New Delhi:  In May 2024, Ankit Gupta, a Delhi resident woke up from his deep slumber, hearing a strong beeping sound; but he couldn't find the source of the sound. Worried about his ears ringing, he entered some of his symptoms into ChatGPT, an Artificial Intelligence (AI)-powered chatbot that scrapes the internet for information and then organizes it based on the questions asked.

"Hearing a sudden, loud beep-like sound”, “ringing noises in the left ear” and “difficulty in hearing distant voices," were some of the symptoms he recalls having entered into the AI powered chatbot.

“These symptoms could indicate tinnitus, which is characterized by hearing ringing, buzzing, or other sounds not caused by external sources,” ChatGPT replied.

Two months later, in July 2024, Gupta was officially diagnosed with Tinnitus by an ENT specialist at AIIMS, New Delhi. 

Can AI be used for self-diagnosis?

There has been a growing interest in developing AI-enabled chatbot-based symptom checker (CSC) apps in the healthcare market. The internet is flooded with AI-powered apps and platforms which are being used to seek preliminary health assessments. 

Now 44, Gupta had started facing difficulty in hearing voices on the phone in his mid 20’s. He initially ignored it but later, though inconsistent, the symptoms grew. He gradually came to suspect it was Tinnitus, but still needed a diagnosis to confirm what he suspected. 

“While it wasn't a diagnosis, I was amazed at the way ChatGPT jumped to the right conclusion,” Gupta said. 

However, experts warn that such AI based apps need to be used with caution. 

“The effectiveness of AI models depends heavily on the quality of data. If the data isn’t accurate, standardized, or clean, the models won’t perform well," said Dr. Sushil Meher, Chief Information Officer (CIO) of the All India Institute of Medical Sciences (AIIMS), New Delhi.

But now, with better data and communication technology, AI’s role in healthcare is expanding, he added.

Healthcare professionals believe that though AI can be an assisting technology, it cannot replace the nuanced judgment and experience of a healthcare professional, especially when it comes to complex, multi-faceted health conditions.

“You can’t just conclude something based on an AI reading,” warned Dr. Meher, who has worked in Information Technology (IT) within the healthcare sector for the last 32 years. 

He added that “healthcare is a different kind of domain... If a model has a 98% accuracy rate—are you ready to fully depend on that model? No, because there’s still a gap.” 

A study titled “Self-Diagnosis through AI-enabled Chatbot-based Symptom Checkers: User Experiences and Design Considerations” investigated the effectiveness of consumer-facing symptom checker (CSC) apps and examined users’ perceptions of these tools. It found that the CSC apps and healthcare chatbots must integrate diverse patient history, offer flexible symptom input, improve response speed, support diverse health conditions, and provide follow-up treatment guidance.

Studies have also examined the risks and challenges associated with self-diagnosis of mental health conditions using AI-driven tools. 

“The biggest problem is the lack of diversity in data that is being fed to AI,” warned Dr. Meher. “AI is being trained with datasets that ignore important factors like gender, demography, geography, etc.”

He explained that AI should be used only in specific situations where you need assistance but still require human intervention. If the doctor is unavailable, AI can be helpful, but it’s never a replacement for professional judgment. 

Training AI for healthcare

AI-based chatbots are built on large language models (LLMs), which are AI neural networks that utilize deep learning to analyze vast amounts of data and perform various natural language processing tasks. 

These models are capable of understanding, summarizing, predicting, and generating content in a way that is user-friendly. However, for AI to perform these tasks effectively, it must undergo a process of training.

AI training involves teaching an AI system to analyze, understand, and learn from data so it can later make informed decisions based on the information it receives. 

This process requires three key elements: a well-structured AI model, vast amounts of high-quality and accurately labeled data, and a robust computing platform. When trained effectively, the potential of AI becomes virtually boundless.

Dr. Bakul Gohel, Assistant Professor at Dhirubhai Ambani Institute of Information and Communication Technology (DA-IICT), Gandhinagar explained that training AI for healthcare involves significant challenges.

"The issue is not the technology itself but the quality and availability of the data. If you have sufficient and well-labeled data, training a model is straightforward,” he said.

However, he raised concerns about the data limitations and biases in AI models. 

The biases occur because there is lack of appropriate labeling in data. For instance, when working with medical images, such as identifying tumors in X-rays or mammograms, there is a need to precisely label the tumor’s location, size, and characteristics. Without accurate labeling, the data cannot be used effectively for training AI models, experts say.

Dr. Gohel pointed out that in the Western countries, private companies are already addressing the issue of labeling, but in India, there is still a lack of organized efforts to label data effectively. “Government should issue some guidelines on maintaining a repository of labeled data,” he said. 

A study titled “Re-Thinking Data Strategy and Integration for Artificial Intelligence: Concepts, Opportunities, and Challenges” comprehensively reviewed and critically examined the challenges of using data for AI, including data quality, data volume, privacy and security, bias and fairness, interpretability and explainability, ethical concerns, and technical expertise and skills.

“Overcoming AI data challenges requires improving data quality, security, fairness, and interpretability. It stresses the importance of interdisciplinary collaboration, education, and ethical AI development to drive future advancements in the field,” the study noted.

Experts believe that integrating AI tools with professional medical judgment is important. “People who are training the AI with datasets also lack medical expertise, that’s a unique challenge as far as AI-training in the healthcare sector is concerned,” said Dr. Gohel, who specializes in brain-computer interaction, cognitive computing and data analysis. 

Future of AI in Indian Healthcare

AI's integration into healthcare systems could revolutionize diagnostic accuracy and treatment protocols, particularly in remote and underserved regions, where access to healthcare professionals is limited. However, experts dismiss the possibility of AI replacing healthcare providers. 

"AI can act as a tool to support doctors, improving decision-making and providing data-driven insights, which can help before actual diagnosis,” said Dr. Meher. 

However he raised a critical concern about the future of AI diagnostics, questioning who approves the AI models and ensures their accuracy before they are used in healthcare. 

“Unlike approved health messages on health websites or Food and Drug Administration (FDA) -certified products, there is no regulatory authority currently overseeing AI models,” Dr. Meher said. 

"When somebody launches an AI model, they claim it themselves," he explained, pointing out that the developers' data and methods remain unclear. He warned that while people are increasingly relying on AI technology, it is essential to be cautious, especially in healthcare, where the stakes are high.

Also read: Explainer: Can artificial intelligence replace doctors? - First Check