AI can improve disease detection but human oversight is needed: experts

1. At a Glance

• AI-driven healthcare uses machine-learning systems for screening, diagnosis and triage. It works best in medical imaging, because digital images give standardised data from which systems can learn patterns [9].
• The main argument: AI can improve disease detection, expand access to healthcare and reduce diagnostic costs. It should be adopted only with clinical validation, data protection and human oversight [9].
• India's governance base has three parts: the ICMR Ethical Guidelines for AI in Biomedical Research and Healthcare (2023), the MeitY AI Governance Guidelines and the DPDP Act, 2023 [1][2].
• For UPSC this links GS-II (health), GS-III (S&T, AI) and GS-IV (ethics of technology: automation bias, accountability, consent).

2. Why in the News

• 13 October 2026 (Tuesday), Chennai: a panel titled 'Artificial Intelligence-Driven Healthcare: Innovation, Ethics, Accessibility and Sustainability' was held as part of The Hindu Sustainability Dialogues 2026, presented by Saveetha Institute of Medical and Technical Sciences (SIMATS) [9].
• Panellists came from Sundaram Medical Foundation (digital health), Deep Tek (research), Skyfire Applied Intelligence (Atom360) and Saveetha Medical College (Transfusion Medicine) [9].
• The panel agreed that AI helps reach larger populations, but adoption must be backed by validation, data protection and human oversight [9].
• Policy context: a PIB document of February 2026, "Transforming Healthcare Delivery Through Artificial Intelligence", sets out the government's AI-in-health deployments and the rules that govern them [1].

3. Background & Evolution

• 2021: WHO issued its first global report, "Ethics and governance of artificial intelligence for health" [8].
• 2023: ICMR released the Ethical Guidelines for Application of AI in Biomedical Research and Healthcare. They cover bias, data handling, autonomy, risk minimisation, professional competence, data sharing and confidentiality, built around ten patient-centric ethical principles [1] (the "ten" count comes from a search summary and was not checked against the ICMR text).
• 2023: the Digital Personal Data Protection (DPDP) Act, 2023 became one of the rules that apply to health-AI deployments [2].
• 18 January 2024: WHO released guidance on Large Multi-Modal Models (LMMs) for health, with 40+ recommendations for governments, technology companies and health-care providers [6].
• 2024–26: NITI Aayog's Frontier Tech Hub documented Indian use cases: AI radiology, AI plus thermal imaging for breast cancer, and AI stethoscopes for rural care [3][4][5].
• February 2026: PIB stated that all AI deployment in health follows the ICMR Ethical Guidelines (2023) and the MeitY AI Governance Guidelines, so that systems are privacy-by-design and exchange data securely [1].

4. Core Static Facts

Item	Fact
Ethics framework for health AI in India	ICMR Ethical Guidelines for AI in Biomedical Research & Healthcare, 2023 [1]
Number of principles	10 patient-centric ethical principles [1] (not checked against the ICMR text)
Cross-sector AI governance	MeitY AI Governance Guidelines [1]
Data-protection statute	DPDP Act, 2023 [2]
Design philosophy	Privacy-by-design and secure data exchange [1]
Example deployment	AI-assisted diabetic retinopathy screening: 38 facilities, 11 states, 14,000+ retinal images, ~7,100 patients [1][2] (exact source unclear, see Section 11)
WHO LMM guidance	Released 18 Jan 2024; 40+ recommendations; five broad application areas, including diagnosis and clinical care [6][7]
WHO base document	Ethics & Governance of AI for Health, 2021 [8]
Strongest AI domain (per the panel)	Medical imaging, because it gives standardised digital data [9]

5. Multi-Dimensional Analysis

Scientific / Technological

• Imaging (radiology, retinal scans, thermal imaging) suits AI because the inputs are standardised [9][3][4].
• New tools push screening outward: AI-powered stethoscopes for rural care [5] and AI plus thermal imaging for early detection of breast cancer [4].
• LMMs bring new risks: inaccurate or biased outputs and training on poor-quality or biased data [6].

Social / Equity

• AI can extend specialist-level screening to under-served and rural populations [9][5].
• WHO warns about how accessible and affordable the best-performing models will be. This could widen the digital health divide [6].
• If datasets are not representative, models may do worse for under-represented groups. ICMR therefore supports trials across diverse populations [4].

Ethical / Governance

• Automation bias: clinicians may defer to machine output too readily. This is the main reason for keeping a human in the loop [6][9].
• Accountability for AI errors, informed consent and confidentiality are core concerns in the ICMR 2023 guidelines [1].
• Clinical validation before deployment and monitoring after deployment are needed for trust [9].

Legal / Constitutional

• The DPDP Act, 2023 governs how personal (health) data is processed [2].
• Health is a State List subject (Entry 6, List II), so national guidelines need State-level uptake for uniform governance.

Economic

• AI can lower diagnostic costs and reduce the workload of scarce specialists [9][3].
• Indian start-ups (e.g. Deep Tek, Atom360) show a domestic health-AI industry emerging [9].

Administrative

• Scaling from pilots (e.g. 38 facilities) to the national level needs interoperable digital infrastructure and trained staff [1].

