development and validation of a risk based post market regulatory assesment framework for ai-enabled medical devices in india- a comparsion assessment of global regulatory approaches

Global regulatory approaches to AI-enabled medical devices: side-by-side comparison

The table below compares the four jurisdictions and the international consensus layer covered by the retrieved evidence. Cells marked 'not reported' reflect genuine silences in the reviewed literature rather than omissions in those systems.

Comparison table

Why AI-enabled devices challenge conventional post-market surveillance

The clearest point of agreement across the corpus is that AI-specific failure modes are systematically invisible to conventional, event-based adverse-event reporting. Analysis of 943 medical device reports linked to 823 FDA-cleared AI/ML devices (2010–2023) found that concept drift, covariate shift and algorithmic instability — the failure modes most characteristic of AI — cannot be captured under malfunction/injury/death categories, and the authors proposed quarterly manufacturer 'deployment conditions' reports and nutrition-style labelling as remedies (Direct; PMID: 40450160). A retrospective simulation study on 130,486 retinal images showed why such monitoring matters: sensitivity for detecting referable diabetic retinopathy ranged from 0.50 to 0.79 across ethnic subgroups, and classifier-based shift-detection tests achieved perfect detection of image-quality and co-morbidity subgroup shifts with about 1,000 deployment samples, whereas shifts in patient sex required more than 30,000 samples — meaning some shifts are practically detectable and others are not (Direct; PMID: 38724581). Cross-border deployment evidence reinforces this: AI models trained on UK National Health Service data achieved AUROCs of only 0.515–0.577 when applied unchanged in two Vietnamese hospitals, improving to roughly 0.650–0.707 with transfer learning, with locally trained models performing second-best (Direct; PMID: 39333515).

What the existing reporting systems do capture is dominated by non-algorithmic issues. Of 266 safety events involving ML-enabled devices reported to FDA, 82% were driven by problems with data input to the device, and while 93% of events involved device problems, the 7% that were use problems were four times more likely to cause harm (relative risk 4.2; 95% CI 2.5–7) (Direct; PMID: 37071804). Similarly, of 429 AI-device-related FDA reports reviewed, only 25.2% were judged potentially AI/ML-related and 34.5% contained too little information to tell (Direct; PMID: 39627534). An empirical portfolio analysis of 1,016 FDA authorisations (736 unique devices) shows what is being surveilled: 84.4% of devices use images as their core AI input, and none involve large language models — the surveillance challenge today is overwhelmingly diagnostic imaging software (Direct; PMID: 40596700).

Premarket evidence quality and its downstream consequences: recalls and adverse events

Across both US and European systems, AI devices have largely entered the market on thin clinical evidence. Among 691 FDA-cleared AI/ML devices through 2023, 46.7% of decision summaries did not report study design, 95.5% reported no demographic information, only 1.6% cited randomised trials and 7.7% prospective studies; post-market, 489 adverse events involved 36 devices (5.2%) and 40 devices (5.8%) were recalled 113 times, with 75.2% of recalls due to software issues (Direct; PMID: 41004181). A generalizability-focused analysis of 903 FDA-listed devices similarly found clinical performance studies reported for 55.9% of devices, 24.1% explicitly with none, only 2.4% randomised, and sex- and age-subgroup data in just 28.7% and 23.2% of studies respectively (Direct; PMID: 40305017). In radiology, of 723 AI devices among 950 authorisations, only 5% underwent prospective testing, 8% included a human-in-the-loop and 29% any clinical testing, with 97% cleared via 510(k) (Direct; PMID: 41201805). Predicate-network analysis found that a third of AI/ML devices were cleared on non-AI/ML predicates in the first generation, with AI/ML tasks changing in 47.8% of radiology predicate connections — 'predicate creep' that raises safety concerns (Direct; PMID: 37625896). A US–Europe comparison of 462 devices (222 FDA, 240 CE-marked) found only 1% qualified as high-risk in either jurisdiction, and of 124 devices approved in both, 80 were CE-marked first, a median of 137 days before 510(k) clearance — a pattern the authors attributed to potentially less rigorous decentralised European evaluation (Direct; PMID: 33478929).

