Hi, @lilmisscupcake
I strongly agree with your input. Having a well-structured output schema is essential for defining a standardized ground truth.
Regarding the explainability of the decision-support system, we should also determine how to report the evidence and inputs that each system relied on when generating its output.
It would be great if you could explore possible solutions for each pathway, including the rule-based, Bayesian, LLM/RAG, and hybrid approaches.
Hi all @AIMLawg ,
We’ll see you in next 15 minutes ,
Link Zoom: Launch Meeting - Zoom
It was a great meeting. Thank you, everyone, for the insightful discussion. I’m looking forward to contributing to the project and collaborating with the team.
Hey Sir, sorry for the inconvenience as I was unable to speak out during the meeting today due to some technical issues but I messaged at the chat regarding joining the subgroup project. I want to join your project and learn from you.
Please look into this.
Manat Raina
Date: 31 July 2026
This session continued the previous roadmap, with the primary focus on product development and software architecture.
The overall pathway will be:
Research → Validation → Development → Product
Publications will be a by-product of this development journey.
The main goal is not limited to SANS. The broader vision is an AI-based ophthalmic triage and predictive decision-support system for detection, risk prediction, and clinical recommendations. SANS will serve as one clinical scenario for testing and validating the framework.
The initial phase will focus on creating realistic simulated clinical scenarios.
Approach:
Literature review to identify:
Relevant clinical variables
Measurement methods
Normal ranges and abnormal thresholds
Develop 4–5 astronaut baseline profiles.
Generate time-point-based clinical scenarios representing disease progression or clinical events.
This step is feasible and provides a foundation for system development and validation.
A group of 4–5 ophthalmology experts will be involved to:
Define and verify clinical scenarios
Provide expert labels and ground truth
Validate system outputs and compare performance against clinical judgment
Suggested workflow:
2 experts: scenario labeling and ground truth generation
2–3 experts: validation and performance assessment
Clinical experts will work closely with programmers and data scientists to ensure simulations are clinically meaningful.
Clinical Lead: Alireza + invited ophthalmologists
Data Science & Engineering Team Leads: Rob Reynolds, Elisabeth
Computational Team: Soujanya, Suryanash
The goal is to generate synthetic clinical datasets based on:
Published literature
Expert-defined variables
Physiological ranges
Clinical relationships
Workflow:
Clinical experts define:
Variables
Measurement approaches
Expected ranges
Computational team generates simulated populations and longitudinal clinical data.
Lead: Rob Reynolds, with support from Elisabeth, Soujanya, and Suryanash
The system will include multiple intelligence approaches for different objectives:
Rule-based CDSS
Bayesian CDSS using Directed Acyclic Graphs (DAGs)
LLM/RAG-based CDSS
Hybrid approaches combining:
Clinical rules
Bayesian inference
Evidence retrieval
Expert knowledge
Each approach will be evaluated using a predefined framework including:
Accuracy
Reliability
Interpretability
False alarm rate
Clinical usefulness
Computational efficiency
Speed and feasibility for edge-computing environments
Clinical and computational teams will collaboratively design the evaluation framework based on literature and clinical requirements.
Literature review for:
Clinical variables
Scenario definitions
Evaluation metrics
AI methodology selection
Establish ophthalmology and clinical expert team.
Create formal GitHub repository under AWG/OSDR for:
Dataset development
Scenario generation
Documentation
Begin clinical scenario and synthetic dataset creation.
Hello! Is it still able to join this subgroup?
Hi ,
Yes of course ,
we are at beginning of building something very cool,
send your email in messages for me, I’ll put you in email list 
Hi @AliReza-H,
I am new to this subgroup. I am willing to join this project and contribute. My email ID: swathigadevic@vt.edu
Thank you. Looking forward!
Hi @Swathiga ,
That’s great , i will add you to email list
See you this Friday !
Hi everyone,
Thank you all for the great discussion and contributions.
The main goal of the CDSS is to detect problems, suggest next steps, and support clinical decision-making, ultimately contributing to greater mission independence and medical autonomy during long-duration spaceflight.
We also suggested Systems Medicine for Human Spaceflight as an important reference for the group.
A major limitation of many current spaceflight datasets, including several OSDR datasets, is that they are mainly focused on multi-omics and research questions rather than practical longitudinal clinical scenarios. Therefore, we need to build a dedicated synthetic astronaut clinical dataset for development and evaluation of the CDSS.
The first major task will be to create realistic longitudinal astronaut profiles.
Each astronaut will have baseline characteristics such as:
Age and sex
Medical history
Anatomy and physiology
Ophthalmic measurements
OCT variables such as RNFL and choroidal thickness
Refraction
Other relevant clinical and mission-related variables
Each simulated astronaut will then be followed across multiple mission timepoints, approximately:
Preflight → Launch → In-flight → Post-flight
For example:
L-30 → L → L+30 → L+90 → L+120
The exact number and timing of measurements should be extracted from the literature. A possible structure may include approximately three preflight, three in-flight, and three post-flight assessments.
