PaperlayerView exact revision on arXiv

Source-linked AI summary

CLAIRE: A Schema-Grounded Hybrid Workflow for Healthcare Administrative Form Completion

Garapati Keerthana, Manik Gupta

arXiv:2609.32787v1cs.AIcs.CLcs.HC

TL;DR

Healthcare administrative staff transfer information from varied sources into dynamic forms, where preserving dependencies, validation, provenance, and oversight is challenging. CLAIRE uses a schema-grounded hybrid workflow with bounded correction and escalation. Under the evaluated synthetic benchmark conditions, full CLAIRE achieved perfect completion metrics, while removing validation lowered stress-suite success.

  • Problem

    Dynamic healthcare forms require preserving field meaning, dependencies, validation constraints, provenance, and human oversight as information moves from heterogeneous sources.

    Source: §1 · p. 1

  • Method

    CLAIRE combines field discovery, source-supported value mapping, deterministic validation, bounded correction, escalation, and audit logging within a schema-grounded hybrid workflow.

    Source: §1 · p. 2, §3.1 · p. 5

  • Results

    Removing schema context reduced success to 0.200; removing validation reduced stress-suite success from 1.000 to 0.500 while leaving clean performance unchanged.

    Source: §5 · p. 9

  • Takeaways & Limitations

    The findings support healthcare administrative automation in which language-assisted mapping is embedded within structured controls and remains subject to validation and traceability.

    Source: §5 · p. 12, §8 · p. 13

    The synthetic evaluation does not capture real organizations’ full diversity, missingness, ambiguity, coding practices, or distribution shifts, and the primary benchmark does not establish autonomous LLM-agent competence.

    Source: §6 · p. 12

Abstract

from arXiv · show

Healthcare administrative staff transfer structured information from electronic health records, referrals, claims systems, provider rosters, and work queues into dynamic forms. We developed and evaluated CLAIRE (Clinical Language and Agentic Intelligence for Reasoning and Entry), a hybrid workflow that separates field-state discovery, source-to-field mapping, deterministic validation, bounded correction, escalation, and audit tracing. We tested five synthetic healthcare administrative schemas, 1,000 source records, four interface variants, two data-quality suites, and six comparators, yielding 24,000 benchmark episodes. A separate strict-output audit evaluated direct mappings from Qwen2.5-1.5B and Qwen2.5-7B, and a trace-derived operational simulation covered 6,000 episodes. Under the evaluated synthetic benchmark conditions, full CLAIRE achieved 1.000 episode success, field accuracy, required-field completion, and dependency completion in both suites; removing validation reduced stress-suite success to 0.500. In the simulation, 100.0% of clean and validation-stress episodes reached a staff-reviewable draft, compared with 68.6% of escalation challenge episodes, unsupported cases were blocked. Scenario-based savings were 149.7-165.5 seconds per case, not observed staff times. The findings support schema-grounded, validation-first healthcare administrative automation in which language-model components assist mapping but do not authorize unsupported or consequential actions.

Highlights

Healthcare administrative form completion must preserve field meaning, dependencies, validation, provenance, and human oversight as information moves from heterogeneous sources into structured interfaces. CLAIRE addresses this with schema-grounded mapping and validation, bounded correction, escalation, and audit tracing.

1. Introduction

  • Administrative form transfers span intake, benefit verification, prior authorization, credentialing, and claims correction, with missing or inconsistent entries able to delay care and create rework.The introduction frames the task as preserving field meaning and oversight across heterogeneous healthcare sources.Source: §1 · pp. 1–2
  • Dynamic forms require automation to reveal dependencies, map only supported values, validate completed state, and stop when evidence is missing.A dependency is a required field revealed by an earlier answer.Source: §1 · p. 2
  • CLAIRE combines field discovery, source-supported mapping, validation, bounded correction, staff review, and logging in a schema-grounded hybrid workflow.Deterministic schema and rule logic authorize benchmark actions; model-generated mappings were evaluated separately.Source: §1 · p. 2
  • The study reports a reproducible synthetic dataset, multiple form families and data-quality suites, interface variants, controlled comparators, a strict-output model audit, and a trace-derived simulation.The contribution list identifies these as separate evaluation components.Source: §1 · p. 2

2. Related work

Prior work studies browser agents, interactive form filling, and document or administrative AI, but leaves a gap in dependable healthcare-specific completion of dynamic forms. This paper positions CLAIRE as a modular workflow with computable schemas, deterministic validation, bounded correction, and traces.

