News of the possibility of “a free AI-powered primary care doctor for every Indian, available 24/7” within the next five years is ambitious. This raises critical questions about India’s feasibility, sustainability, and preparedness to tackle such a huge undertaking.
Primary health care (PHC) guarantees the right to the highest attainable level of health by bringing integrated services closer to the community. It addresses health needs, addresses the wider determinants of health through multisectoral action, and empowers individuals to manage their health. We risk undermining this fundamental aspect of PHC by relying on Artificial Intelligence (AI) because it is impersonal, making people passive recipients of care rather than active participants.
AI excels at processing and automating repetitive tasks but lacks characteristics of human intelligence such as understanding the physical world, retrieving complex information, maintaining continuous memory, and engaging in reasoning and planning. This is all fundamental to medicine, where understanding the patient’s condition goes beyond pattern recognition.
Delivering health care requires a human-centric approach of empathy and cultural understanding. Consciousness – awareness and understanding of the real-world environment – supports human decision-making, distinguishing human intelligence from AI. AI cannot replicate the moral and ethical judgments that come from conscious experience. Unlike other domains, healthcare data is scattered, incomplete and often inaccessible for AI training, making it difficult to train models.
Data, models and problems
Naegele’s rules of obstetrics, which have been used for more than 200 years, can be used to highlight challenges in health care. This is based on the 18th century reproductive habits of European women, which may not apply today. This method is used to predict the date of birth of a child during pregnancy. It depends only on the length of the last menstrual cycle and has an accuracy of 4%. Failure to account for critical factors such as maternal age, parity, nutrition, height, race, and uterine type, which are essential for accurate prediction. Developing a better prediction model than Naegele’s rule requires a lot of personal data, which is the right of the patient. This illustrates the paradox inherent in the development of AI in healthcare – the need for extensive data collection to improve accuracy is incompatible with privacy and ethical concerns.
The costs involved in building the infrastructure to capture, collect, and train this data are enormous. As levels of reproductive health and fertility change, continuous AI model changes are necessary, resulting in ongoing costs. Health care data is complex and personal, making it difficult to standardize across populations.
India’s diversity complicates the problem further. This diversity means that data for AI models must be extensive and highly contextual, but generating that data requires access to personal and behavioral information.
Utility of AI in healthcare
AI can play an important role in certain defined tasks in health care, especially through narrow intelligence, diffusion models and transformers. Narrow intelligence focuses on specific tasks such as predicting hospital kitchen supply needs, managing biomedical waste, or optimizing drug procurement. Diffusion models, which are good at predicting patterns from complex data sets, can help filter histopathological slides or display only a subset of the population using medical images.
Large Language Models (LLMs) and Large Multimodal Models (LMMs) are emerging as powerful tools in medical education and research writing. It can provide rapid access to medical knowledge, simulate patient interactions, and support the training of health care professionals. By offering personalized learning experiences and simulations of complex clinical scenarios, LLMs and LMMs complement traditional medical education.
An important problem with AI in healthcare is the “black box” problem, where the AI algorithm’s decision-making process is neither transparent nor easy to understand. This poses a risk in healthcare, where understanding the rationale for the diagnosis or treatment plan is critical. Healthcare providers are left in the dark about certain conclusions, leading to a lack of trust and potential harm if AI makes incorrect or inappropriate recommendations.
Google DeepMind’s mysterious AI algorithm that beat world-class players in the GO game (board game) can be celebrated. While these achievements are acceptable for gaming, they raise concerns about actual health care decisions. The stakes are serious in human health, where the consequences of mistakes can be life-threatening.
India and AI governance issues
A recent petition in the Kenyan Parliament by content moderators against OpenAI’s ChatGPT has highlighted the ethical complexities in AI development, pointing to the exploitation of unpaid workers in the training of AI models. This raises concerns about the exploitation of vulnerable populations in AI training. This underscores the importance of safeguarding the interests of Indian patients as the data required to train the model legitimately becomes a patient.
While population-level data generated by health systems can be useful, they are prone to ecological errors. India does not have comprehensive regulations or legislation on AI such as the European Union’s Artificial Intelligence Act, making it even more critical. AI tools in healthcare must be developed and deployed with the core medical ethics of “Do No Harm”.
AI-powered healthcare in India promises efficiency and reduced error rates. Advanced AI technologies require significant investment in research, data infrastructure, and continuous updates – costs that people must bear. India cannot leapfrog into AI-driven healthcare without addressing fundamental issues in its healthcare system. The complexity of patient care, the need for high-quality data, and the ethical implications of AI require a more measured approach.
Dr. C. Aravinda is an academic and public health physician. The opinions expressed are personal
Published – 13 September 2024 12:08 IST
