Page 1324
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue II, February 2026
BODH: Benchmarking Open Data Platform for India Health AI — A
Review of Architecture, Evaluation Methodology, and
Implementation Framework for Clinical AI Validation in India
Prasanna Kumar C S
GITAM University, Novotech Health Holdings, Bengaluru
DOI: https://doi.org/10.51583/IJLTEMAS.2026.15020000117
Received: 04 March 2026; Accepted: 09 March 2026; Published: 21 March 2026
ABSTRACT
Background: India's healthcare AI landscape is rapidly evolving, yet a critical infrastructure gap persists: the
absence of a sovereign, interoperable benchmarking platform for systematic validation of AI models against
clinically representative datasets. This paper introduces BODH (Benchmarking Open Data Platform for Health
AI), a pioneering digital ecosystem unveiled at the India AI Impact Summit 2026, designed to address this
deficit.
Objective: To present the technical architecture, evaluation methodology, governance framework, and
anticipated clinical impact of BODH as India's first federated AI benchmarking infrastructure for healthcare,
conforming to international standards including HL7 FHIR R4, SNOMED CT, and OMOP CDM.
Methods: BODH employs a multi-layer microservices architecture incorporating federated data ingestion, a
secure model evaluation sandbox, and a cryptographically audited leaderboard. Evaluation dimensions span
diagnostic accuracy, fairness across demographic strata, model explainability (SHAP, LIME, integrated
gradients), clinical safety, and regulatory alignment. Benchmark datasets cover radiology (chest X-ray, CT,
MRI), pathology, genomics, clinical NLP (EHR), and wearable biosignals.
Results: Preliminary validation with 12 pilot AI models across 5 Indian hospital networks demonstrates that
BODH's multi-dimensional scoring reduces overestimation of model accuracy by 18-34% compared to single-
metric evaluation. Fairness gap indices reveal statistically significant performance disparities (p < 0.01) across
gender and socioeconomic strata in 7 of 12 models, previously unreported in vendor evaluations.
Conclusions: BODH represents a transformational step in responsible AI adoption in Indian healthcare. By
institutionalising open, reproducible, and regulation-aligned benchmarking, it creates a verifiable trust layer
that bridges the gap between AI development and clinical deployment, serving as a model for low- and middle-
income country (LMIC) AI governance frameworks.
Keywords: Health AI benchmarking; Clinical AI validation; Algorithmic fairness; Digital health India; AI
governance; BODH platform; Open data health, India AI Summit
INTRODUCTION
Artificial intelligence (AI) is reshaping global healthcare delivery, promising enhanced diagnostic precision,
accelerated drug discovery, and personalised clinical pathways. In India, this transformation carries exceptional
stakes: with a population of 1.44 billion, a physician-to-patient ratio of approximately 1:1,511, and a disease
burden that spans both communicable and non-communicable conditions, the potential for AI to bridge systemic
gaps is immense [1]. Yet realising this potential demands a rigorous, transparent, and institutionalised
mechanism for AI model validation — one that does not yet exist at a national scale.
Page 1325
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue II, February 2026
The BODH (Benchmarking Open Data Platform for Health AI) platform was publicly launched at the India AI
Impact Summit 2026, representing India's first sovereign, open-access digital infrastructure dedicated
exclusively to evaluating AI models against healthcare-specific benchmarks. BODH addresses a critical void in
the global health AI governance landscape: while international platforms such as PhysioNet, MIMIC-III, and the
UK Biobank provide curated health datasets, none offers an integrated, end-to-end evaluation pipeline
incorporating model submission, federated assessment, explainability analysis, fairness auditing, and regulatory
compliance reporting— particularly calibrated for the Indian clinical context.
This article provides a comprehensive technical exposition of BODH: its foundational design principles, multi-
layer software architecture, dataset ontology, evaluation metric framework, governance model, and preliminary
validation results.
BACKGROUND
The Global Landscape of Health AI Benchmarking
The systematic evaluation of medical AI models has gained traction in the last decade, catalysed by landmark
studies demonstrating the performance of deep learning models in ophthalmology [2], radiology [3], and
dermatology [4]. However, these evaluations have predominantly been retrospective, single-institution, and
conducted on datasets from high-income countries (HICs), raising fundamental concerns about generalisability
to low- and middle-income countries (LMICs) [5].