Recall data connect weak premarket evidence to post-market burden. A 27-year US recall analysis found software design was the single largest root cause for AI/ML device recalls (42%), and design-plus-software-design causes accounted for 50% of recalls, higher than for general 510(k) or software-related devices (Direct; PMID: 40644609). A cross-sectional study of 950 AI-enabled devices found 60 (6.3%) associated with 182 recalls, 43.4% occurring within the first 12 months — roughly double the rate for all 510(k) devices — and lack of reported clinical validation (OR 2.8; 95% CI 1.6–4.7) and publicly traded manufacturer status (OR 5.9; 95% CI 2.4–14.6) independently associated with recall (Direct; PMID: 40844774). Notably, this association is contested in direction: the 691-device analysis reported adverse-event and recall rates for AI devices lower than for non-AI 510(k) devices, and the authors interpreted the low absolute numbers as a warning about inadequate surveillance rather than evidence of safety (Direct; PMID: 41004181). Predicate-network work likewise concluded it was too early to establish whether AI/ML devices are recalled more often than other devices (Direct; PMID: 37625896). Transparency evaluations — which are opinion- and audit-style rather than outcome studies — consistently report low disclosure: modest, incomplete improvement after the FDA's 2021 Good Machine Learning Practice guidance (Derived; PMID: 41249460); race/ethnicity and prospective post-market data rarely reported across 692 clearances (Derived; PMID: 39362934); and median transparency far below complete for CE-certified radiology AI products, with the largest gaps in training data, ethics and deployment caveats (Derived; PMID: 38455991). FDA's own workshop on transparency framed the stakeholder needs such disclosures must serve, while noting this is guidance-level rather than enforceable evidence (Direct; PMID: 38273098).

The EU model in depth: MDR obligations, the AI Act overlay, and professional-society proposals

The EU provides the most fully articulated lifecycle architecture in the corpus. Legal analyses describe how AI-enabled medical devices become 'high-risk' AI systems, triggering quality management, risk management, data governance, transparency, human oversight and post-market monitoring obligations under the AI Act in addition to MDR conformity assessment; the AI Act's QMS requirement may be integrated with existing ISO 13485 systems, and its extraterritorial reach captures non-EU manufacturers (Direct; PMID: 39242831). Implementation risks are documented: notified-body capacity constraints, ambiguity about 'substantial modification' of adaptive systems, and the reality that several years after MDR introduction large shares of previously certified devices lacked new certification (Direct; PMID: 41583580).

Professional-society guidance operationalises these duties. A European Society of Radiology Delphi (14 of 16 invited experts; 12 consensus statements) recommends that a PMS platform be available at deployment, that data collection occur at institutional rather than individual-volunteer level, that providers publish baseline accuracy metrics with uncertainty measures such as 95% confidence intervals, and that periodic PMS reviews be performed every 6–12 months; AI Act deployers must additionally retain system logs for at least six months (Direct; PMID: 41385025). A companion ESR statement on the AI Act, containing advocacy positions rather than empirical findings, calls for harmonising QMS requirements between the MDR and AI Act, model cards for transparency, domain-expert involvement in national sandboxes, and population-level infrastructures for post-market monitoring, noting over 200 CE-certified AI imaging applications already exist (Direct; PMID: 39948192). A doctrinal case study of the two EU regimes found convergence on documentation and performance monitoring but divergence — the AI Act carrying human-oversight duties the MDR lacks, the MDR requiring reporting of non-serious incident patterns the AI Act does not — plus guidance gaps on system updates and healthcare-provider responsibilities (Derived; PMID: 41714355). Health-sector commentaries concur that implementation guidance remains unclear and resource-dependent across member states (Derived; PMID: 40703136; PMID: 38778162).

