Healthcare and life sciences
How hospitals, life science companies and medical device makers apply NVIDIA AI to genomics, drug discovery, device software, clinical documentation agents and research compute, with patient privacy, clinical validation and device regulation planned in from the start.
The problem
Hospitals hold years of scans, pathology slides, genome files and clinician notes, but most of it sits in departmental systems with different formats and access rules. Turning that archive into a model means settling ethics approval, consent terms and data use agreements long before any GPU is involved.
Clinical staff carry a heavy documentation and phone load that takes time away from patients, and research groups wait on slow secondary analysis of each sequencing run. Device makers face a different limit: software inside a surgical robot or an ultrasound system must respond within fixed time budgets and must pass medical device review before it reaches an operating room.
Drug discovery teams search very large chemical and protein spaces where every wet-lab experiment is costly, so they want computation to shorten the candidate list before anything is synthesized.
The approach
Each part of the sector draws on a different part of NVIDIA's stack. Genomics labs use Parabricks to run accelerated versions of common open source secondary analysis steps (alignment with BWA-MEM, GATK processing and DeepVariant calling); NVIDIA ships it as a free public container and sells enterprise support through NVIDIA AI Enterprise. Drug discovery teams use BioNeMo models, recipes and NIM microservices to design molecules, screen compound libraries in silico and predict protein structures. Imaging researchers often build on MONAI, an open community framework begun by King's College London together with NVIDIA and now developed by more than 30 institutions.
Medical device makers use the Holoscan SDK to process sensor and video streams with low latency on hardware such as NVIDIA IGX, and robotics teams use Isaac for Healthcare to simulate anatomy, sensors and hospital settings. For clinical language work, NVIDIA positions Nemotron open models for ambient healthcare agents and clinical research agents, and publishes an ambient agents example that drafts structured visit notes from recorded conversations. Health systems with very large image archives may train their own foundation models on dedicated clusters, as Mayo Clinic does on a DGX SuperPOD.
Most providers do not need to build any of this. A cleared commercial product, a hosted genomics service or rented cloud GPUs is usually simpler, and scheduling, billing and many analytics tasks run well on CPUs. Using NVIDIA software does not by itself satisfy medical device rules: the manufacturer still has to validate the product and obtain whatever clearance the regulator requires.123456
Conceptual architecture
Diagram as a list
Applications & solutions
- Secondary genomic analysis (Parabricks)Aligns reads and calls variants for research and laboratory pipelinesConnects to Clinician sign-off and quality management system
- Ambient documentation and research agents (Nemotron)Draft notes, handle patient intake and summarize literature for staff to checkConnects to Clinician sign-off and quality management system
Models & frameworks
- Foundation model and drug discovery training (BioNeMo, MONAI)Builds pathology, imaging and molecular models on approved datasetsConnects to Model serving inside the organization's network (NIM)
Inference & runtime software
- Model serving inside the organization's network (NIM)Exposes trained or open models through standard APIs without sending data to outside servicesConnects to Ambient documentation and research agents (Nemotron)
- Device and surgical robot software (Holoscan on IGX)Processes camera and sensor streams in real time inside a regulated deviceConnects to Clinician sign-off and quality management system
Operations & orchestration
- Hospital data intake with consent and de-identification controlsPulls images, notes, sequences and signals from PACS, EHR, lab systems and sequencers under approved data use termsConnects to Secondary genomic analysis (Parabricks), Foundation model and drug discovery training (BioNeMo, MONAI)
- Clinician sign-off and quality management systemKeeps a qualified person accountable and records validation and post-market evidence
Accelerated computing
- Hospital GPU cluster, DGX SuperPOD or cloud GPUsSupplies capacity for training, genomics batches and servingConnects to Secondary genomic analysis (Parabricks), Foundation model and drug discovery training (BioNeMo, MONAI), Model serving inside the organization's network (NIM)
Technologies and their roles
NVIDIA Holoscan SDK78
Medical device software runtime
Runs low-latency sensor and video pipelines; Moon Surgical says its FDA-cleared ScoPilot feature runs locally on the Maestro robot with Holoscan.
