Medical AI Use Case¶
The Problem: Medicine in Space¶
Astronauts operating in space face a medical environment fundamentally different from anything on Earth. Four factors converge to create a scenario where an on-board AI medical assistant is not merely convenient but potentially life-saving.
Communication Delay¶
As humanity extends its reach beyond low Earth orbit to the Gateway lunar station and future Mars missions, round-trip communication times with ground-based medical personnel increase significantly:
| Destination | One-Way Signal Delay | Round-Trip (Question + Answer) |
|---|---|---|
| Gateway (lunar orbit) | ~1.3 seconds | ~3 seconds |
| Mars (closest approach) | ~4 minutes | ~8 minutes |
| Mars (farthest) | ~24 minutes | ~48 minutes |
Even on the Gateway, where communication delay is near-real-time (~1.3 seconds), an on-board medical AI provides immediate decision support without depending on ground contact availability. Ground-based flight surgeons may not be on console 24/7, communication windows may be limited, and in an emergency, every second matters. Having an AI assistant that provides instant guidance eliminates the need to wait for a ground link.
For future Mars missions, the value becomes even more critical: a single exchange with a physician on Earth takes up to 48 minutes. In a medical emergency -- a cardiac event, a severe allergic reaction, a traumatic injury -- that delay can be the difference between life and death. An on-board medical AI provides immediate guidance with zero communication latency.
Limited Crew Medical Training¶
Crew members receive approximately 40 hours of structured medical training before launch, including basic life support, wound care, and use of the on-board medical kit. However, they are not physicians. Complex diagnoses -- distinguishing between appendicitis and a kidney stone, identifying a pulmonary embolism, managing anaphylaxis with limited drug supplies -- require expert medical knowledge that exceeds the crew's training level.
A medical AI trained on clinical literature can serve as a decision-support tool, helping crew members work through differential diagnoses, identify appropriate treatments from available supplies, and follow established clinical protocols.
No Internet Beyond Earth Orbit¶
Any system that relies on cloud-based AI services (ChatGPT, Google's medical AI, etc.) is fundamentally unusable on a deep-space mission. Even on the Gateway, bandwidth is limited and connection reliability cannot be guaranteed. On a Mars mission, maintaining a persistent internet connection is not feasible. A medical AI system must run entirely on hardware carried aboard the spacecraft.
Hardware Failure in Space¶
Space is an unforgiving environment for electronics. Cosmic radiation can cause single-event upsets (bit flips) in memory and processors. Thermal cycling between sun-facing and shadow-facing orientations stresses solder joints and connections. Launch vibration can loosen connectors and fatigue materials. A single server running a medical AI is a single point of failure -- and in a space environment, hardware failures are not a matter of "if" but "when."
Astra addresses this with triple redundancy: three independent clusters, each capable of running the full medical AI stack, with automatic failover that preserves all conversation history and medical context.
Medical AI Models¶
Astra runs two specialized medical AI models alongside its general-purpose models. These models are specifically trained on biomedical and clinical literature, giving them domain expertise that general-purpose models lack.
Meditron¶
| Property | Value |
|---|---|
| Developer | EPFL (Switzerland) and Yale University |
| Model size | 3.8 GB |
| Parameter count | 7 billion |
| Training data | Medical literature, clinical guidelines, PubMed abstracts |
| Capability | Medical text Q&A |
| Runs on | All three AI servers |
Meditron was developed by researchers at the Ecole Polytechnique Federale de Lausanne (EPFL) and Yale University. It was trained on a carefully curated corpus of medical text including:
- Clinical practice guidelines from major medical associations
- Medical textbooks and reference materials
- Peer-reviewed research articles from PubMed
- Medical licensing examination materials
Meditron excels at text-based medical questions: symptom analysis, treatment protocols, drug interactions, and clinical decision support. It is designed for use by non-specialists (such as astronauts with basic medical training) who need expert-level guidance.
At 3.8 GB, Meditron runs comfortably on all three AI servers, including the memory-constrained Jetson Orin Nano. On the Jetson GPU, it produces responses at approximately 25-30 tokens per second.