India's current foundations and their gaps

India possesses the raw institutional components for an AI-device post-market framework but they are fragmented. The MvPI, launched on 6 July 2015, coordinates device adverse-event reporting through roughly 380 reporting centres, yet a knowledge–attitude–practice survey of 370 healthcare workers at a Gujarat tertiary hospital found adequate knowledge overall but a wide cadre gap (doctors scored 5.7 ± 1.21 vs 3.76 ± 1.50 for paramedical staff on a 7-point scale, P < 0.05), lack of knowledge cited by 37.30% as the main reason for underreporting, and only 25.10% of doctors able to name even one reporting centre — despite 85.13% expressing willingness to report (Direct; PMID: 41221570). Consistent with this, a northern-Indian ambispective hospital study reported that reporting to the national MvPI was negligible during its retrospective phase, and an ICU study of 90 respiratory-disorder patients documented 152 device adverse events, of which 22.4% were serious, with comorbidities and hyper-polypharmacy significantly associated with events (P = 0.04) (Direct; PMID: 38881137; PMID: 40878347). Abstract-level evidence additionally indicates that cardiac stent adverse events reported through the national system were predominantly serious, and that individual-level, organisational and resource barriers depress reporting generally (Derived; PMID: 39828893; PMID: 39259332). At the system level, the comparative database analysis cited above concludes that the MvPI 'lacks the comprehensiveness of platforms like EUDAMED' and recommends a centralised national registry supported by digital-health initiatives (Direct; PMID: 41100654).

Assessment capacity is emerging around, rather than within, device regulation. An expert elicitation for a digital-health health-technology-assessment (DHT-HTA) framework in India (17 of 20 invited experts responding) found the highest-weighted needs included a framework for classifying digital health products by patient risk (15/17), data-security protocols (13/17) and collaboration with end-users (14/17), while noting that the Materiovigilance infrastructure is a foundation but that integration into decision-making processes remains difficult (Direct; PMID: 39494593). A rapid HTA of AI-assisted chest X-ray interpretation for TB screening demonstrates that accelerated, India-adapted assessment methods covering clinical and economic evidence are already being applied to AI diagnostic tools in the Indian public system (Direct; PMID: 41425285). A specialty-level signal also exists: an IADVL Academy position statement for AI in Indian dermatology sets out recommendations on clinician-in-the-loop use, prospective multicentric validation in representative Indian populations, skin-of-colour inclusivity and post-market surveillance — an example of professional-body governance forming in parallel to national regulation (Derived; PMID: 42714465). India's broader digital-health ecosystem — the Ayushman Bharat Digital Mission, the Global Initiative on Digital Health launched in 2023, and domestic digitally enabled clinical platforms — provides infrastructure on which surveillance could be built, though the reviewed evidence does not establish that these currently support device-level AI performance monitoring (Direct; PMID: 39882493). LMIC-focused analyses identify the wider constraints such a framework must plan around: imaging-device scarcity, unstable power and bandwidth, small and non-representative datasets, workforce shortages and regulatory-capacity gaps (Direct; PMID: 41128758; PMID: 38186931).