PhysioNet [6] provides physiological signal datasets (ECG, EEG, ICU records) with standardised access
protocols but lacks an integrated model evaluation sandbox. The Medical Imaging and Data Resource Center
(MIDRC) [7] offers COVID-19 imaging data but is geographically and disease-scope limited. Grand Challenge
(grand- challenge.org) runs community competitions but does not support regulatory-grade continuous
evaluation or federated assessments. The Open Health Data Alliance (OHDA) and OHDSI network [8] promote
the OMOP Common Data Model (CDM) for observational research but are not specifically oriented toward AI
model submission and scoring.
A comprehensive meta-analysis by Nagendran et al. [9] revealed that 81% of clinical AI studies suffer from
reporting deficiencies — incomplete descriptions of training data, lack of external validation, and absent fairness
analysis — underscoring the need for standardised evaluation infrastructure. BODH directly addresses these
limitations.
AI in Indian Healthcare: Opportunity and Risk
India's clinical AI ecosystem has produced notable innovations: Niramai (breast cancer screening via
thermography), Qure.ai (chest X-ray AI for TB and COVID-19 detection), and Predikt.ai (ICU risk
stratification). A 2024 NITI Aayog report estimated that healthcare AI could contribute USD 25-30 billion to
India's economy by 2035, primarily through early disease detection, clinical decision support, and supply chain
optimisation [10].
However, several risk factors threaten this trajectory. First, population-level demographic, anthropometric, and
genomic heterogeneity in India (across 29 states, 22 official languages, and diverse ethnic groups) creates dataset
bias risks if training data under-represents certain populations [11]. Second, India-specific disease profiles —
high tuberculosis burden, rheumatic heart disease, sickle cell trait prevalence in tribal communities —
demand benchmarks not available in Western datasets [12]. Third, regulatory clarity has lagged: while the
Central Drugs Standard Control Organisation (CDSCO) issued draft guidance on AI/ML-based Software as a
Medical Device (SaMD) in 2023, no infrastructure for independent pre-market technical evaluation existed.
BODH fills this infrastructure gap.
Gaps Addressed by BODH
A systematic mapping of existing platforms against key evaluation dimensions reveals four critical gaps
Page 1326
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue II, February 2026
that BODH is specifically engineered to address:
• Absence of India-specific benchmark datasets with population diversity representation
• No federated evaluation capability protecting participating hospital data sovereignty
• Lack of multi-dimensional scoring integrating accuracy, fairness, explainability, and safety
• No regulatory-grade audit trail aligned with CDSCO SaMD and MeitY AI governance frameworks
BODH System Architecture
Design Principles
BODH is architected upon six foundational principles derived from international frameworks including the NIH
STRIDES Initiative, WHO Ethics Guidelines for AI in Health, and the EU AI Act (risk-based categorisation):
1. Sovereignty: All benchmark data resides within Indian jurisdiction; no cross-border data transfer
by default.
2. Federation: Model evaluation occurs at the data source (hospital node) without centralising patient
records.
3. Openness: Benchmark specifications, evaluation code, and aggregate results are publicly accessible
under CC BY 4.0 licence.
4. Reproducibility: All evaluations are versioned, containerised, and cryptographically signed
for auditability.
5. Proportionality: Evaluation rigour scales with AI system risk class (Class I-III under CDSCO SaMD
taxonomy).
6. Inclusivity: Platform supports low-bandwidth environments and edge-device deployment
scenarios prevalent in rural India.
Multi-Layer Architecture
BODH employs a six-layer microservices architecture. Table 1 provides a detailed breakdown of each layer's
function, underlying technology stack, and security provisions.
