Large Language Models for Structured Clinical Data Analysis: Dual-Agent Grounding and Validation

1 Introduction

• The HIE analysis table harmonizes two multicenter trials across 21 sites and contains 532 de-identified infant records with approximately 1,300 variables per record.The study domain combines treatment assignments, physiological measurements, imaging variables, and longitudinal outcomes.Source: §1 · p. 2
• Exact columns, joins, filters, and aggregation operators matter because prompt-serializing the clinical table does not scale as a general query mechanism.The answer must also preserve the cohort and numerical meaning when translating between natural language, SQL, and results.Source: §1 · p. 2
• CLEAR-Med pairs an Invocation Agent that generates and executes SQL with a separately invoked Clinical Validation Agent that checks the draft, query, and result.The system can accept the draft, request bounded repair, or abstain after checks.Source: §1 · p. 2
• The evaluation covers CLEAR-Med configuration and scalability and Invocation Agent accuracy and consistency across 25 query templates with repeated trials.The introduction describes an oracle-scored development benchmark with task-level breakdowns.Source: §1 · p. 2
• For a stated property, sound executable checks constrain emitted answers to satisfy it, while unresolved cases abstain.The guarantee does not extend to clinical judgments that have not been encoded as executable checks.Source: §1 · p. 2

2 Related Work

• Biomedical language models have been adapted for medical dialogue, consultation, multilingual question answering, and domain knowledge [10] [11] [12].Healthcare-agent research extends toward planning, tool use, and multi-agent collaboration.Source: §2 · p. 3
• Clinical text-to-SQL work includes UniQA for electronic health-record questions [14], MedT5SQL, and Criteria2Query 3.0 for clinical-trial eligibility queries.Other work examined context-aware pharmacovigilance SQL and generalization across medical text-to-SQL datasets [17] [18].Source: §2 · p. 3
• CLEAR-Med treats query generation as a first stage, retaining executed SQL, raw database results, and the natural-language draft for a separate validation decision.Its design combines explicit provenance, executable checks, independent review, bounded repair, and abstention for relational aggregates.Source: §2 · p. 3

3.1 Overview

• CLEAR-Med is a dual-agent framework for natural-language analysis of structured disease-specific clinical data.The overview points to separate descriptions of the database, agents, and complete control flow.Source: §3.1 · pp. 3–4

3.2 Formal Dual-Agent Guarantee

• The formalization maps each attempt’s SQL, executed result, draft, and checker outcomes to accept, repair, or abstain, with outputs r0, r1, or ⊥.It permits at most one repair; ⊥ denotes abstention.Source: §3.2 · p. 4
• Under sound checkers, faithful database execution, and fail-closed control flow, every unresolved path terminates in abstention after at most one repair.The guarantee applies to properties encoded by the checkers.Source: §3.2 · p. 4

3.5 Clinical Validation Agent

• CLEAR-Med combines deterministic SQL/result and clinical-consistency checks with independent accept, repair, or abstain decisions, allowing one repair before fail-closed abstention.The validator cannot author replacement numbers; a repaired query is executed and revalidated, and unresolved failures produce explicit abstention.Source: Figure 1 · p. 4, §3.5 · p. 6

3.3 Convert Excel to SQLite Database

• 532 participant rows and approximately 1,300 variables in the Excel table were converted to SQLite; adaptation may be needed for datasets with different variable definitions.The 500 × 1300 size used later is a nominal synthetic scalability configuration, not the analysis table’s observed size.Source: §3.3 · pp. 4–5

3.4 Invocation Agent

• The Invocation Agent uses schema-guided text-to-SQL rather than document-retrieval RAG, with the executed relational database serving as the source of numerical evidence.The approach requires no fine-tuning.Source: §3.4 · p. 6

3.6 Pseudocode for CLEAR-Med

• The pseudocode preserves provenance through invocation and validation, emits an accepted draft as the final output, and never emits an unaccepted draft.The workflow includes one repair and fail-closed abstention.Source: §3.6 · p. 6

3.7 Evaluation Design

• The evaluation comprised configuration screening, end-to-end scalability comparison, and a 25-question development benchmark repeated five times per condition; full-pipeline validation and abstention were reserved for prospective study.The benchmark compared Invocation Agent responses with an ungrounded ChatGPT baseline, yielding 125 responses per condition.Source: §3.7 · p. 7

