various issues in the transferability of deep learning models as they get released into the wild and tested across different hospitals and communities. Bias can arise due to inadequate representation of certain demographics, biased collection methods and equipment, and inappropriate use of algorithms. These biases can lead to inaccurate diagnoses, incorrect treatment recommendations and other potentially severe consequences. Integrating AI systems into local human, organizational and technical contexts starts at the data collection level (Belenguer, 2022).

Additionally, data augmentation, which is the method of artificially expanding datasets by creating modified copies of the original data has been used to reduce some forms of bias and expand datasets for supervised learning models. Generative AI also offers promising improvements to the transferability of deep learning models and issues caused by data bias. Different modalities of synthetic clinical data can be produced by generative AI to expand the available datasets and modify them for wider representations (Chlap et al., 2021).

 

mri synthetic.png

Source: Han et al.

Several promising examples have demonstrated the power of generative AI. An ML class of frameworks called Generative Adversarial Networks (GANs) are effectively used to produce synthetic clinical data to improve the diagnostic reliability of AI models and diversify the available data. Chest x-rays, mammograms and brain MRIs are being synthetically created with high degrees of accuracy that even expert physicians have a hard time distinguishing them from the real samples (Han et al., 2018).

 

Ethical Issues

The sensitive nature of health records raises important ethical concerns in clinical AI. Ethical use, data handling, privacy and security are challenging areas that are nuanced and hard to navigate. A multidisciplinary approach is needed to define and approach these issues.

One key issue is the protection and privacy of patients’ data. As healthcare data becomes more digitized and AI systems are integrated into healthcare delivery, the risk of data breaches and unauthorized access to sensitive patient information increases. Developing protocols around anonymization, encryption and securitization of data is crucial for the future of medical AI in MENA. Other major factors to be considered include informed consent to use data, safety and transparency, algorithmic fairness, and data privacy (Naik et al., 2022).

Source: Lepri et al.

Another significant ethical issue is related to the previously discussed area of data bias in AI systems. For instance, racial bias has been widely reported in AI systems and algorithms. A study in the United States found that an algorithm designed to allocate patients to healthcare programs was less likely to refer black people than white people to programs they equally qualified for. Electronic health records and medical images vary across different patient phenotypes, so the data used to train AI models have to be inclusive enough to reduce bias in the generated results (Ledford, 2019).

The lines between ethical and technical issues are often blurry in the context of AI systems. Therefore, standardization, regulation and collaboration among stakeholders are elements of paramount importance to grow an effective and responsible medical data and AI ecosystem. Communities that stand to benefit from the promise of clinical AI advancements can only do so with a high-quality, bias-free, inclusive and safe data infrastructure.  

 

 

References

AD Department of Health (2019). Department of Health – launches Abu Dhabi Healthcare Information. [online] www.doh.gov.ae. Available at: https://www.doh.gov.ae/en/news/department-of-health-launches-abu-dhabi-healthcare-information-and-cyber-security.

Belenguer, L. (2022). AI bias: exploring discriminatory algorithmic decision-making models and the application of possible machine-centric solutions adapted from the pharmaceutical industry. AI and Ethics. doi:https://doi.org/10.1007/s43681-022-00138-8.

‌CCE Birzeit University. (2023). Center for Continuing Education | Birzeit University: AI and Data for Health. [online] Available at: https://ccelearning.net/course/index.php?categoryid=140.

Chlap, P., Min, H., Vandenberg, N., Dowling, J., Holloway, L. and Haworth, A. (2021). A review of medical image data augmentation techniques for deep learning applications. Journal of Medical Imaging and Radiation Oncology, 65(5), pp.545–563. doi:https://doi.org/10.1111/1754-9485.13261.

Dealroom. The State of Healthtech in the MENA region (2022). Available at: https://dealroom.co/uploaded/2022/04/Emerge-GHI-Report-April-2022.pdf.

Han, C., Hayashi, H., Rundo, L., Araki, R., Shimoda, W., Muramatsu, S., Furukawa, Y., Mauri, G. and Nakayama, H. (2018). GAN-based synthetic brain MR image generation. 2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018). doi:https://doi.org/10.1109/isbi.2018.8363678.

Heavy AI. (2022). What is Data Quality? Definition and FAQs | HEAVY.AI. [online] Available at: https://www.heavy.ai/technical-glossary/data-quality.

‌Javaid, S (2022). AI Data Collection: Quick Guide, Challenges & Top 4 Methods. [online] Available at: https://research.aimultiple.com/data-collection/.

Kharbanda, V and Linnenbank, P (2022). AI in healthcare: Key lessons for the Middle East. [online] Available at: https://gulfnews.com/uae/health/ai-in-healthcare-key-lessons-for-the-middle-east-1.1664869314913.

‌Kolyshkina, I. and Simoff, S. (2021). Interpretability of Machine Learning Solutions in Public Healthcare: The CRISP-ML Approach. Frontiers in Big Data, 4. doi:https://doi.org/10.3389/fdata.2021.660206.

Ledford, H. (2019). Millions of black people affected by racial bias in health-care algorithms. Nature, [online] 574. doi:https://doi.org/10.1038/d41586-019-03228-6.

Li, R., Yang, Y. and Lin, H. (2021). The critical need to establish standards for data quality in intelligent medicine. Intelligent Medicine, 1(2), pp.49–50. doi:https://doi.org/10.1016/j.imed.2021.04.004.

‌‌Naik, N., Hameed, B.M.Z., Shetty, D.K., Swain, D., Shah, M., Paul, R., Aggarwal, K., Ibrahim, S., Patil, V., Smriti, K., Shetty, S., Rai, B.P., Chlosta, P. and Somani, B.K. (2022). Legal and Ethical Consideration in Artificial Intelligence in Healthcare: Who Takes Responsibility? Frontiers in Surgery, [online] 9, p.862322. doi:https://doi.org/10.3389/fsurg.2022.862322.