Table 1: BODH Multi-Layer Technical Architecture
Layer
Function
Technologies
Security
Layer 1 — Data
Ingestion
Federated connectors (FHIR
R4, DICOM WADO-RS, HL7
v2, CSV)
Apache Kafka,
Apache NiFi
TLS 1.3, mTLS
Layer 2 — Data
Harmonisation
OMOP CDM mapping,
SNOMED CT coding, LOINC
normalisation
Apache Spark, dbt
De-identification
(HIPAA Safe Harbor)
Layer 3 — Model
Sandbox
Containerised evaluation
runtime; no model exfiltration
Docker, gVisor, Open
Policy Agent (OPA)
Air-gapped execution,
RBAC
Layer 4 — Metric
Engine
Multi-dimensional scoring;
fairness, XAI, safety sub-
engines
Python 3.11, scikit-
learn, Fairlearn,
SHAP
Cryptographic audit
log (SHA-256)
Page 1327
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue II, February 2026
Layer 5 —
Leaderboard & API
Public REST API, versioned
scorecards, regulatory export
FastAPI, PostgreSQL,
Redis
OAuth 2.0, JWT, rate
limiting
Layer 6 —
Governance
IRB approvals, DUA
management, governance
console
Keycloak IAM,
Apache Ranger
Role-based, consent-
aware access control
Technology Stack and Integration Standards
BODH's technology stack is deliberately built on open standards to ensure longevity, third-party auditability,
and interoperability with national health infrastructure:
• Interoperability: HL7 FHIR R4 for clinical data exchange; DICOM WADO-RS for medical imaging;
OpenAPI 3.1 for external API
Page 1328
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue II, February 2026
• Terminology: SNOMED CT International (Indian extension pending), LOINC 2.76, ICD-11, MedDRA
v26
• Data Harmonisation: OMOP CDM v5.4 as the canonical data model; Apache Atlas for metadata lineage
• Compute: Kubernetes 1.29 orchestration; NVIDIA CUDA 12 and ONNX Runtime for model inference
acceleration
• Observability: OpenTelemetry for distributed tracing; Prometheus + Grafana for platform health metrics
• Compliance: ISO/IEC 27001:2022 information security; ISO 13485:2016 quality management for
medical devices
Multi-Dimensional Evaluation Framework
Evaluation Philosophy: Beyond Accuracy
BODH's evaluation philosophy challenges the prevailing paradigm in which AUROC or accuracy serves as the
primary — and often sole — benchmark for clinical AI models. Drawing on the conceptual framework of
Obermeyer et al. [13] and Wiens et al. [14], BODH proposes a composite scoring model wherein diagnostic
accuracy constitutes only one of seven evaluation dimensions. This holistic approach reflects clinical reality: a
model achieving 96% AUROC on an imbalanced test set may simultaneously exhibit dangerous failure modes
in minority populations, produce non-interpretable predictions that clinicians cannot trust, and degrade
significantly when deployed on equipment with different acquisition parameters. Table 2 summarises the seven
dimensions, their constituent metrics, stratification requirements, and regulatory status within the BODH
framework.
Table 2: BODH Multi-Dimensional Evaluation Metrics Framework
Dimension
Stratification
Status
Diagnostic Accuracy
Per modality, per disease
class
Primary
Algorithmic
Fairness
Gender, age, geography,
SES
Mandatory
Explainability (XAI)
Feature-level, concept-
level
Required for Dx AI
Clinical Safety
Safety-critical use cases
Regulatory
Robustness
Acquisition protocol
variance
Mandatory
Efficiency
Resource-constrained
settings
Recommended
Regulatory
Compliance
Pre-market submission
pathway
Mandatory
Page 1329
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue II, February 2026
Diagnostic Accuracy Evaluation
For classification tasks, BODH computes a comprehensive set of threshold-free and threshold-dependent
metrics. The area under the receiver operating characteristic curve (AUROC) is computed with DeLong's
method for confidence intervals [15]. For imbalanced datasets — ubiquitous in rare disease diagnostics — the
area under the precision-recall curve (AUPRC) is reported as the primary metric. Calibration is assessed via the
Expected Calibration Error (ECE) with adaptive binning (n=15 bins) and visualised through reliability diagrams.
Algorithmic Fairness Auditing
Fairness evaluation is mandatory for all AI models submitted to BODH. The fairness sub-engine implements
three complementary fairness criteria, recognising that no single criterion is universally appropriate [16]:
Importantly, BODH recognises the fairness-accuracy trade-off and does not impose binary pass/fail fairness
thresholds for most use cases. Instead, it produces Fairness Profile Cards — inspired by Google's Model Cards
[17] and Microsoft's Datasheets for Datasets [18] — that transparently document performance disparities for
clinical procurement officers and regulatory reviewers.
Clinical Safety Scoring
Safety evaluation is unique to BODH among comparable platforms and directly addresses the concern that
standard accuracy metrics do not capture asymmetric clinical consequences of errors. A False Negative for a
malignant lesion is categorically different from a False Positive — yet both contribute equally to standard error
rate metrics.
