REGULATION OF DECENTRALIZED AI TRAINING ACROSS HOSPITALS

1. INTRODUCTION

Decentralized artificial intelligence (AI) training across hospitals involves developing machine-learning models using medical data held by multiple healthcare institutions without necessarily transferring their raw patient records to a central repository. A common approach is federated learning, under which participating hospitals train a shared model locally and transmit model updates to a coordinating server or network.

This technology can improve medical-image analysis, disease prediction, diagnostic accuracy and clinical decision support while reducing the need to pool sensitive health information. However, decentralized training does not automatically guarantee privacy or legal compliance. Model updates may reveal information about individual patients, and trained models may reproduce biases, memorise sensitive data or perform poorly on underrepresented populations.

In India, the principal legal framework includes the Digital Personal Data Protection Act, 2023 (DPDP Act), applicable provisions of the Information Technology Act, 2000, the Information Technology (Reasonable Security Practices and Procedures and Sensitive Personal Data or Information) Rules, 2011, where applicable, and healthcare confidentiality obligations. The commencement and transitional applicability of particular provisions must be checked against the relevant notifications. Additional requirements may apply to clinical investigations, medical devices and institutional research.

2. LEGAL AND REGULATORY FRAMEWORK

A. DIGITAL PERSONAL DATA PROTECTION ACT, 2023

The DPDP Act establishes obligations concerning the processing of digital personal data, including lawful processing, notice and consent where required, security safeguards, breach-related duties and accountability of data fiduciaries.

Hospitals participating in decentralised AI projects must determine their respective roles. A hospital may act as a Data Fiduciary when determining the purpose and means of processing, while a coordinating organisation may act as a Data Processor or may itself be a Data Fiduciary depending on its actual decision-making role.

Keeping raw records inside a hospital does not necessarily take the project outside data-protection law. Gradients, parameter updates, identifiers and model outputs may still constitute personal data if they relate to identifiable individuals.

B. INFORMATION TECHNOLOGY ACT, 2000

Section 43 addresses specified unauthorised access, copying and other computer-related acts, while Section 66 criminalises conduct under Section 43 when committed dishonestly or fraudulently. Section 72 addresses specified breaches of confidentiality and privacy, and Section 72A addresses disclosure of information in breach of a lawful contract under its prescribed conditions.

These provisions may become relevant where a hospital's AI infrastructure is accessed without authority, patient information is improperly extracted or confidential information is disclosed unlawfully.

C. HEALTHCARE CONFIDENTIALITY AND PROFESSIONAL DUTIES

Hospitals and medical practitioners must comply with applicable confidentiality, professional-conduct and record-management obligations. Research involving patients' medical information should undergo appropriate institutional ethical review where required, and patients should receive clear information about the use of their data when the applicable legal or ethical framework requires it.

D. MEDICAL DEVICE REGULATION

Where a decentralised AI model is intended for diagnosis, treatment recommendations or other regulated clinical functions, its classification and intended use must be assessed under the Medical Devices Rules, 2017, and applicable Central Drugs Standard Control Organisation requirements. A model's development through federated learning does not exempt the resulting clinical product from applicable regulatory controls.

3. CORE LEGAL REQUIREMENTS

A compliant hospital network should implement the following safeguards:

LAWFUL DATA PROCESSING: Establish the applicable legal basis, permitted purpose and responsibilities of each participating institution.

DATA MINIMISATION: Use only the records and attributes reasonably necessary for the defined research or clinical objective.

PRIVACY-PRESERVING TRAINING: Evaluate secure aggregation, differential privacy, encryption and access controls according to the project's risks.

ETHICAL OVERSIGHT: Obtain institutional ethics committee review where required and document any applicable consent or waiver.

CYBERSECURITY: Protect local datasets, training servers, model updates, credentials and communication channels against unauthorised access.

MODEL VALIDATION: Evaluate performance across hospitals, demographic groups, clinical settings and relevant patient subpopulations.

ACCOUNTABILITY: Maintain training logs, dataset provenance, model versions, incident records and documented procedures for responding to patient complaints.

THIRD-PARTY GOVERNANCE: Define contractual responsibilities for software vendors, cloud providers, research partners and model coordinators.

These safeguards should be established before training begins and maintained throughout deployment, updates and retirement of the model.4. IMPORTANT CASE LAWS

There is no established reported Indian judgment directly determining the legality of federated AI training across hospitals. The following decisions are therefore identified as constitutional, healthcare-confidentiality or analogous digital-evidence authorities.

1. JUSTICE K.S. PUTTASWAMY (RETD.) v. UNION OF INDIA, (2017) 10 SCC 1

Authority: Direct constitutional authority on privacy; analogous application to AI training.

A nine-judge Constitution Bench of the Supreme Court recognised privacy as a fundamental right protected under Article 21 and other provisions of Part III of the Constitution. The judgment recognised informational privacy and the need to balance individual rights with legitimate public interests.

Legal principle: Hospitals cannot disregard patients' privacy merely because medical data may improve diagnostic models. Data processing must satisfy applicable legal requirements, and any constitutionally impermissible interference must be assessed against legality, legitimate purpose and proportionality. The Court did not specifically decide federated learning. <Cite refs={["turn233885search0","turn233885search2"]}/>

2. JUSTICE K.S. PUTTASWAMY (RETD.) v. UNION OF INDIA, (2019) 1 SCC 1

Authority: Analogous constitutional authority on proportionality.