For every astronaut, different longitudinal clinical scenarios can then be generated, including:
Normal/stable state
Progressive changes
Persistent/steady abnormalities
Missing modality
Poor-quality data
Incomplete information
Other clinically relevant scenarios
This will create many combinations of astronauts, timepoints, and clinical states for testing the CDSS.
The simulation should be transparent, reproducible, and evidence-based.
Literature review → Structured spreadsheet → LLM-assisted extraction → Paper-by-paper verification → Simulation → Computational validation → Clinical expert validation
From the literature, we will extract variables such as:
Clinical and physiological variables
Ophthalmic and OCT variables
Demographics
Health conditions
Measurement methods
Normal and abnormal ranges
Longitudinal changes
Relationships between variables
These will first be recorded in a shared spreadsheet.
LLMs such as ChatGPT or Claude can help extract and organize the information, but every extracted value should remain linked to its source and then be rechecked against the original paper. This is important to prevent unsupported or hallucinated values and make the AI-assisted process scientifically reproducible.
After defining variables and ranges, we can compare different methods of patient simulation rather than relying only on simple random generation.
Clinical experts, particularly ophthalmologists for the SANS scenarios, will then assess whether the generated longitudinal cases are realistic and clinically meaningful.
We should remain open to using real and analogue datasets.
If we obtain a well-structured clinical dataset, even if incomplete, an important question will be how to combine it with our simulation framework.
Examples include bed-rest studies and other terrestrial spaceflight analogues.
The goal is not to replace real data with simulation, but to use simulation to extend existing datasets, generate missing clinical trajectories, and create rare or difficult-to-observe scenarios.
The initial scientific question is:
How safely and accurately can a CDSS detect clinical changes and suggest the appropriate next steps during spaceflight?
SANS can be the first focused use case.
Different CDSS approaches can then be compared:
Rule-based systems
LLM-based systems
Hybrid approaches
Other probabilistic or causal methods as the project develops
The same simulated scenarios should be presented to each system, and their outputs compared against expert-defined reference standards using a structured scoring framework.
The longer-term pathway is:
Astronaut simulation → CDSS validation → Software development → Deployment-oriented architecture
The software should eventually consider real mission constraints such as:
Edge computing
Fast response
Limited connectivity
Computing resources
Appropriate programming architecture
Mission independence
The first broader product can become an Eye Triage CDSS, with SANS as one important scenario. The framework could later expand toward other human-system risks such as thrombosis, embolism, renal stones, radiation-related problems, and other classes defined within NASA’s Human System Risk framework.
A shared spreadsheet should be available to interested members so that literature extraction and dataset development can happen collaboratively.
Supervision / Advising: @ElisabethAslinger, @RobertReynolds
Volunteer: Soujanya Pittala
Collect all relevant papers in one shared folder.
Prepare the structured variable-extraction spreadsheet.
Begin evidence-based extraction of variables, ranges, timepoints, and clinical relationships.
Explore computational methods for validating synthetic patient/astronaut data.
Establish an official CDSS repository within the AWG GitHub for code, data structures, documentation, and version tracking.
Thank you everyone for the discussion and contributions during the meeting.
The main focus of this session was continuing the Astronaut Profile Simulation work for the SANS-focused CDSS, particularly the literature-to-variable workflow, defining the minimum clinical dataset, and planning how synthetic astronaut profiles should eventually be generated and validated.
Soujanya presented progress on the literature review and variable extraction.
She reviewed approximately 35 papers, with useful information extracted from 27 relevant papers. The literature covered several domains relevant to astronaut ocular and systemic health, including:
Retina
Optic nerve
Visual function
Ocular biometry
Intracranial and neurological variables
Cerebral variables
Cardiovascular variables
Astronaut demographics and cohort characteristics
Measurement timepoints
One important limitation was that many publications report information at the cohort level rather than individual astronaut level, making reconstruction of complete individual astronaut profiles difficult.
The current spreadsheet therefore represents an important first step but still needs further cleaning, harmonization, and expansion.
Soujanya will share the literature resources, search strategy, and variable-extraction spreadsheet with the group.
A major discussion was that before generating synthetic astronauts, we first need to clearly define:
What information should a clinically meaningful astronaut profile contain?
Visual acuity or a few isolated ophthalmic measurements are not sufficient.
The group agreed that we should first develop a minimum required SANS dataset, containing the variables necessary for meaningful clinical assessment.
This should include the important ophthalmological findings and relevant systemic or physiological variables that influence SANS.
Matthew volunteered to prepare a comprehensive table containing:
Variable → What it represents → Why it matters for SANS → Plain-language explanation
This table can become a reference for both the clinical and engineering teams.
An important point raised during the discussion was that synthetic data generation cannot be separated from the intended use of the CDSS.
We therefore need to clearly define what the system is expected to answer.