Candidate framework components available in the literature

Several tested or consensus-based instruments could seed an Indian risk-based post-market framework. FUTURE-AI, built by 117 experts from 50 countries through eight Delphi rounds with under 5% final disagreement, offers 30 lifecycle recommendations across fairness, universality, traceability, usability, robustness and explainability — including continuous risk management, periodic auditing with defined datasets and timelines, privacy-preserving usage logging, and explicit assignment of accountability for AI-related errors among clinicians, institutions, developers and manufacturers (Direct; PMID: 39909534). At the institutional tier, FAIR-AI provides an operational triage: ten low-risk screening questions, an in-depth review for solutions that fail screening, three risk categories (low/moderate/high) with escalation of high-risk tools to a governance committee, a periodic monitoring attestation by the business owner, and an 'AI Label' for high-risk deployments; in early use about half of reviewed solutions were triaged low-risk (Direct; PMID: 40790350). The multi-society radiology statement (ACR, CAR, ESR, RANZCR, RSNA) adds concrete monitoring mechanics: at minimum annual re-evaluation of every deployed model, surveillance of input-data drift triggers (equipment, protocols, demographics), defined escalation and resolution pathways on detected degradation, and the caution that such monitoring is a position statement, not an empirically validated intervention (Direct; PMID: 38251899). A scoping review of 72 guidance documents for AI prediction models maps quality criteria across six phases (data preparation, development, validation, software development, impact assessment, implementation) and flags that software development, impact assessment and implementation phases have received far less guidance than the first three — a gap directly relevant to post-market framework design (Derived; PMID: 35013569); the TRIPOD+AI checklist supplies the corresponding reporting standard (Derived; PMID: 38626948).

For the AI-specific surveillance layer, the evidence supports augmenting event reporting with periodic, population-level performance monitoring. The classifier-based distribution-shift detection methods evaluated on retinal images are described by their authors as readily implementable components of a manufacturer's post-market surveillance, with the important caveat that detecting a shift does not itself establish harm and thresholds require calibration (Direct; PMID: 38724581). The FDA-focused governance analysis proposes quarterly manufacturer reports on training-data updates and deployment conditions, plus nutrition-label disclosure of training populations (Direct; PMID: 40450160); the critical FDA report advocates mandated post-market monitoring with approval updates on detected drift, training-data disclosure and subgroup performance evaluation — again, an advocacy position rather than evaluated policy (Direct; PMID: 40471897). Ethics-level requirements from the WHO-convened forum — AI impact assessments, environmental impact consideration, community engagement and fair international partnerships — define the governance envelope within which such technical mechanisms should sit, particularly for LMICs (Direct; PMID: 38637857). Health-system survey evidence (abstract-level) indicates that even in high-income systems, immature tools are the most-cited barrier to adoption, tempering expectations for rapid framework uptake (Derived; PMID: 40323320).

Synthesis and outlook: designing and validating an India-specific framework

Where the field agrees. Across jurisdictions and disciplines, the reviewed evidence converges on four propositions: (1) AI-specific failure modes — drift, shift and subgroup instability — require population-level performance monitoring that event-based adverse-event reporting cannot deliver (Direct; PMID: 40450160; PMID: 38724581); (2) premarket evidence for cleared AI devices is thin and inconsistently disclosed in both the US and EU, and weak validation is associated with early recalls (Direct; PMID: 41004181; PMID: 40305017; PMID: 40844774); (3) risk-stratified obligations, institutional-level data collection and periodic (6–12-month) performance review are the emerging operational core of credible oversight (Direct; PMID: 40790350; PMID: 41385025; PMID: 38251899); and (4) India's MvPI, HTAIn and digital-health infrastructure constitute a usable foundation, but are not yet an integrated AI-device post-market regime (Direct; PMID: 41221570; PMID: 41100654; PMID: 39494593).

Where the field disagrees or is uncertain. Whether AI devices are intrinsically less safe than conventional 510(k) devices is unresolved: one analysis finds lower reported adverse-event rates and warns this reflects surveillance inadequacy, another finds recall rates roughly double those of all 510(k) devices in the first year, and a third judges the follow-up too short to conclude anything (Direct; PMID: 41004181; PMID: 40844774; PMID: 37625896). Jurisdictional stringency is likewise contested — the EU-first pattern of dual-approved devices may reflect weaker European evaluation, but the comparison's own authors state that evidence-based conclusions require more analysis (Direct; PMID: 33478929). Finally, no framework in this corpus has been empirically validated: FUTURE-AI, FAIR-AI and the multi-society guidance are consensus or institutional positions, and the distribution-shift detection tools have been demonstrated only in simulation (Direct; PMID: 39909534; PMID: 40790350; PMID: 38724581).