3.8.1 Choosing the ChatGPT configuration

• Configuration screening prioritized successful SQL generation over speed when the two conflicted.ChatGPT Thinking and ChatGPT Instant were compared on time-based and operational diagnostics.Source: §3.8.1 · p. 8

3.8.2 Optimizing other hyperparameters

• Four hyperparameters—iteration limit, execution-time limit, agent type, and Top K—were varied individually, prioritizing successful SQL generation and execution, then fewer errors and lower execution cost.SQL length and complexity were descriptive diagnostics.Source: §3.8.2 · p. 8

3.9.1 Language-Model Comparators and Interface

• Five biomedical, medical-dialogue, and general-purpose models were compared end to end against CLEAR-Med, which used SQLite with schema-guided SQL tools while comparators processed serialized tables in context.All systems used the same six nominal table sizes, five-minute timeout, and success definition.Source: §3.9.1 · p. 8

3.9.2 Scalability Metric

• A scalability configuration counted as successful only if it completed within 300 seconds without GPU out-of-memory failure; reporting included completion percentage, largest successful cell count, and elapsed time.The measures covered six configurations.Source: §3.9.2 · p. 9

3.9.3 Accuracy and Consistency Metrics

• For nonzero deterministic oracles, accuracy metrics were signed relative error, mean absolute error over defined predictions, and the percentage of attempts within 5% of the oracle.Oracle-zero outputs were excluded from Err; missing and nonnumeric outputs were also excluded from Err but counted as failures for GT5%.Source: §3.9.3 · p. 9
• Repeated-run consistency was summarized by the within-query range and standard deviation of signed relative error, averaged across queries with at least two defined predictions.Source: §3.9.3 · p. 9

4.1.1 Choosing the ChatGPT configuration

• ChatGPT Thinking took 167.44 s versus 29.23 s for ChatGPT Instant, but produced usable SQL while Instant returned no valid SQL.Both made two database invocations; Thinking completed without errors, while Instant produced three errors and hit the iteration limit.Source: §4.1.1 · p. 9

4.1.2 Setting up other key hyperparameters

• The reported configuration selected openai-tools for lower execution cost and a 60 s maximum execution time because only that tested limit reached a database invocation.Errors remained constant at three per run across the tested time limits.Source: §4.1.2 · p. 10
• The reported configuration retained Top K = 10 and Max Iterations = 50; at Top K = 10, 50 iterations was fastest (134.19 s), used one database invocation, and retained high complexity.Source: §4.1.2 · pp. 9–10

4.2.1 Scalability

• CLEAR-Med completed six of six fixed scalability configurations, versus four of six or one of six for the comparison models.It processed the nominal 500 × 1300 table; the largest-table run used approximately 19.71 GB of 39.56 GB GPU memory and took 28 s.Source: Table 2 · p. 10, §4.2.1 · p. 10

4.3 CLEAR-Med Development-Benchmark Results

• 39 of 60 Invocation Agent responses were within 5% of the oracle on percentage-valued queries, versus 4 of 60 baseline responses.These 60 responses covered 12 percentage-valued queries; Table 3 reports accuracy and consistency across five repetitions.Source: Table 3 · p. 11, §4.3 · p. 11
• 66.4% of Invocation Agent responses were correct versus 12.0% for the ungrounded ChatGPT baseline, a paired improvement of 54.4 percentage points.The benchmark included 25 queries repeated five times; overall accuracy was reported with query-cluster bootstrap confidence intervals.Source: Table 4 · p. 11, §4.3 · p. 11

5.1 Principal Findings and Interpretation

• The 25-query development benchmark yielded 66.4% Invocation Agent accuracy against 12.0% for the ungrounded ChatGPT baseline, while CLEAR-Med completed all six nominal scalability configurations.The accuracy comparison reflects the full invocation configuration, including database access, schema guidance, and prompt design.Source: §5.1 · pp. 11–12