In the Aadhaar judgment, the Supreme Court examined privacy, informational autonomy and the proportionality of measures involving personal information.

Legal principle: A hospital network should be able to justify why particular patient data are required, why the chosen processing method is suitable and whether less intrusive alternatives are available. Federated learning may reduce centralised exposure, but its privacy benefits must be evaluated against the actual design and risks of the system.

3. MR. X v. HOSPITAL Z, (1998) 8 SCC 296

Authority: Direct healthcare-confidentiality authority; analogous to data-sharing disputes.

The Supreme Court considered disclosure of a patient's HIV status and recognised the importance of medical confidentiality while also examining circumstances in which competing rights and public-health interests may justify disclosure.

Legal principle: Medical information is subject to confidentiality obligations, and disclosure requires a legally defensible justification. Hospitals collaborating on AI research should not assume that sharing or processing patient information is permissible merely because the project has potential public-health benefits. <Cite refs={["turn233885search5","turn233885search27"]}/>

4. DISTRICT REGISTRAR AND COLLECTOR, HYDERABAD v. CANARA BANK, (2005) 1 SCC 496

Authority: Analogous constitutional authority.

The Supreme Court considered privacy interests in relation to access to private records and emphasised the importance of protecting personal information against unjustified intrusion.

Legal principle: Access to sensitive records should be limited by lawful authority and appropriate safeguards. For hospital AI networks, this supports role-based access controls, restricted research permissions and documented authorisation for accessing clinical records.

5. SHREYA SINGHAL v. UNION OF INDIA, (2015) 5 SCC 1

Authority: Analogous digital-law authority.

The Supreme Court examined provisions of the Information Technology Act, including Section 79 concerning intermediary liability, and invalidated Section 66A. The judgment clarified important aspects of online expression and intermediary regulation.

Legal principle: Digital technology must be governed by the actual statutory provisions applicable to the activity. In decentralised healthcare AI, the roles of hospitals, platform operators and technology vendors should be distinguished rather than assuming that every participant bears identical legal responsibility. This case did not determine liability for federated learning or patient-data breaches.

6. ARJUN PANDITRAO KHOTKAR v. KAILASH KUSHANRAO GORANTYAL, (2020) 7 SCC 1

Authority: Direct electronic-evidence authority; analogous application to AI audit records.

The Supreme Court examined the admissibility requirements for electronic records under Section 65B of the Indian Evidence Act, 1872, as it then applied. It clarified the certificate requirements for computer-generated evidence in the circumstances covered by that provision.

Legal principle: Where a dispute involving AI training reaches litigation, electronic records such as server logs, model-version histories, access records and audit trails must be preserved and produced in accordance with the applicable evidentiary law. The judgment concerns electronic evidence, not the substantive legality of AI training.

5. LIABILITY AND RISK MANAGEMENT

Liability may arise when a participating institution processes patient information without appropriate authority, fails to protect its systems, discloses confidential records unlawfully or deploys a clinically unsafe model in breach of applicable obligations.

Potential consequences include regulatory proceedings, compensation claims where legally available, contractual liability, professional disciplinary action and, where the elements of an offence are established, criminal prosecution. Responsibility should be allocated according to each institution's actual conduct, legal role and applicable statutory duties.

Particular attention is required for model inversion, membership inference and data leakage. These attacks can potentially reveal whether a patient's data were used in training or reconstruct aspects of sensitive records from model outputs or updates. Secure aggregation, differential privacy, encryption and access controls may reduce these risks, but none should be treated as an unconditional guarantee of anonymity.

6. PRACTICAL COMPLIANCE FRAMEWORK

Hospitals should adopt a written collaboration agreement specifying the research purpose, permitted data uses, security responsibilities, model ownership, retention periods, incident reporting and procedures for withdrawal or deletion where applicable.

Before training, institutions should conduct a privacy and security assessment, establish appropriate ethics review, document the legal basis for processing, validate the model across participating sites and assess the possibility of re-identification. They should also maintain procedures for responding to data breaches, correcting model defects and suspending unsafe deployments.

Important current-law qualification: India's DPDP Act, 2023 and DPDP Rules, 2025 have a phased commencement schedule. The principal processing obligations under Sections 3–17 of the Act are scheduled to commence 18 months after the 13 November 2025 notification, on 13 May 2027. Accordingly, as of 27 October 2026, hospitals must distinguish provisions already in force from those not yet commenced and comply with the applicable existing legal framework during the transition. <Cite refs={["turn738774search24","turn738774search0","turn738774search3"]}/>

7. CONCLUSION

Decentralised AI training across hospitals can advance medical research while reducing the need to centralise sensitive clinical records. Nevertheless, decentralisation alone does not eliminate privacy, cybersecurity, discrimination or clinical-safety risks.

Indian constitutional jurisprudence, particularly K.S. Puttaswamy, establishes the importance of privacy and proportionality, while Mr. X v. Hospital Z reinforces the significance of medical confidentiality. Electronic-evidence and digital-law decisions provide additional guidance for governance and accountability. A sound regulatory framework must combine lawful data processing, effective security, ethical oversight, transparent institutional responsibilities and rigorous clinical validation to ensure that technological progress remains consistent with patients' rights.