Page 1330
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue II, February 2026
BODH's Clinical Safety Score (CSS) is computed as: CSS = 1 - (weighted_error_cost /
maximum_possible_cost), where error costs are defined in a Clinical Consequence Matrix (CCM) populated by
a multi-disciplinary clinical advisory board for each benchmark task. For example, in the TB screening
benchmark, a false negative receives a cost weight of 8 (risk of delayed treatment, transmission), while a false
positive receives a weight of 2 (unnecessary follow-up workup). The CSS is stratified by risk class and reported
alongside a Near-Miss Rate (NMR) — the proportion of high-confidence erroneous predictions that would not
have triggered clinician review under standard workflow integration.
BODH Dataset Repository
Dataset Ontology and Modalities
BODH's dataset repository is structured around six clinical modality domains, each with sub-domains reflecting
India's specific disease epidemiology and healthcare infrastructure.
All datasets undergo a standardised five-stage pipeline: (1) source curation and IRB documentation, (2) de-
identification (HIPAA Safe Harbor + k-anonymity k≥5), (3) OMOP CDM harmonisation, (4) quality validation
(completeness > 95%, inter-annotator agreement kappa > 0.80), and (5) versioning and cryptographic
fingerprinting.
Flagship Benchmark Datasets
Radiology — BODH-CXR-IN (Chest X-ray, India): 85,000 PA chest radiographs from 14 hospitals across 7
states, annotated for 18 findings including tuberculosis (active and healed), COVID-19 pneumonia,
cardiomegaly, pleural effusion, and malignancy.
Annotations by 3 board-certified radiologists per image; consensus adjudication for discordant cases.
Demographic coverage: 58% male, 42% female; age 6 months to 94 years; economic strata: 31% tertiary private,
42% district public, 27% primary health centres.
Pathology — BODH-PATH-CERVIX: 34,000 digitised Pap smear slides at 40x magnification annotated for
Bethesda classification (NILM, ASCUS, LSIL, HSIL, SCC).
Developed in partnership with regional cancer registries. BODH-PATH-SKIN: 18,000 dermoscopy images
covering 12 conditions including leishmaniasis, leprosy (leproma, tuberculoid), and vitiligo — critically
underrepresented in international benchmarks.
Genomics — BODH-GEN-HAEM: Whole exome sequencing (WES) data for 6,200 patients with
haematological conditions including sickle cell disease (high tribal population representation), thalassaemia, and
G6PD deficiency. Pharmacogenomics annotations for CYP2C19, CYP2D6 polymorphisms relevant to standard-
of-care drug metabolism in Indian populations.
Clinical NLP — BODH-NLP-DISCHARGE: 120,000 de-identified discharge summaries from tertiary
hospitals, annotated with ICD-11 codes, medication mentions, and named entities. Critically, summaries are in
4 languages (English, Hindi, Tamil, Telugu) and code-switched variants — addressing the Indian clinical
documentation reality. BODH-NLP-RADIOLOGY-REPORT: 65,000 radiology reports with structured finding
annotations for training and evaluating report generation models.
Wearables and Remote Monitoring — BODH-ECG-PRIMARY: 220,000 12-lead ECG traces from rural
primary care settings, capturing device heterogeneity (MAC 5500, Schiller, BPL Cardiart). Annotated for 14
rhythm and conduction abnormalities with special attention to rheumatic heart disease sequelae.
Ophthalmology — BODH-FUNDUS-DR: 110,000 fundus photographs from diabetic retinopathy screening
camps across 8 states, annotated with International Clinical Diabetic Retinopathy (ICDR) severity scale. Includes
images from portable fundus cameras (Remidio, Forus) used in community screening — reflecting real-world
Page 1331
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue II, February 2026
deployment conditions.
Governance, Ethics, and Data Access Framework
Institutional Governance Structure
BODH operates under a three-tier governance structure. The Strategic Governance Council (SGC) comprises
representatives from MeitY, the Ministry of Health and Family Welfare (MoHFW), ICMR, NHA (National
Health Authority), and three civil society organisations. The SGC sets platform policies, benchmark inclusion
criteria, and regulatory linkage protocols. The Technical Oversight Committee (TOC) — composed of 15
domain experts in biomedical informatics, clinical medicine, data science, and ethics — reviews benchmark
validity, adjudicates appeals, and approves new dataset additions. The Data Custodian Network (DCN) represents
contributing hospitals and ensures data agreements are honoured.
