Artificial Intelligence in Clinical Diagnosis: Promise and Privacy

AI in medicine is not magic. It is a model trained on many labelled examples.

1. What is AI, and who named it?

Artificial intelligence, or AI, means machines that do tasks which need human intelligence. John McCarthy wrote the 1955 proposal for a summer research conference on “artificial intelligence”. That was the first use of the phrase in print. He then organised the 1956 Dartmouth summer project.

Machine learning is a part of AI. A machine learning model learns patterns from data. Deep learning is a part of machine learning. It uses layered neural networks.

2. How does a deep learning model learn to spot disease?

Developers feed the model many scans. Each scan has a label, such as “cancer” or “no cancer”. The model finds patterns that match the labels. No doctor writes the rules inside it.

A lung study shows the idea. Researchers trained a model on low-dose chest computed tomography (CT) scans, which are detailed X-ray images.

It reached 94.4% area under the curve on 6,716 cases from the National Lung Cancer Screening Trial. It did about as well on an independent set of 1,139 cases.

3. Can AI match a radiologist on a scan?

On narrow tasks it can. In 2019 the same lung model was tested against six radiologists. When no earlier scan was available, the model did better than all six. It cut false positives by 11% and false negatives by 5%.

When an earlier scan was available, the model did about as well as the radiologists. So AI works best as a helper, not a replacement.

4. How does AI help with slides and eye photos?

A pathology study used 44,732 slides from 15,187 patients. The system let pathologists skip 65% to 75% of slides while keeping 100% sensitivity. That saves expert time for the hard cases.

In 2018 Google described a tool called LYNA, short for Lymph Node Assistant. It checks lymph node slides for breast cancer that has spread. It told slides with cancer from slides without cancer 99% of the time.

Tools must prove themselves before doctors may use them. The first approval for an autonomous AI screen came in 2018.

5. What was IDx-DR, and why did it matter?

On 11 April 2018 the US Food and Drug Administration (FDA) permitted the marketing of IDx-DR (DR stands for diabetic retinopathy).

It was the first medical device to use AI to detect more than mild diabetic retinopathy in adults with diabetes. Diabetic retinopathy is damage to blood vessels in the retina.

The software reads eye images taken with a Topcon camera. It gives a result without a specialist reading the image. That lets a primary care office run the screening.

6. How accurate was IDx-DR in its trial?

The FDA studied images from 900 patients at 10 primary care sites. IDx-DR found more than mild disease 87.4% of the time. It found no such disease in 89.5% of patients who did not have it.

The company’s own trial paper gives close numbers. Sensitivity was 87.2% and specificity was 90.7%. Sensitivity means catching the sick. Specificity means clearing the healthy.

India has set a national plan, funded a national mission, and issued health-specific ethics rules.

7. What did NITI Aayog say in its AI strategy?

NITI Aayog (National Institution for Transforming India), the government’s policy think tank, published the National Strategy for Artificial Intelligence in 2018. It uses the tag #AIforAll. The strategy names healthcare as its first focus area, ahead of agriculture and education.

It also has a chapter on ethics, privacy and security. The Union Budget speech for 2018-19 had asked NITI Aayog to set up a national programme on AI.

8. What is the IndiaAI Mission?

The Union Cabinet approved the IndiaAI Mission in March 2024. The budget outlay is Rs 10,371.92 crore. It plans public AI compute with 10,000 or more GPUs.

The mission aims to make computing easier to access and to improve data quality. It also backs home-grown AI tools and AI skills.

9. What do the ICMR guidelines add?

The Indian Council of Medical Research (ICMR) issued ethical guidelines for AI in biomedical research and healthcare in 2023. They aim to guide ethical decisions while AI tools are built, deployed and adopted.

They are written for doctors, researchers and developers. They sit beside the Digital Personal Data Protection (DPDP) Act, which covers all personal data.

Health records show illness, habits and family history. Misuse can cost a person a job or insurance.

10. What does India's DPDP Act, 2023 say?

The Digital Personal Data Protection Act, 2023 covers digital personal data. Consent must be free, specific, informed, unconditional and unambiguous. It needs a clear affirmative action.

The Act has no separate “sensitive data” class. Health data gets the same rules as other personal data. The Act does allow processing without consent in a medical emergency. It also allows it for treatment during an epidemic or other threat to public health.

A failure to keep reasonable security safeguards can bring a penalty of up to Rs 250 crore.

11. How do the EU and WHO treat health data?

The EU’s General Data Protection Regulation (GDPR) puts data about health in a special category. Article 9 bans processing it, unless an exception applies. One exception is the person’s explicit consent.

The World Health Organization (WHO) issued guidance on the ethics of AI for health on 28 June 2021. It sets out six consensus principles so that AI serves the public good.

12. Is removing the name enough to protect a patient?

No. Latanya Sweeney showed this in 2000. She found that 87% of the US population could be picked out by only three facts. These were a 5-digit ZIP (postal) code, gender and date of birth.

So data without names can still point to one person. Joining two datasets can rebuild an identity.

13. How can hospitals protect patients and still use AI?

Federated learning is one answer. The model trains inside each hospital. Only model details, such as parameters, are shared. Patient data stays behind the hospital’s firewall.

Another problem is the black box. Many models cannot explain their results. A Nature Machine Intelligence paper argues that for high-stakes decisions we should use models that are interpretable.

The 2023 Mains asked for both sides. It wanted the benefits and the privacy threat.

14. What did UPSC Mains 2023 ask, and how do you answer it?

The question came in UPSC Civil Services Mains 2023, General Studies Paper III (the third General Studies paper), as Q5. It carried 10 marks and a 150-word limit.

Answer in three blocks. First define AI. Then give the diagnosis gains, such as scans, slides and eye screening. Last, weigh the privacy threat with consent, the DPDP Act and ethics guidance.