The initial objective should probably not be simply:
“Does this astronaut have SANS?”
A more useful CDSS could answer questions such as:
Is this astronaut showing a concerning change?
What is the current risk level?
Is the finding stable or progressing?
Is immediate action required?
Should another examination or modality be obtained?
Is the available information insufficient or unreliable?
Should the Crew Medical Officer be contacted?
How confident is the system?
What should be done next?
This supports our previous concept of building the CDSS primarily as an early-warning, triage, and decision-support system, rather than only a diagnostic classifier.
The group discussed creating synthetic longitudinal astronaut profiles after the variables and their clinically meaningful ranges have been established.
The intended workflow remains:
Literature → Variables → Measurement definitions → Ranges/relationships → Baseline astronaut profiles → Longitudinal clinical scenarios
Synthetic profiles could represent different levels and trajectories of SANS-related risk, including:
Normal profiles
Mild abnormalities
Progressive changes
Higher-risk profiles
Stable abnormalities
Missing measurements
Low-quality measurements
Incomplete modalities
This allows us to create controlled clinical scenarios even when complete real astronaut datasets are unavailable.
The group emphasized that the generated values must remain physiologically and clinically plausible, rather than simply being independent random numbers.
The group also discussed limitations in available spaceflight datasets.
Many datasets do not contain all the clinical modalities that would normally be useful for SANS assessment, particularly detailed ophthalmic imaging.
Possible alternatives or complementary measurements, such as astronaut-performed vision testing, may provide useful monitoring information but cannot necessarily replace imaging modalities such as OCT or detailed optic-disc assessment.
This means that the simulation framework should eventually represent both:
complete clinical profiles and realistic incomplete profiles where some modalities are unavailable.
That is particularly relevant to evaluating CDSS behavior under mission constraints.
Validation remains one of the major scientific challenges.
A model developed using terrestrial ophthalmic datasets may not automatically generalize to astronauts or the spaceflight environment.
The group discussed the difficulty of obtaining adequate astronaut ground truth, particularly for imaging-based AI systems.
Therefore, our validation strategy should not depend on a single method.
The planned framework should combine:
Literature-supported physiological constraints + real/analogue datasets where available + computational validation + expert clinical review
Expert-labelled synthetic scenarios can provide the reference standard for evaluating the CDSS when adequate real astronaut cases do not exist.
Soujanya: Share the literature review resources, search strategy, and current variable-extraction spreadsheet with the group.
Matthew: Create a comprehensive table of important SANS-related variables with clear explanations of each variable and its clinical relevance.
Continue cleaning and harmonizing the literature-to-variable spreadsheet.
Define the minimum required astronaut/SANS clinical dataset before starting large-scale simulation.
Define the initial CDSS use cases and structured outputs, especially risk detection, warning level, next-step recommendations, uncertainty, and escalation.
Determine appropriate ranges and relationships between variables from literature and expert knowledge.
After these steps, begin development and validation of synthetic longitudinal astronaut profiles.
Hi everyone,
As discussed, I’ll share the literature review resources, search strategy, and the current variable-extraction spreadsheet with the group.
Hi everyone, sharing the literature resources I used for the literature review. These include the papers reviewed and the relevant sources used for variable extraction.
Hello @AliReza-H ! Apologies for being MIA for the past couple of weeks. I am still interested in contributing to the CDSS project. Please let me know how I can help the team. As I am motivated to polish my coding skills, please assign me tasks on GitHub. I will appreciate it.
Hi, @alavia
That’s great!
Would you like to help Souja with the Literature to Variable task?
Of course, there will be many more tasks in the future, and you can wait until then for something more suitable if you’re not interested
Sure! I am interested. Please connect me with her so we can collaborate.
Hi everyone! 
Thank you, AliReza, for inviting me to the subgroup. The project looks very interesting, and I’m looking forward to learning more and contributing where I can.
Hi @AliReza-H,
Circling back on the explainability question you raised, across rule-based, Bayesian, LLM/RAG, and hybrid pathways. I’ve been digging into this and wanted to share progress.
Each approach reports evidence differently: rule-based systems are fully traceable via the rule that fired, Bayesian systems need a specific explanation structure since raw probability shifts aren’t intuitive on their own (there’s an existing taxonomy that splits this into explaining the model, the reasoning, and the evidence separately), and LLM/RAG systems explain via retrieved context plus generated text, though hallucination and source traceability are still real unsolved risks there, not something to assume is handled.
The most promising direction for us is on the hybrid side: rule-augmented RAG (symbolic rules baked into the prompt so outputs show both the activated rule and the retrieved evidence) and LLM-as-interface architectures (where the LLM only presents results, an underlying validated model makes the actual call using feature attribution). The second pattern specifically cuts hallucination risk by not letting the LLM drive the decision, which feels important for something mission-critical.
Want me to write this up properly as a subsection, possibly tying into Ayse’s AI Governance proposal? Let me know how you’d like this to plug in