CDSS Group Meeting Summary — 25 September 2026
Thank you everyone for another thoughtful discussion. This meeting helped us narrow the direction of the CDSS subgroup and connect several pieces of work that have been developing in parallel over the past few months.
Main direction: SANS as the first CDSS use case
We agreed that Spaceflight-Associated Neuro-ocular Syndrome (SANS) should remain the first focused clinical use case for the CDSS.
One important point raised by Ritika was that we already have a considerable amount of work spread across different places: literature reviews, datasets, prototypes, UI/UX ideas, engineering work, governance discussions, and earlier meeting outputs. Before adding more layers, it would be useful to bring these together into one reference document so everyone can see clearly what has already been done, what is still missing, and where each person can contribute.
Dr. Chintan Dave also emphasized something fundamental: before choosing a model, dataset, or technical architecture, we need to define the exact clinical and research question.
This led to an important distinction in the discussion. An AI model may predict or classify a condition, but a CDSS should do more than diagnosis. It should bring together clinical evidence, guidelines, sequential information, uncertainty, and practical next steps to help a medical officer decide what to do.
For SANS, this could mean supporting questions such as whether a finding is concerning, whether it is progressing, what additional test is needed, how urgently action is required, and when expert or Crew Medical Officer involvement is necessary.
The system should also be designed around the actual constraints of spaceflight: limited computing resources, intermittent connectivity, delayed access to specialists, and the need for at least partially autonomous clinical decision-making.
We also discussed reviewing existing Earth-based CDSS platforms and open-source clinical systems rather than rebuilding every component from the beginning.
SANS biomarkers and clinical interpretation
Farnaz presented the outline of her review on AI-based detection of ocular biomechanical alterations in SANS.
The discussion covered several potential imaging and physiological biomarkers, including OCT-derived measures such as optic disc edema and retinal nerve fiber layer changes, fundus and retinal vascular imaging, OCT angiography, orbital ultrasound, and MRI.
A useful clarification from Dr. Chintan was the need to keep three concepts separate:
Symptoms are what the astronaut reports.
Signs are objective findings identified through examination or imaging.
Digital biomarkers are measurable quantitative features that can potentially be used for detection, monitoring, or prediction.
This distinction is particularly important in SANS because objective ocular changes may appear before an astronaut reports noticeable symptoms.
The group also discussed reversibility. Some effects related to fluid redistribution may improve after return to Earth, while structural abnormalities such as choroidal folds may persist for longer periods. The review should therefore look not only at detection, but also at which findings are reversible, which may persist, and what their possible long-term significance is.
Patient-reported symptoms should still be included in the CDSS, but they need to be interpreted alongside objective measurements and individual variability. The group also discussed the possibility that symptoms may sometimes be underreported because astronauts could be concerned about the implications for future flight eligibility.
Tabular foundation models and digital-twin work
Jian introduced tabular foundation models, particularly TabPFN 3.5, and discussed how they might contribute to several parts of the project.
Potential uses include classification, prediction, missing-data imputation, causal analysis, and synthetic-data generation. One attractive feature is that these models may allow several different analytical tasks to be approached within a common framework rather than developing an entirely separate pipeline for every task.
The proposed first experiment is intentionally simple: test TabPFN on an OSDR tabular dataset, such as OSD-255, and understand its practical strengths and limitations before considering more complex clinical applications.
There was also discussion about how these models might eventually contribute to a digital-twin framework, particularly if local or edge inference becomes feasible with relatively small datasets and limited computing resources.
Atul raised an important limitation: a model trained largely on Earth-based data may not adequately represent biological responses under microgravity, radiation, and other space-specific conditions.
Jian therefore proposed exploring a complementary approach in which foundation models are combined with methods such as conditional variational autoencoders (cVAE) and explicit spaceflight-related conditioning variables. This could provide a way to incorporate space-specific context rather than assuming that terrestrial relationships transfer directly to space.
Where we are heading
The discussion is gradually bringing several parts of the subgroup into one common framework:
Clinical question → SANS variables and biomarkers → astronaut longitudinal data → synthetic/real data integration → AI and probabilistic models → CDSS logic → clinically meaningful recommendations
The important point is that we are not trying to build an AI model simply because a model is available. The goal remains to develop a system that can eventually support real clinical reasoning under the unusual constraints of spaceflight.
Next steps
Ritika will prepare a forum update asking the group for input on the research question, project direction, and development tasks, and will also coordinate with Alireza on application development and the addition of Adolfo Viguera to the engineering team.
The subgroup will begin consolidating the existing literature, datasets, prototypes, UI/UX work, and governance material into a common reference resource. We will also review existing terrestrial CDSS systems and open-source clinical guidelines to identify components that may be reusable.
Farnaz will continue refining the SANS biomarker review, particularly the distinction between signs, symptoms, and digital biomarkers, as well as the evidence regarding reversibility and longer-term structural effects.
Jian will share the digital-twin presentation, datasets, notebooks, and relevant Google Drive/GitHub resources, and will organize a practical coding session for TabPFN 3.5. Group members can then test it on a simple OSDR tabular dataset and bring their findings to the next meeting.
Max and the existing VAE contributors will continue developing the conditional VAE and related digital-twin work.