MedGemma 1.5¶
| Property | Value |
|---|---|
| Developer | |
| Model size | 3.3 GB (quantized to Q4_K_M) |
| Parameter count | 4 billion |
| Training data | Medical imaging datasets, clinical text, biomedical literature |
| Capability | Medical imaging analysis and medical text Q&A |
| Best on | Jetson Orin Nano (CUDA GPU inference) |
MedGemma 1.5 is Google's medical AI model designed for both medical imaging analysis and text-based medical reasoning. Key capabilities include:
- Medical image analysis: Can process medical images (X-rays, dermatological photographs, fundoscopy images) and provide preliminary assessments
- Clinical text reasoning: Answers medical questions with reasoning grounded in biomedical literature
- Multimodal integration: Can analyze an image and text prompt together, such as interpreting an image alongside the patient's symptoms
The model is quantized to Q4_K_M format, which reduces the model size from its full-precision form while maintaining diagnostic accuracy. This quantization is essential for running on the Jetson's 8 GB unified memory.
Why MedGemma runs best on the Jetson
Medical image analysis is computationally intensive. The Jetson's CUDA GPU accelerates both the image processing and text generation components of MedGemma, producing results 5-10 times faster than CPU-only inference on the Pi 5 units.
Example Medical Prompts for Demos¶
The following prompts demonstrate the medical AI capabilities during live presentations. They are designed to showcase clinically relevant responses that a non-specialist crew member might need.
Symptom Assessment¶
"A crew member reports sudden onset of sharp pain in the lower right abdomen, mild fever of 38.2C, and nausea that started 6 hours ago. What are the most likely diagnoses, and what examination steps should I perform with the medical kit available on the station?"
Emergency Response¶
"A crew member is experiencing difficulty breathing, swelling of the face and lips, and a rapidly spreading rash after eating a meal containing tree nuts. They have a known tree nut allergy. Walk me through the emergency treatment protocol step by step."
Medication Guidance¶
"We have the following medications in the station medical kit: acetaminophen, ibuprofen, diphenhydramine, loperamide, and amoxicillin. A crew member has a suspected urinary tract infection with symptoms of painful urination and lower abdominal discomfort. What is the appropriate treatment with available medications?"
Wound Management¶
"A crew member sustained a 4cm laceration on the forearm from a sharp metal edge during EVA suit maintenance. The wound is clean but deep enough to see subcutaneous tissue. I have suture kits, butterfly strips, and antiseptic solution available. Guide me through wound closure and aftercare."
Altitude/Environmental Medicine¶
"What are the symptoms of acute mountain sickness, and how should it be treated in a remote environment with limited medical supplies?"
Model selection for medical demos
For text-only medical questions, use meditron on any server. For questions involving medical images (if applicable), use medgemma-1.5-4b-it on the Jetson (ai3 prefix) for fastest processing.
Important Disclaimer¶
AI is a decision-support tool, not a replacement for medical professionals
The medical AI models in Astra are designed as decision-support tools to assist crew members with basic medical training. They are not a substitute for professional medical judgment.
Key limitations:
- AI models can produce incorrect or misleading medical information (hallucination)
- Models are trained on general medical literature and may not account for the unique physiological effects of microgravity
- No AI model has been FDA-approved or CE-marked for clinical diagnosis
- AI responses should always be cross-referenced with established medical protocols and, when possible, verified with ground-based medical personnel
In a real space station deployment, the medical AI would be one component of a comprehensive medical support system that includes:
- Ground-based flight surgeon consultation (when communication allows)
- Printed medical reference materials
- Structured clinical protocols for common scenarios
- Crew medical officer training programs
The value of on-board medical AI is significant even when communication is near-real-time (as on the Gateway), because it provides immediate decision support without depending on ground contact availability. It becomes critical when communication delays make real-time physician consultation impossible, as on Mars missions. In those situations, AI-assisted decision support is better than no expert guidance at all.
Integration with OpenWebUI¶
Medical models are accessed through the same OpenWebUI interface as general-purpose models. Users select the desired model from the model dropdown menu in the chat interface:
- Models with the
ai1.,ai2., orai3.prefix route to specific AI servers - Meditron and MedGemma appear in the model list on all servers where they are installed
- Conversation history with medical models is stored in PostgreSQL and replicated across all three clusters, ensuring that medical context is preserved through failover events
This design means that if an astronaut is in the middle of a medical consultation when a hardware failure occurs, the failover preserves the entire conversation history. The crew member can continue the consultation on the surviving cluster without repeating any information.