We introduce EHR-TempSQL, a benchmark for temporal database state reasoning in medical text-to-SQL, comprising 2K dual-turn interactions constructed from three publicly available EHR databases, including MIMIC-III, MIMIC-IV, and eICU, where database states are separated by an average 1.42 hour interval. We construct the dataset through a multi-stage pipeline that transforms static queries into temporal progression scenarios where queries track accumulated information and detect updates between states.
The dataset contains the following fields:
- scenario_id: Unique identifier for each sample.
- turn: Indicates the turn number in the dual-turn interaction.
- T_init, T_curr: Timestamps of the database snapshots.
- question: Natural language question.
- query: Corresponding SQL query.
- sql_result: Result of executing the SQL query.
- answer: Final answer to the question.
- q_tag: The question template.
- value: Sampled values from the database.