6. Recent Developments (last 12–18 months)

• Feb 2026: PIB note "Transforming Healthcare Delivery Through AI". It lists deployments and the governing rules: ICMR 2023, MeitY guidelines and privacy-by-design [1].
• 2025–26: a PIB release (PRID 2199422) on steps taken to include AI-based diagnostic tools in healthcare, which cites the DPDP Act among applicable rules [2].
• 25 March 2025: WHO's publications page lists the full LMM guidance report [7].
• 2025–26: NITI Frontier Tech case studies on AI radiology, AI thermal imaging (with ICMR-backed trials) and the AI stethoscope [3][4][5].
• 13 Oct 2026: The Hindu Sustainability Dialogues 2026 panel in Chennai on AI-driven healthcare [9].

7. Prelims Hooks

• ICMR issued its Ethical Guidelines for AI in Biomedical Research and Healthcare in 2023 [1].
• The ICMR 2023 guidelines set out ten patient-centric ethical principles [1] (count not checked against the ICMR text).
• Government health-AI deployments follow the ICMR 2023 guidelines plus the MeitY AI Governance Guidelines [1].
• The guiding design principle for government health-AI is privacy-by-design [1].
• The DPDP Act, 2023 applies to AI-based diagnostic tools that handle personal data [2].
• AI diabetic retinopathy screening: 38 facilities in 11 states, 14,000+ retinal images [1][2].
• WHO released LMM ethics and governance guidance on 18 January 2024 [6].
• The WHO LMM guidance has more than 40 recommendations [6].
• The WHO LMM guidance covers five broad applications of LMMs in health, including diagnosis and clinical care [6].
• WHO's first report on Ethics and Governance of AI for Health came out in 2021 [8].
• Automation bias (over-reliance on machine output) is a risk named in WHO guidance [6].
• ICMR is backing trials of AI plus thermal imaging for early detection of breast cancer [4].
• The Hindu Sustainability Dialogues 2026 was presented by SIMATS (Saveetha Institute of Medical and Technical Sciences) [9].

8. Why a Good Pilot Does Not Prove a Safe National Rollout

• The biggest Indian example is still small
• The AI diabetic retinopathy screening (an eye disease caused by diabetes) covered about 7,100 patients in 38 facilities [1][2].

• India has hundreds of millions of people across very different States. A tool that worked for 7,100 people has not yet been tested on most of them.

• An AI model learns only the people it was trained on

• WHO says it "can be difficult for AI models to accurately represent the diversity of populations" [10].
• A model trained mostly on city patients, or on one machine brand, can miss disease in a tribal district or on a cheaper camera. Its accuracy drops quietly, and nobody sees a warning sign.

• This is why ICMR is backing trials of AI thermal imaging across diverse populations, not just one site [4].

• External validation is the missing step between pilot and scale

• External validation means testing the tool on new patients from places it has never seen before.
• WHO lists external validation and a clear "intended use" (exactly what job the tool is approved to do) among its six areas for regulating health AI [10].
• A tool built to say "refer this patient or not" should not slowly turn into a tool that gives the final diagnosis.

9. Guidelines Advise, but Only Rules Can Be Enforced

• The ICMR and MeitY documents are guidelines, not law
• PIB says government health-AI follows the ICMR 2023 guidelines and the MeitY guidelines [1].
• A guideline tells you what good practice is. It carries no licence, no penalty and no recall power.

• The binding layer is different: the Medical Devices Rules, 2017, enforced by CDSCO (Central Drugs Standard Control Organisation, India's drug and device regulator), plus the DPDP Act [11][2].

• The device rules were written for fixed products

• The 2017 Rules follow the GHTF framework (Global Harmonisation Task Force, an older international model for device rules) [12].
• For AI-enabled devices, the applicant must submit full technical documentation under these Rules [11].

• But some AI keeps learning after approval. WHO names this "continuous learning" as a risk regulators must manage [10]. If the model changes after the licence, the approved product and the product in use are no longer the same.

• Licensing is split between the Centre and the States

• Lower-risk devices (Class A and B) get manufacturing licences from State authorities. Higher-risk ones (Class C and D) go to the central authority [12].
• So how strictly a health-AI tool is checked can depend on its risk class and on which body licenses it.

10. The Strongest Case Against Insisting on a Doctor Every Time

• The objection: in many villages there is no doctor to oversee
• AI's main promise is reaching places with no specialist, like rural clinics using AI stethoscopes [5].
• If every AI result needs a specialist to sign it, the tool becomes useless where it is needed most.

• For that patient, the real choice is AI screening or no screening at all. It is not AI or a doctor.

• What is right about this objection

• A rule that always requires a specialist can protect patients on paper while denying them care in practice.

• Human oversight can also become a formality. Because of automation bias (trusting the machine too easily), a tired doctor may just approve whatever the machine says [6]. Then the doctor gets blamed, but nobody really checked.

• The answer: match the level of oversight to the risk of the job

• WHO treats "human intervention" as a risk-management choice that depends on intended use. It is not one fixed rule for every tool [10].
• A screening tool that only says "send this patient to the district hospital" needs lighter oversight. A tool that decides treatment needs a qualified clinician.
• This keeps the access benefit while making clear who answers for an error.