NVIDIA BioNeMo1
Biology and drug discovery models
NVIDIA names protein binder design, virtual screening and biofoundation model building among BioNeMo uses.
NVIDIA Nemotron1
Clinical language and voice agents
NVIDIA names ambient healthcare agents and clinical research agents as Nemotron use cases.
NVIDIA Isaac1
Healthcare robotics simulation
Isaac for Healthcare offers synthetic data and sensor simulation for endoscopy, ultrasound and CT, aimed at hospital automation and surgical robotics.
NVIDIA DGX6
Training infrastructure for large clinical archives
Mayo Clinic deployed a DGX SuperPOD with DGX B200 systems, first aimed at foundation models in pathomics and drug research.
NVIDIA NIM1
Self-hosted model serving
NVIDIA includes NIM microservices in BioNeMo, so models can be served behind an API inside the organization's own environment.
What you need first
- Ethics or institutional review board approval and a data use agreement for every training dataset
- A de-identification process that covers DICOM headers, free-text notes and genomic files
- Clinical experts with protected time to label data and review model output
- For device software, a quality management system and a regulatory plan for each market (for example FDA premarket review in the United States or CE marking in the European Union) agreed before develop
- Integration skills for hospital standards such as DICOM, HL7 and FHIR
- Bioinformatics or computational chemistry staff for genomics and drug discovery projects
- A security review that settles where models run and whether any patient data may leave the hospital network
Risks and how to reduce them
- Exposure of protected health information during training or inference
- De-identify before data leaves clinical systems, keep inference on infrastructure the organization controls where possible, and follow HIPAA, GDPR or the local equivalent.
- Software that influences diagnosis or therapy being handled as a research tool8
- Define the intended use early; if it is a device function, plan validation and a regulatory submission, and describe any clearance exactly as the regulator lists it (for example a 510(k) decision of substantial equivalence).
- Models that perform worse for some patient groups, sites or scanners
- Validate on data from several sites, devices and demographic groups, and keep monitoring after go-live.
- Staff over-relying on AI-drafted notes or summaries
- Require clinician review and sign-off of every drafted note and keep a record of edits.
- Reference examples built on components whose status changes5
- Check the lifecycle of each listed component before reuse; the ambient healthcare agents example, for instance, lists NeMo Microservices among its parts.
Documented examples
Moon Surgical · Medical devices (surgical robotics)
Moon Surgical: FDA-cleared ScoPilot camera control running on NVIDIA Holoscan
Moon Surgical's ScoPilot feature lets the laparoscope on its Maestro surgical robot follow a chosen instrument tip, running locally on NVIDIA Holoscan. It received FDA clearance in March 2025; Moon Surgical says Maestro had been used in more than 1,100 patients.
In production
Mayo Clinic · Healthcare and medical research
Mayo Clinic: a DGX SuperPOD with DGX B200 for pathology foundation models
Mayo Clinic deployed an NVIDIA DGX SuperPOD with DGX B200 systems in July 2025 and says its first work will be building foundation models for pathology, drug discovery and precision medicine. Mayo states the system cuts four weeks of slide analysis and model work to one; no measurement is published.
In production
Related
Sources
- NVIDIA healthcare and life sciences industry page (opens in a new tab)
- About Parabricks (Parabricks v4.7.1 documentation) (opens in a new tab)
- Project MONAI: About (opens in a new tab)
- NVIDIA Holoscan SDK (developer page) (opens in a new tab)
- Ambient Healthcare Agents developer example (GitHub README) (opens in a new tab)
- Mayo Clinic deploys NVIDIA Blackwell infrastructure to drive generative AI solutions in medicine (Mayo Clinic release via Newswise) (opens in a new tab)
- Moon Surgical receives FDA clearance for ScoPilot on Maestro, powered by NVIDIA Holoscan (company release, linked from moonsurgical.com) (opens in a new tab)
- FDA 510(k) Premarket Notification database: K242323 (opens in a new tab)
Thank you. Your correction was sent.
The editors check it against the sources. If you left an email address, they may reply about it.