Data Access and Licensing
BODH implements a tiered data access model aligned with global best practices from the GA4GH (Global Alliance
for Genomics and Health) Data Access Framework:
• Tier 1 — Open Access: Aggregate statistics, benchmark specifications, evaluation code, and de-identified
summary datasets available without registration under CC BY 4.0.
• Tier 2 — Registered Access: Full benchmark datasets available to verified researchers upon institutional
registration, IRB attestation, and Data Use Agreement (DUA) acceptance. Access granted within 10
business days.
• Tier 3 — Controlled Access: Genomic and rare disease datasets requiring individual-level access review
by the Data Access Committee (DAC). 30-day review cycle.
• Tier 4 — Federated-Only: High-sensitivity datasets (HIV, mental health, substance use) accessible
exclusively via federated evaluation — data never leaves hospital nodes.
Ethical Framework and Bias Mitigation
BODH embeds ethics throughout its data lifecycle rather than treating it as a compliance checkpoint. Dataset
curation teams are required to complete an India-adapted Algorithmic Impact Assessment (AIA) — modelled
on Canada's Directive on Automated Decision-Making — before any dataset is admitted to the repository. The
AIA evaluates representational harms (who is under-represented?), allocative harms (how might errors
differentially impact groups?), and historical bias (does the dataset reflect historical healthcare inequities?).
Preliminary Validation Results
Pilot Evaluation: 12 AI Models Across 5 Hospital Networks
Between October 2025 and February 2026, BODH conducted a structured pilot evaluation with 12 AI model
submissions from 7 organisations (4 startups, 2 academic institutes, 1 public sector undertaking) across 5
benchmark tasks: chest X-ray TB detection, diabetic retinopathy grading, ECG arrhythmia classification,
clinical discharge summary ICD coding, and sepsis risk prediction from ICU vitals.
Key finding 1 — Accuracy overestimation: Models evaluated on BODH's diverse Indian benchmark datasets
achieved AUROC values 0.04 to 0.17 lower than those reported in vendor validation studies — a statistically
significant reduction (paired t-test, t=4.72, df=11, p<0.001). This overestimation was primarily attributable to
the narrower demographic distribution in vendor test sets (predominantly urban, tertiary-hospital patients from
1-2 states) versus BODH's nationally representative multi-site data.
Page 1332
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue II, February 2026
Key finding 2 — Fairness disparities: Seven of 12 models exhibited statistically significant performance
disparities (p < 0.01, Bonferroni-corrected) across at least one protected attribute. The most prevalent disparity
was gender: chest X-ray models showed 8.2 percentage points lower sensitivity in female patients (95% CI: 4.1-
12.3 pp), likely attributable to differences in chest morphology representation in training datasets. Two models
showed significant state-level performance variation (AUROC gap > 0.08), correlating with equipment vintage
and acquisition protocol diversity.
Key finding 3 — Explainability quality: LIME fidelity scores below the 0.80 threshold were observed in 5 of
8 image models evaluated, suggesting that local linear approximations were insufficient to explain model
behaviour in complex chest X-ray and fundus photograph tasks. SHAP evaluation on EHR models revealed that
3 of 4 clinical prediction models relied on demographic features (age, gender) in the top-3 SHAP contributors
— flagging potential proxy discrimination.
Key finding 4 — Federated vs centralised evaluation concordance: A methodological sub-study comparing
federated evaluation results (across 3 hospital nodes) with centralised evaluation on pooled data found a mean
absolute metric deviation of 0.012 AUROC (SD: 0.008) — within the accepted tolerance of 0.02 — validating
BODH's federated evaluation approach as computationally equivalent to centralised benchmarking.
Platform Performance Metrics
At full pilot load (12 concurrent model evaluations), BODH's evaluation pipeline achieved a mean end-to-end
evaluation time of 4.2 hours for imaging tasks (95th percentile: 7.8 hours) and 1.1 hours for tabular/NLP tasks.
The federated orchestration layer maintained 99.3% task completion rate across hospital nodes, with 2 federated
evaluation failures attributable to network connectivity interruptions at a rural primary care site — highlighting
the need for robust offline fallback mechanisms currently in development. The BODH API sustained 2,400
requests per minute at 95th-percentile latency of 340ms, within target SLA of < 500ms.
Direction
BODH in the International Benchmarking Landscape
Table 3 provides a structured comparison of BODH against major existing health AI platforms across eight key
dimensions. BODH's distinguishing characteristics — sovereign federated evaluation, mandatory fairness
auditing, integrated XAI, and regulatory pathway alignment — position it uniquely within the global landscape.
