The EnaResearch AI research assistant suggests suitable tests for a research question, interprets approved analysis results and drafts report text. It is optional and off by default; it never computes and never sees participant rows.
It does not run unless the organization turns it on in system management and a provider key is set on the server.
Suggests suitable tests from the research question and the variable dataset; every suggestion is validated on the server.
Interprets results of approved analyses and drafts text for methods, findings and discussion.
Variable definitions, counts and test output are sent; participant rows, pseudonyms, names and birth dates are not.
Every number in a reply is checked against the statistics output; replies with unverifiable numbers are rejected.
Requests are written to the assistant history and the audit log, including provider, model and unverified numbers.
The assistant ships turned off. The organization enables it in the AI assistant setting under system management, and a provider key must also be set on the server. While it is off, the rest of EnaResearch, including statistics, works the same way.
The assistant is for research use only, not for clinical decisions. An organization should complete its own legal assessment before using it with real patient data.
The assistant does three things on the Analyses page. For analysis suggestions, it proposes suitable tests from the research question and the variable dataset; variables and eligibility counts come from the statistics module, every suggested plan is validated on the server, and invalid ones are shown with the reason. For interpretation, it explains the result, meaning and limitations of an approved analysis that has a result. For report drafts, it writes copyable text for methods, findings and discussion from selected approved analyses.
The assistant never computes. Every number, such as a mean, p-value, confidence interval or regression coefficient, comes from the statistics module. The assistant does not create, approve or run analyses: the researcher creates a draft analysis from a suggestion and approves it through the usual approval step.
Only aggregated context is sent to the provider: variable definitions, counts and test output. Counts below 5 are sent as “<5,” and other parts of the same total are coarsened so a hidden count cannot be derived. Minimum and maximum values, which could reveal a single person’s value, and suppressed cells are not sent.
Participant rows, pseudonyms, names and birth dates are never sent.
Every number in a reply must match a number in the statistics output. If a number does not match, the model is warned once; if it still does not match, the reply is rejected and not shown. A statement such as “p < 0.05” passes only if the actual p-value supports it.
Every request is written to the assistant history and the audit log, including the provider, model, token usage and any unverified numbers.
The assistant works through an interface that is not tied to a specific model. The first supported provider is Claude by Anthropic. The provider key is set on the organization’s own server, and the rules for the context sent are the same whichever provider is used.
Versioned variables, missing values with a reason, and statistics from analysis plans approved before results.
Versioned definitions, event windows, pseudonymized CSV, Excel, SPSS and R output, a manifest and audit.
Approved lifecycle, PI and co-investigators, change requests, metric selection and dashboards.
We'll show what EnaResearch does today, share the roadmap openly and look at your studies and data sources together.
A separate deployment for every organization · Turkish and English