Where the field is silent. No reviewed source reports an AI-specific risk classification, adaptive-update pathway or post-market performance-monitoring obligation for India; none quantifies the performance of any AI device post-deployment in Indian clinical settings; and China's adaptive-update and AI-performance surveillance obligations are not reported at all in the retrieved evidence.

Aims for the proposed project. Against this gap, the logical aims of a development-and-validation project are: to specify tiered post-market obligations per risk tier, combining event reporting through MvPI with periodic performance/deployment reporting and drift detection proportional to tier (Direct; PMID: 40450160; PMID: 38724581; PMID: 41385025); to integrate institutional-level oversight using FAIR-AI-style screening, monitoring attestations and AI labels, feasible within Indian hospital IT constraints documented by the LMIC evidence (Direct; PMID: 40790350; PMID: 41128758); and to validate the framework through a modified Delphi with Indian domain experts and, where feasible, pilot application to deployed tools already assessed by HTAIn — mirroring the validation methods used for the Indian DHT-HTA framework and the ESR PMS statements (Direct; PMID: 39494593; PMID: 41385025). Any such framework must be treated as a living instrument: every reviewed governance author notes that AI technology and regulation are evolving faster than the evidence base supporting them (Direct; PMID: 38778162).

Landscape check

The literature landscape maps how the gathered papers connect through shared genes, chemicals, diseases and processes, reported mechanisms, and citations. These notes compare the brief with that map.

Central papers not cited
- The state of artificial intelligence-based FDA-approved medical devices and algorithms: an online database (2020): linked to 0 other papers in this set.
- The European artificial intelligence strategy: implications and challenges for digital health (2020): linked to 0 other papers in this set.

Sources set aside as unconnected
- Trust in and Acceptance of Artificial Intelligence Applications in Medicine: Mixed Methods Study (2024) shares no gene, chemical, disease, process, citation or topic with any other paper gathered here, so it was not used in the answer.
- Federated benchmarking of medical artificial intelligence with MedPerf (2023) shares no gene, chemical, disease, process, citation or topic with any other paper gathered here, so it was not used in the answer.
- Toward a responsible future: recommendations for AI-enabled clinical decision support (2024) shares no gene, chemical, disease, process, citation or topic with any other paper gathered here, so it was not used in the answer.
- The Vital Role of Materiovigilance in Orthopedic Practice (2025) shares no gene, chemical, disease, process, citation or topic with any other paper gathered here, so it was not used in the answer.
- Digital Health Revolution in India: Transforming Health and Medicine (2024) shares no gene, chemical, disease, process, citation or topic with any other paper gathered here, so it was not used in the answer.
- Elevating Orthopaedic Care through Materiovigilance (2025) shares no gene, chemical, disease, process, citation or topic with any other paper gathered here, so it was not used in the answer.

Cited sources unconnected to the rest of this set
- Privacy Protection in Using Artificial Intelligence for Healthcare: Chinese Regulation in Comparative Perspective (2022).
- Global trends in post-market surveillance of high-risk medical devices: An empirical analysis based on regulatory data (2025).
- Cost-effectiveness analysis of AI-assisted chest X-ray interpretation tools for TB screening: a rapid HTA (2025).

Research notebook

What specific risks, performance monitoring challenges, and post-market evidence requirements are unique to AI/ML-based medical devices (algorithm drift, dataset shift, bias, real-world performance monitoring, adaptive algorithms)?

Distribution shift is a measurable, monitorable driver of AI device performance degradation: a retrospective simulation study of distribution shift detection for postmarket surveillance of medical AI algorithms demonstrates that shift detection methods applied to routinely collected deployment data can flag performance decay, supporting population/instance-level monitoring as a complement to event-based adverse event reporting.