5.2 Comparison with Related Approaches

• CLEAR-Med extends clinical text-to-SQL into a selective evidence pipeline with distinct operations, explicit terminal decisions, and bounded repair after independent validation.Its architectural claim is not that a second model replaces a database oracle, but that an independently invoked validator can review the complete evidence record after deterministic checks.Source: §5.2 · p. 12

5.3 Applicability

• The demonstrated use case is retrospective research analysis of a harmonized neonatal HIE schema; other governed relational datasets require replacement of schema guidance, constraint cards, and oracles.Transfer calls for dataset-specific validation.Source: §5.3 · p. 12

5.4 Limitations

• The study is limited to one harmonized HIE schema, sequential configuration screening, and a 25-query development benchmark whose errors informed prompt and oracle corrections.The conditional guarantee covers only properties represented by sound executable checkers under faithful database execution; open-ended clinical reasoning, individualized prognosis, and schema transfer remain outside demonstrated scope.Source: §5.4 · p. 12

• CLEAR-Med combines SQL provenance, deterministic checks, independent validation, bounded repair, and explicit abstention in a dual-agent framework for structured clinical data.Source: §6 · p. 13
• The SQL interface scaled to the full nominal HIE table, and the Invocation Agent substantially outperformed an ungrounded ChatGPT baseline on oracle-scored development queries.Source: §6 · p. 13
• Held-out testing and full-pipeline evaluation of validation and abstention remain the next stage of this work.Source: §6 · p. 13
• Acceptance contingent on an inspectable evidence record defines a route to auditable numerical answers in governed relational data.Source: §6 · p. 13

• The intended use is research and educational analysis of structured neonatal HIE trial data through natural-language questions.Source: §A.1 · p. 13
• ChatGPT Thinking handles invocation and Claude provides independent validation; the system uses no fine-tuning and retains requested and returned identifiers and access timestamps per run.Source: §A.1 · p. 13
• The de-identified harmonized database combines the Optimizing Cooling and Late Hypothermia trials across 21 sites, with variables defined by established NRN protocols.Access is restricted.Source: §A.1 · p. 13
• Reported measures include configuration diagnostics, scalability success, numerical accuracy, oracle error, repeated-run dispersion, and task- and difficulty-stratified accuracy.Source: §A.1 · p. 13
• Unresolved or malformed cases abstain after at most one repair, and accepted numbers retain SQL provenance and pass deterministic and independent validation.Source: §A.1 · p. 13

A.3 Illustrative SQL Examples

• SQLite examples use grouped SQL to summarize cognitive, language, and motor scores by HIE severity, an analysis described as cumbersome in Excel.Source: §A.3 · pp. 13, 16
• SQLite is described as able to manage millions of records, with an example query aggregating HIE clinical measures across time points where Excel may struggle.Source: §A.3 · p. 16

A.4 Formal Analysis of the Dual-Agent Selective Pipeline

• The appendix formalizes the bounded selective pipeline, defines endpoints for prospective full-pipeline evaluation, and limits its guarantee to sound executable checks.Source: §A.4 · p. 16

A.4.1 State and Transition Semantics

• The pipeline tracks the question, generated SQL, executed result, draft, and repair history; a separate cross-provider validator may accept, repair, or abstain.Source: §A.4.1 · pp. 16–17

A.4.2 Proof of the Encoded-Validity Guarantee

• Every emitted response satisfies the encoded properties, giving zero probability of an encoded-property violation for any admissible question.The guarantee depends on deterministic checks being sound and applies only to represented properties.Source: §A.4 · p. 16, §A.4.2 · p. 17

A.4.3 Selective Evaluation Quantities

• Selective evaluation defines coverage, incorrect-emission risk, accepted-output accuracy, and paired changes using resampling that preserves repeated trials from each question template.Source: §A.4.3 · pp. 17–18

A.4.4 Dependence-Aware Residual Risk

• Residual incorrect-acceptance risk is decomposed across no-repair and repair branches without assuming independence between agents.For errors caught by a sound checker, the deterministic-pass factor is zero; the bound does not extend to unencoded clinical judgments.Source: §A.4.4 · p. 18

A.4.5 Conditional Predictive Uncertainty

• Conditional predictive uncertainty separates model-induced variability from irreducible clinical outcome uncertainty; the control flow checks executable properties and routes unresolved cases to abstention.Source: §A.4.5 · p. 18