Nodes
A clinical question enters intent analysis, then SQL generation when answerable. The generator performs basic syntax checks and repair; execution and result comparison happen in the evaluator.
← Back to timeline
Team project · Imperial College London
Clinical Multi-agent Text-to-SQL
Oct 2025 – Dec 2025Answered 17 of 20 questions correctly in the team’s project evaluation.
Role: Planner-agent design, LangGraph integration and failure-case evaluation.
Problem
Translate clinical questions into schema-grounded SQL over MIMIC-IV, with validation, repair and safe abstention.
System representation
Agent State Graph / Runtime Loop
Ask a question without having to write a database query
Hospital records are spread across tables. The team's assistant turns a question into SQL—the instructions used to retrieve records. My work focused on stepwise planning, LangGraph integration and failure-case testing.
Example and implementation details
QUESTION FROM THE PROJECT PRESENTATION
How many patients were admitted in 2018?
Find: admissionsTime: 2018Return: a count
1 · Interpret the request
The intent step identifies the kind of question. It can stop with no query when the request is not answerable from the database.
ANSWERABLE
2 · Generate SQL using the table definitions
Gemini 2.5 Pro receives the question, its intent and the available schema: the names and structure of the tables.
3 · Check, fix, and try again when needed
Basic syntax checks catch malformed queries. The generator can retry after an error; unresolved failures return no SQL.
↺ Repair happens inside the generator
NOT ANSWERABLE
Stop without inventing an answer
A request for treatment advice is different from a request to retrieve a recorded fact. The system can return null instead of SQL.
TEAM EVALUATION
Does the query retrieve the right result?
The evaluator executes generated and reference SQL in SQLite and compares the returned results. Passing syntax checks alone does not establish that the question was answered correctly.
Implementation boundary: agents, retries and validation
The supplied notebook has two LangGraph nodes: intent_analyzer and generator. The generator makes at most two attempts in total. Checks cover basic SELECT structure, parentheses and incomplete clauses; table-name checks provide warnings rather than blocking output. SQL execution is part of the separate diagnostic evaluator. These checks do not certify clinical correctness.
Technical walkthrough
Shared state
LangGraph coordinates shared state, nodes and transitions.
Branches and repair loop
Answerable intents route to the generator; unanswerable intents end the graph. Syntax repair and retries take place inside the generator before output is returned.
Safe abstention
The system abstains when it cannot produce a safe answer.
Core Ideas / Concept Primer
The ideas behind the implementation
SQL and Text-to-SQL
SQL expresses a query over relational tables. Text-to-SQL translates a natural-language question into that query; a syntactically valid query can still answer the wrong question.
Clinical questions need a plan
The planner decomposes a question into cohort, time window, admissions and aggregation. Making these choices explicit helps separate interpretation of the question from query construction over MIMIC-IV.
State, nodes and edges
State carries information through a run. The supplied implementation has intent-analysis and generation nodes, with a conditional route to generate or end. The repair loop is internal to generation.
Validation is a decision point
Basic syntax checks can trigger another generation attempt. Table-name checks are advisory, while execution is tested separately. Neither a parseable query nor a matching response format guarantees the correct answer.
Implementation details
Clinical question
Natural-language request
Planner agent
Intent analysis and stepwise decomposition
SQL generation
Schema-grounded query construction
Validate and repair
Syntax, schema and execution checks with repair
Execute or abstain
Execute valid queries; abstain when unresolved
Results and Demonstrations
Questions answered correctly in team evaluation
17 / 20 (85%)Evaluation details
- Sample size
- 20 questions
Ownership
Designed and implemented
Designed stepwise planner logic; integrated LangGraph state, nodes and transitions.
Evaluated
Performed failure-case testing within the team evaluation, which answered 17 of 20 questions correctly.
Team context
Imperial College London team project using MIMIC-IV.