Status: verified • Confidence: medium

• Analysis of 943 MDRs linked to 823 FDA-cleared AI/ML devices (2010-2023) found extreme missing data (event location missing in 100% of AI/ML reports vs 90% overall; reporter health-professional status missing 73% vs 43%), misclassified events (e.g. analyst error reported as device injury), and no mechanism to report concept drift, covariate shift, or algorithmic instability, which the authors argue requires periodic deployment-condition reporting or a new post-market regime. (PMID 40450160, results)
• Analysis of FDA medical device reports identified AI-related safety issues, showing that most AI device adverse event reports capture malfunction-type problems rather than the algorithmic performance failures most characteristic of AI devices. (PMID 39627534, abstract)
• An analysis of safety events involving ML-enabled medical devices reported to the FDA found that user errors and the way AI/ML devices are applied in practice are particularly likely to cause harm, indicating that post-market signals are dominated by human-device interaction rather than model failure per se. (PMID 37071804, abstract)
• A retrospective simulation study developed and evaluated distribution shift detection methods for postmarket surveillance of medical AI algorithms, showing that monitoring deployment data for covariate/concept shift can identify performance degradation that individual adverse-event reporting cannot capture. (PMID 38724581, abstract)
• The authors identify concept drift, covariate shift and algorithmic stability as the three root AI-specific failure modes invisible to individual-level adverse event reporting, and propose quarterly manufacturer 'deployment conditions' reports as a remedy. (PMID 40450160, results)
• A generalizability assessment of AI models across hospitals in a low-middle and a high income country demonstrated significant performance degradation when models were deployed in settings different from their training environment, quantifying the dataset-shift risk in real-world deployment. (PMID 39333515, abstract)

How does the US FDA regulate AI-enabled medical devices and SaMD, including post-market surveillance, Predetermined Change Control Plans (PCCPs), and the 510(k)/De Novo pathways for adaptive AI?

Post-market signal data further challenge the current FDA regime: 27 years of AI/ML-enabled device recall records and analyses of early recalls show that AI devices experience recalls linked to premarket clinical validation gaps, and systematic appraisals of FDA AI approvals find limited and inconsistently reported clinical validation and benefit-risk information - together providing empirical grounding for risk-proportionate post-market obligations in any new framework.

Status: verified • Confidence: medium

• A taxonomy across 1,016 FDA authorizations characterized how AI is used in FDA-authorized medical devices, providing the most complete map of the authorized AI device portfolio. (PMID 40596700, abstract)
• A review of FDA approvals of AI/ML devices in radiology found that radiology accounts for the majority of AI device authorizations and that most are cleared via 510(k) with limited premarket clinical evidence. (PMID 41201805, abstract)
• An analysis of ML-enabled devices authorized by the FDA in 2024 described regulatory characteristics, predicate lineage, and transparency reporting, finding continued reliance on 510(k) predicate chains and inconsistent transparency about model details. (PMID 41463017, abstract)
• Analysis of FDA-cleared AI/ML-based medical devices and their 510(k) predicate networks showed that most AI devices trace to predicates through long chains, raising questions about how safety and performance evidence accumulates over iterative clearances. (PMID 37625896, abstract)
• Guidance-oriented work on Predetermined Change Control Plans describes how manufacturers can pre-specify, in the marketing submission, the anticipated model modifications and the validation/update protocols, enabling iterative AI improvement without new premarket submissions. (PMID 41171641, abstract)
• A 27-year analysis of AI/ML-enabled medical device recalls in the United States characterises recall reasons and trends and draws implications for governance, providing empirical evidence that AI devices generate conventional post-market safety signals (recalls) that must be tracked alongside algorithmic performance monitoring. (PMID 40644609, abstract)
• An analysis of early recalls in AI-enabled medical devices links early post-market recalls to gaps in premarket clinical validation, supporting the argument that inadequate clinical validation at authorisation translates into downstream post-market burden. (PMID 40844774, abstract)
• A report on AI healthcare product approvals to the FDA ('the illusion of safety') documents limited clinical validation evidence in AI device authorisations, arguing that current approvals convey an unwarranted impression of established safety. (PMID 40471897, abstract)
• An assessment of benefit-risk reporting for FDA-cleared AI-enabled devices finds inconsistent or limited benefit-risk communication in authorisation summaries, a transparency gap relevant to post-market oversight design. (PMID 41004181, abstract)

How does the EU regulate AI-enabled medical devices under MDR/IVDR and the AI Act (risk classification, conformity assessment, post-market surveillance, PMS reports/PMCF)?