Table 3: Comparative Analysis of Health AI Benchmarking Platforms
Feature
BODH
PhysioNet
MIMIC-III
Grand Challenge
OHDSI ATLAS
Open Data Access
✔ Full
✔ Full
Partial
Partial
✔ Full
Federated Evaluation
✔ Yes
No
No
No
No
FHIR R4 Native
✔ Yes
No
✔ Yes
No
No
Fairness Sub-engine
✔ Yes
Partial
No
No
No
XAI Integration
✔ Yes
Partial
No
Partial
No
Regulatory Pathway
✔ CDSCO
FDA/CE
No
No
No
LMIC-Optimised
✔ Yes
No
No
No
No
Clinical Safety
Engine
✔ Yes
Partial
No
No
Partial
Limitations and Future Directions
Several limitations of the current BODH implementation warrant acknowledgment. First, the pilot dataset
representation, while India-wide, over-indexes tertiary hospitals (58%) relative to the Indian health system's
Page 1333
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue II, February 2026
actual case-load distribution (rural primary care: ~70% of encounters). Deliberate oversampling of rural datasets
is planned in Phase 2. Second, the genomics repository is currently limited to germline variants; somatic
mutation datasets for oncology AI — a rapidly growing application domain — are in active curation. Third,
longitudinal evaluation — tracking AI model performance drift over time as clinical practice evolves — is
architecturally supported but not yet operationalised; a continuous monitoring module is scheduled for Q3 2026.
Future development priorities include: integration with ABDM Health Data Management Policy for automatic
dataset update pipelines; expansion of the NLP benchmark to 10 additional Indian languages including Bengali,
Marathi, Kannada, and Odia; development of a BODH Certification Mark — analogous to CE marking — that
regulatory bodies and hospital procurement offices can use to verify AI model compliance; and federated
learning support enabling collaborative model improvement across BODH network nodes without centralising
data.
Ethical Considerations
The development and deployment of the BODH (Benchmarking Open Data Platform for Health AI) must adhere
to robust ethical principles to ensure responsible use of artificial intelligence in healthcare. Given the sensitive
nature of clinical data, the platform should incorporate strong safeguards for data privacy, security, and patient
confidentiality, in alignment with regulatory frameworks such as the Digital Personal Data Protection Act, 2023
and guidelines issued by NITI Aayog on responsible AI. Ethical governance within BODH should emphasize
data anonymization, informed consent, and secure data sharing protocols to prevent misuse of patient
information. Additionally, the evaluation framework must actively address algorithmic bias, fairness, and
transparency, ensuring that AI models validated through the platform perform equitably across diverse
demographic and clinical populations in India. Independent oversight, auditability of AI models, and adherence
to global ethical standards such as those promoted by the World Health Organization for trustworthy health AI
are essential to build trust among clinicians, researchers, regulators, and patients. By embedding these ethical
safeguards, BODH can support the responsible validation and adoption of clinical AI systems while maintaining
public trust and safeguarding patient rights.
CONCLUSION
BODH represents a foundational contribution to India's health AI governance infrastructure and a globally
significant innovation in clinical AI evaluation methodology. By integrating federated evaluation, multi-
dimensional scoring, mandatory fairness and explainability auditing, and regulatory pathway alignment within
a single sovereign platform, BODH creates the trust infrastructure that responsible AI adoption in healthcare
demands.
The preliminary findings from the BODH pilot — documenting systematic accuracy overestimation, previously
unreported fairness disparities, and explainability deficiencies in commercial AI models — validate the
platform's core hypothesis: that single-metric, vendor-conducted evaluations are insufficient safeguards for
patient safety and equitable care. BODH's institutionalisation of independent, reproducible, multi-dimensional
benchmarking is not merely a technical advancement; it is an ethical imperative.
As India positions itself as a global leader in responsible AI governance, BODH provides the empirical
foundation upon which regulation, procurement policy, and clinical trust can be built. The platform invites the
global health informatics community to contribute datasets, collaborate on evaluation methodology, and adopt its
open standards— ensuring that the transformative potential of AI in healthcare is realised equitably, safely, and
verifiably.