11. Who Should Do What Next

• CDSCO should ask for proof that a tool works on Indian patients before licensing it
• It can follow WHO's advice: make datasets "intentionally made representative" and report patient features such as gender, race and ethnicity [10].

• A developer would then have to show accuracy for women, older patients and different regions separately, not just one average number.

• ICMR should make multi-site trials the normal route to scale

• The breast-cancer thermal imaging trials across diverse populations [4] are a model. The retinopathy tool should go through the same kind of trial before it grows beyond its 38 facilities [1].

• Regulators should watch tools after approval, not only before

• WHO asks for documentation across the full product lifecycle (from building the tool to its retirement) [10].

• In practice, hospitals using AI should report missed cases back, so that a model whose accuracy is falling gets caught.

• State health departments must adopt the rules themselves

• Health is a State List subject, and State authorities license lower-risk devices [12].

• Without State uptake, the national guidelines stop at the State border.

• Consent must work for a patient who cannot read the form

• WHO calls consent requirements under data laws such as GDPR (Europe's data-protection law) complex, even in rich countries [10].
• Under the DPDP Act [2], consent in a rural clinic must be explained in the patient's own language, or it is consent only on paper.

12. Anchors for Answers

• Data: AI diabetic retinopathy screening covered ~7,100 patients and 14,000+ retinal images across 38 facilities in 11 States: a pilot, not national proof [1][2]
• Report/Committee: WHO, Regulatory considerations on AI for health (October 2023): six areas, including external validation, data quality and lifecycle documentation [10]; ICMR Ethical Guidelines for AI in Healthcare (2023) [1]
• Law/Case: Medical Devices Rules, 2017 (CDSCO; Class A/B licensed by States, C/D by the Centre) [12][11]; DPDP Act, 2023 [2]; Entry 6, List II (health as a State subject)
• Comparison: WHO cites Europe's GDPR and the US's HIPAA (its health-privacy law) to show how hard consent and jurisdiction become for health AI [10]
• Scheme: ICMR-backed multi-site trials of AI thermal imaging for breast cancer: a model for validating tools before scale-up [4]

13. Mains Relevance

• GS-II: Issues relating to development and management of the Social Sector/Services relating to Health; Government policies and interventions.
• GS-III: Science & Technology: developments and their applications; awareness in IT, computers, robotics; indigenisation of technology.
• GS-IV: Ethics in public administration; ethical issues in technology (accountability, consent, bias).
• Possible question stems: 1. "AI can democratise diagnostics in India but cannot replace the clinician." Critically examine, with reference to the regulatory framework for AI in healthcare. 2. Discuss the ethical challenges of large multi-modal AI models in healthcare. How far do the ICMR (2023) and WHO (2024) frameworks address them? 3. Data protection and clinical validation are prerequisites for scaling AI in public health. Analyse in the light of the DPDP Act, 2023.

14. Related Topics to Study Next

• Ayushman Bharat Digital Mission (ABDM): the digital health infrastructure that AI tools plug into.
• DPDP Act, 2023: how health data is protected and when consent is needed.
• IndiaAI Mission: national compute and datasets for AI, including health use cases.
• ICMR Ethical Guidelines (2017) for biomedical research: the parent ethics framework.
• Medical Devices Rules, 2017 / CDSCO: how software counts as a medical device.
• WHO Global Strategy on Digital Health 2020–25: the international policy context.
• Telemedicine (eSanjeevani): a parallel tool for extending access to care.
• Algorithmic bias and explainable AI (XAI): the technical side of GS-IV ethics questions.

15. NCERT Concepts to Revise

• Digital Personal Data Protection Act 2023 (Polity): The law governing how patients' health data is processed; central to data protection in health AI.
• Accountability (Polity): Answerability for AI errors is a core ethical concern, needing clear responsibility and human oversight.
• Automation (Economy): Helps explain automation bias, where clinicians over-trust machines, and why AI must support rather than replace doctors.
• State List (Polity): Public health is a State subject, so national AI guidelines need State-level uptake.
• NITI Aayog (Polity): Its Frontier Tech Hub documented Indian AI health cases; it is a knowledge platform, not a regulator.
• Information and communication technology (Economy): Gives the digital technology background for AI tools, digital health infrastructure and extending care to remote areas.

16. Common Errors / Trap Areas

• ICMR vs MeitY: health-specific AI ethics guidelines come from ICMR (under the Department of Health Research). MeitY issues cross-sector AI governance guidelines [1].
• Year confusion: the WHO general AI-health guidance is from 2021; the LMM-specific guidance is from January 2024 [8][6]. ICMR's AI guidelines are from 2023, not 2017 (2017 is the general biomedical research ethics guidelines).
• "40+ recommendations" belongs to the WHO LMM guidance, not the ICMR guidelines [6].
• Human oversight means AI supports clinical decisions; it does not replace them. Statements implying AI can diagnose on its own are usually wrong [9][6].
• NITI Frontier Tech Hub case studies are documentation and knowledge platforms, not regulatory instruments [3].