The EU regulates AI-enabled medical devices through the interlocking MDR/IVDR and the AI Act: MDR imposes risk-class-based conformity assessment with mandated post-market surveillance plans, PSURs and PMCF, while the AI Act adds a horizontal, risk-based regime in which most AI-enabled medical devices are high-risk, requiring quality/risk management, data governance, transparency, human oversight and post-market monitoring - with residual ambiguities about how the two regimes overlap.

Status: verified • Confidence: medium

• Analysis of the EU AI Act's implications for regulated digital medical products explains how AI-enabled medical devices fall into the high-risk category and must satisfy AI Act requirements in addition to MDR conformity assessment, and identifies open questions about coordinating the two regimes. (PMID 39242831, abstract)
• A Health Policy analysis of the EU Artificial Intelligence Act (2024) sets out its risk-based architecture and its implications for healthcare, including obligations for high-risk medical AI and post-market monitoring duties. (PMID 39244818, abstract)
• A mapping of the regulatory landscape for AI in health in the EU shows that the MDR already requires manufacturers to run post-market surveillance plans, periodic safety update reporting and PMCF, and situates the AI Act's additional duties on top of this. (PMID 39191937, abstract)
• European Society of Radiology consensus recommendations on AI medical device post-market surveillance call for structured, risk-proportionate PMS including performance monitoring plans for AI devices under MDR, noting that conventional PMS tools inadequately capture AI performance issues. (PMID 41385025, abstract)
• Commentary on the passage of the EU AI Act argues that its implications for digital medicine remain unclear, particularly how sectoral device law and horizontal AI law will interact in conformity assessment and notified body practice. (PMID 38778162, abstract)

What is the current regulatory landscape and post-market surveillance framework for medical devices (especially AI-enabled/SaMD) in India (CDSCO, Medical Device Rules 2017, Drugs and Cosmetics Act amendments)?

India-specific assessment instruments for AI medical devices are beginning to appear: rapid health technology assessment methods have been applied to AI-assisted diagnostic tools in the Indian health system, and Indian specialty societies have started issuing their own position statements and policy frameworks for safe AI adoption - indicating that the building blocks (HTAIn, professional-body frameworks, materiovigilance) exist but remain fragmented and device-sector-specific rather than an integrated risk-based post-market regime.

Status: verified • Confidence: medium

• A comparative analysis of post-market incident trends across EUDAMED, MAUDE, BfArM and ANSM concludes that 'the Indian context is particularly relevant: while the Materiovigilance Programme of India (MvPI) provides a foundation, it lacks the comprehensiveness of platforms like EUDAMED', and recommends establishing a centralised national registry supported by digital health initiatives and regulatory reform to enhance early detection, field safety corrective action coordination and international alignment. (PMID 41100654, discussion)
• An expert elicitation (17/20 respondents) building a digital-health HTA framework for India found that the highest-weighted needs included risk classification of digital health products (endorsed by 15/17 experts), regulatory frameworks for digital diagnosis/monitoring software, data storage/security protocols, and adverse-event/materiovigilance mechanisms, and documents India's institutional landscape (HTAIn, NDHM/ABDM, NITI Aayog's National Strategy for AI) alongside capacity gaps. (PMID 39494593, results)

What are the regulatory approaches of other jurisdictions (UK MHRA, Health Canada, Japan PMDA, Singapore HSA, Australia TGA, WHO guidance) for AI medical devices, and what common principles and divergences emerge from global comparisons?

