Studying patient similarity and cognitive change over time, with Alzheimer's disease as the focus. SOMA combines longitudinal research inputs with federated matching and privacy-preserving methods.
SOMA is an Alzheimer's research initiative by Curadai. The work uses retrospective research cohorts to investigate patient representations, cognitive trajectories, and privacy. It is a research foundation, not a clinical diagnostic service. The work is conducted independently of the institutions that contributed the research data.
Alzheimer's researchers need to compare cognitive change across participants and time. Fragmented records, privacy requirements, and incomplete measurements make it difficult to identify participants with similar trajectories.
Patient records siloed across institutions with incompatible EHR systems and clinical vocabularies.
Research permissions and privacy requirements shape how sensitive participant data can be accessed and shared.
Existing approaches treat patients as frozen in time, losing critical trajectory and progression information.
Alzheimer's research visits may have cognitive scores without every scan, genetic measure, or biomarker.
SOMA investigates patient representations, cognitive change, and privacy using retrospective Alzheimer's research data. Current experiments focus on progression among participants with mild cognitive impairment, using ADNI as the primary cohort and NACC for an external research comparison. Clinical usefulness and the value of the resulting matches require further independent evaluation.
SOMA learns patient representations from cognitive assessments, APOE genetics, selected PET measures, and CSF and plasma biomarkers. Current ADNI experiments evaluate how these representations relate to progression from mild cognitive impairment. MRI extraction is available for further study but is not part of the current 33-feature reference configuration.
SOMA evaluates differential privacy in representation training and protected population summaries. Experiments examine the effect of privacy noise on utility and the feasibility of combining results from ADNI and NACC. The architecture is designed for locally held source records; privacy across the complete workflow and deployment at independent sites require further validation.
SOMA examines whether repeated cognitive measurements improve patient representations and the identification of comparable research participants. Existing retrospective experiments give mixed results. Further evaluation must use a defined observation window and keep future measurements and evaluation outcomes out of model fitting and matching references.
SOMA jointly encodes cognitive, genetic, PET and biomarker features and evaluates their contribution by retraining with selected input groups removed. Their contribution varies by task and privacy setting. Comparisons between alternative fusion architectures and their value for research matching remain to be established.
Controlled ADNI experiments examine what happens when selected measurements are withheld, including a feature mask corresponding to inputs unavailable in the NACC comparison. Results depend on whether measurements are missing from the query participant, the reference group, or both. Broader evaluation of naturally incomplete records is still needed.
ADNI provides the primary research setting. A frozen ADNI-trained representation has been evaluated on NACC participants with mild cognitive impairment. Differences in cognitive instruments, available biomarkers and observation timing limit how this transfer result should be interpreted. Independent prospective evaluation remains a separate step.
SOMA's ongoing questions include how updates from another site might change previously learned patient relationships. Existing work includes simulated drift monitoring and site-update infrastructure. Retention of earlier relationships after model updates remains to be evaluated directly.
Longitudinal cognitive change, incomplete visits, and differences between research cohorts shape how we evaluate patient similarity.
Patient matching depends on how well the representation captures Alzheimer's-related clinical structure. Privacy methods, temporal matching, and input fusion must be assessed against that foundation.
A single assessment cannot describe an entire trajectory. Repeated cognitive measurements and visit timing help frame the question of which research participants are progressing in similar ways.
Cognitive tests, imaging, genetics, and biomarkers should not be treated as interchangeable signals. Their contribution to Alzheimer's patient similarity needs to be measured for the specific research task.
Incomplete visits are part of longitudinal Alzheimer's research. Evaluate matching with the inputs actually available, and distinguish measured observations from any estimated values.
Results depend on the cohort, available measurements, and evaluation setup. A promising retrospective Alzheimer's result is a reason for further study, not a substitute for independent or prospective evaluation.
The Alzheimer's research pipeline maps cognitive assessments, imaging, genetics, and biomarkers into representations for similarity analysis. Source records and the representations used for matching have separate privacy requirements.
In the Alzheimer's research configuration, cognitive scores, MRI, PET, genetic inputs, and cerebrospinal fluid biomarkers are encoded into a shared 64-dimensional representation. Similarity and temporal information support research matching; privacy controls govern what a participating site may share.
In the research architecture, site nodes prepare representations locally and the Soma Core compares Alzheimer's trajectories. The diagram and simulation illustrate the design; they do not show a live hospital network.
Research-site nodes prepare Alzheimer's patient representations from locally held assessments, imaging, genetics, and biomarkers. The architecture separates source records from the representations used for matching.
The core studies similarity and cognitive trajectories within Alzheimer's research. It brings together representations and temporal information to identify comparable research participants.
The onboarding protocol is designed to calibrate representations from another research site before matching. Access, compatibility, and privacy controls remain part of the evaluation.
Alzheimer's experiments examine protected population summaries, privacy budgets, and their effect on similarity. Representations require privacy safeguards even when source records remain local.
ADNI provides the primary longitudinal research setting. NACC provides an external research comparison with partial feature coverage. These cohorts support research evaluation, not a claim of a deployed clinical network.
Alzheimer's Disease Neuroimaging Initiative data supports longitudinal research using cognitive assessments, APOE genetics, cerebrospinal fluid and plasma biomarkers, and PET imaging.
NACC provides an external Alzheimer's research comparison with partial feature coverage and different cognitive instruments. Transfer evaluation uses a frozen ADNI-trained representation; differences in available biomarkers and observation timing limit interpretation.
The next steps stay focused on Alzheimer's disease: stronger longitudinal evaluation, independent research partnerships, and a clearer understanding of performance over time.
Expand the assessment of cognitive trajectories, participant subgroups, and incomplete visits within the Alzheimer’s research scope.
Explore independent evaluation with Alzheimer’s researchers, including site-specific data permissions, cohort compatibility, and prospective study design.
Study how changes in visit schedules, measurements, and research populations affect the stability of Alzheimer’s patient matching over time.
SOMA is the Alzheimer's research initiative of Curadai — a company building AI infrastructure for non-profit and healthcare institutions. Contact us to discuss Alzheimer's research collaboration, longitudinal data, or evaluation of federated patient similarity.