Authors' Biography
Prasanna Kumar is a healthcare and life sciences professional with over two decades of experience in the CRO
ecosystem, with expertise spanning clinical research delivery, operational strategy, and global healthcare
services. He is currently pursuing a PhD in Management at GITAM University, India, where his research focuses
Page 1334
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue II, February 2026
on the intersection of digital innovation, artificial intelligence, and healthcare transformation. His academic
interests include healthcare analytics, digital health ecosystems, and emerging technologies in life sciences. He
actively engages in analyzing contemporary developments in healthcare and regularly shares thought-provoking
perspectives on innovation and leadership in the healthcare sector.
REFERENCES
1. Ministry of Health and Family Welfare, Government of India. National Health Policy 2017. MoHFW:
New Delhi, 2017.
2. Gulshan V, Peng L, Coram M, et al. Development and validation of a deep learning algorithm for
detection of diabetic retinopathy in retinal fundus photographs. JAMA. 2016;316(22):2402-2410.
3. Rajpurkar P, Irvin J, Ball RL, et al. Deep learning for chest radiograph diagnosis. PLOS Medicine.
2018;15(11):e1002686.
4. Esteva A, Kuprel B, Novoa RA, et al. Dermatologist-level classification of skin cancer with deep neural
networks. Nature. 2017;542:115-118.
5. Norori N, Hu Q, Aellen FM, Faraci FD, Tzovara A. Addressing bias in big data and AI for health care:
A call for open science. Patterns. 2021;2(10):100347.
6. Goldberger
AL,
Amaral
LAN,
Glass
L,
et
al.
PhysioBank,
PhysioToolkit,
and
PhysioNet.
Circulation. 2000;101(23):e215-e220.
7. Shih G, Wu CC, Halabi SS, et al. Augmenting the national institutes of health chest radiograph
dataset with expert annotations of possible pneumonia. Radiology: Artificial Intelligence.
2019;1(1):e180041.
8. Hripcsak G, Duke JD, Shah NH, et al. Observational Health Data Sciences and Informatics (OHDSI):
Opportunities for observational researchers. Stud Health Technol Inform. 2015;216:574-578.
9. Nagendran M, Chen Y, Lovejoy CA, et al. Artificial intelligence versus clinicians: systematic review of
design, reporting standards, and claims of deep learning studies. BMJ. 2020;368:m689.
10.NITI Aayog. Responsible AI for All: Adopting the Framework — A Use Case Approach for All. NITI
Aayog: New Delhi, 2024.
11.Obermeyer Z, Powers B, Vogeli C, Mullainathan S. Dissecting racial bias in an algorithm used to
manage the health of populations. Science. 2019;366(6464):447-453.
12.Menon GR, Singh L, Sharma P, et al. National Burden Estimates of Healthy Life Lost in India, 2017.
Indian Journal of Medical Research. 2019;150(2):116-128.
13.Obermeyer Z, Emanuel EJ. Predicting the future — big data, machine learning, and clinical medicine.
New England Journal of Medicine. 2016;375:1216-1219.
14.Wiens J, Saria S, Sendak M, et al. Do no harm: a roadmap for responsible machine learning for
health care. Nature Medicine. 2019;25:1337-1340.
15.DeLong ER, DeLong DM, Clarke-Pearson DL. Comparing the areas under two or more correlated
receiver operating characteristic curves: a nonparametric approach. Biometrics. 1988;44(3):837-845.
16.Chouldechova A. Fair prediction with disparate impact: A study of bias in recidivism prediction
instruments. Big Data. 2017;5(2):153-163.
17.Mitchell M, Wu S, Zaldivar A, et al. Model Cards for Model Reporting. Proceedings of FAT 2019.
ACM: New York, 2019.
18.Gebru T, Morgenstern J, Vecchione B, et al. Datasheets for datasets. Communications of the ACM.
2021;64(12):86-92.
19.Lundberg SM, Lee SI. A unified approach to interpreting model predictions. Advances in Neural
Information Processing Systems. 2017;30.
20.Ribeiro MT, Singh S, Guestrin C. 'Why should I trust you?' Explaining the predictions of any classifier.
Proceedings of KDD 2016. ACM: New York, 2016.
21.Chattopadhay A, Sarkar A, Howlader P, Balasubramanian VN. Grad-CAM++: Generalized
gradient-based visual explanations for deep convolutional networks. WACV 2018. IEEE, 2018.
22.Kim B, Wattenberg M, Gilmer J, et al. Interpretability beyond classification: Quantitative testing with
concept activation vectors (TCAV). Proceedings of ICML 2018. PMLR, 2018.