Beyond the FDA and EU, other jurisdictions regulate AI medical devices through distinct instruments - China has an increasingly specified AI SaMD regulatory framework with global manufacturer implications, and international comparative work shows convergence on risk-proportionate oversight and lifecycle monitoring but divergence in pathways, transparency requirements and evidence expectations; consensus initiatives such as FUTURE-AI articulate transversal trustworthy-AI principles spanning jurisdictions.

Status: verified • Confidence: medium

• A comparative analysis of China's regulatory frameworks for AI-enabled medical device software reviews NMPA requirements, classifies China's approach relative to other jurisdictions, and draws out implications for global manufacturers, documenting an increasingly formalised, jurisdiction-specific pathway for AI SaMD. (PMID 39073860, abstract)
• A comparative analysis of AI/ML-based medical device approvals in the USA and Europe (2015-2020) found substantial differences in device classification and evidence expectations between the FDA and EU pathways even for the same technologies, evidencing cross-jurisdictional divergence. (PMID 33478929, abstract)
• The FUTURE-AI international consensus guideline (co-developed by multidisciplinary stakeholders across multiple continents) establishes consensus recommendations for trustworthy and deployable AI in healthcare covering technical, clinical and governance dimensions, providing a transversal reference framework usable across regulatory jurisdictions. (PMID 39909534, abstract)
• A comparative analysis of privacy protection in healthcare AI, centred on Chinese regulation in comparative perspective, shows how data-governance rules differ substantially across jurisdictions and interact with device-level AI regulation. (PMID 36292325, abstract)

What frameworks, maturity models, or validated assessment tools have been proposed in the literature for risk-based oversight/audit of AI medical devices that could inform an India-specific framework, and what gaps exist in low- and middle-income countries?

The literature offers several operational, risk-based frameworks for oversight/audit of AI medical devices that an India-specific framework could draw on: a general governance framework for marketed AI/ML devices built on lifecycle surveillance, a practical hospital-facing FAIR-AI implementation/review framework, multi-society guidance covering the full deployment cycle from procurement to ongoing monitoring, and consensus quality criteria for AI prediction models - while LMIC-focused analyses document persistent regulatory-capacity, infrastructure and data gaps that such frameworks must be adapted to.

Status: verified • Confidence: medium

• A general framework for governing marketed AI/ML medical devices proposes lifecycle post-market governance that complements premarket authorisation, arguing adverse-event reporting must be augmented with periodic performance/deployment reporting for AI-specific failure modes. (PMID 40450160, abstract)
• The FAIR-AI framework provides a practical structure for appropriate implementation and review of AI in healthcare, covering institution-level assessment, implementation and periodic re-review of deployed AI tools - a hospital-facing oversight instrument that complements national regulation. (PMID 40790350, abstract)
• A multi-society statement from the ACR, CAR, ESR, RANZCR and RSNA sets out practical considerations for developing, purchasing, implementing and monitoring AI tools in radiology, including acceptance testing, ongoing performance monitoring and vendor accountability - a template for institutional risk-based AI oversight. (PMID 38251899, abstract)
• A scoping review of guidelines and quality criteria for AI-based prediction models in healthcare consolidates the quality criteria (transparency, validation, reporting standards) against which AI medical device frameworks and audits can be benchmarked. (PMID 35013569, abstract)
• A perspective on AI implementation in medical imaging in LMICs identifies challenges (infrastructure, workforce, data representativeness, regulatory preparedness) and priorities/strategies for LMIC-specific AI oversight, documenting the gaps an India-specific framework must address. (PMID 41128758, abstract)
• A literature review of AI in laboratory medicine in LMICs enumerates challenges and recommendations - including regulatory and quality-management gaps - for safe AI adoption in resource-constrained settings. (PMID 38186931, abstract)