2.1. Web agents and computer-use benchmarks

  • Browser-agent benchmarks cover reproducible web, enterprise, and desktop tasks, while reported reliability and safety weaknesses motivate evaluating validation and execution behavior.The paper reports validation and execution behavior rather than accuracy alone.Source: §2.1 · p. 4

2.2. Interactive form filling

  • FormFactory evaluates field–value alignment and graphical actions across layouts, whereas CLAIRE evaluates a healthcare-specific modular workflow with schemas, validation, correction, and traces.The adapter tests semantic mappings, not screenshot-coordinate trajectories, so the paper does not claim a full reproduction.Source: §2.2 · p. 4

2.3. Document AI and form prediction

  • Document-understanding and form-prediction methods support extraction or suggestion but do not themselves reveal conditional controls, validate dynamic form state, or block unsafe submission.The cited examples include LayoutLM, LayoutLMv3, and LAFF.Source: §2.3 · p. 4

2.4. Healthcare administrative AI

  • Healthcare administrative AI work supports individual capabilities, but dynamic form filling still requires source-grounded values, rule checks, bounded correction, and escalation when unresolved.The paper identifies this as a healthcare-specific gap.Source: §2.4 · p. 4

3. Methods

CLAIRE uses a shared schema and source record to map supported values, validate form state, correct only with supported evidence, and route unresolved cases to staff. Its evaluation combines synthetic benchmark conditions, a trace-derived operational simulation, and a strict-output audit of direct LLM mappings.

3.1. Architecture and workflow framing

  • CLAIRE’s Selector finds fields and dependencies, the Filler maps supported values, and the Validator checks rules; bounded correction is revalidated, while unresolved cases are blocked and logged.All roles use a shared source record and computable schema; model-generated content did not authorize benchmark actions.Source: §3.1 · pp. 5–6

3.2. Dataset and terminology

  • The dataset paired 500 clean records with 500 validation-stress records, including invalid identifiers, reversed dates, malformed claim identifiers, and negative amounts.Stress records retained explicit corrected values, allowing correction to use supplied evidence rather than guesswork; four interface variants changed presentation without changing meaning.Source: §3.2 · p. 6, §3.3 · p. 6

3.3. Internal benchmark design and comparators

  • The benchmark crossed 1,000 records, four interface variants, and six comparators for 24,000 episodes, measuring episode success alongside field, required-field, dependency, and validation outcomes.Variants included synonym labels, obfuscated identifiers, and shuffled layouts; episode success required all expected fields to match with no final validation error.Source: §3.3 · p. 6

3.4. Operational simulation

  • The operational simulation used 6,000 trace-derived episodes to test routine, validation-stress, and escalation routes under explicit per-action time assumptions.The escalation suite removed supplied corrections from selected stress cases; its timing assumptions were reproducible scenario inputs, not participant measurements.Source: §3.4 · pp. 6–7

3.5. Strict-output LLM audit and external alignment

  • The strict-output audit tested whether Qwen2.5-1.5B and Qwen2.5-7B returned parseable JSON mappings across labels-only, schema-loop, and schema-plus-correction conditions.A separate 230-example FormFactory check assessed mapping alignment, not screenshot coordinates or action trajectories.Source: §3.5 · pp. 7–8

4. Results

Full CLAIRE achieved perfect benchmark metrics across both synthetic suites, while removing validation reduced stress-suite success. The simulation produced reviewable drafts or blocked escalations, and direct LLM mapping did not match the guarded executor across both suites.

4.1. Internal benchmark comparison

  • Across 24,000 internal episodes, full CLAIRE achieved 1.000 episode success, field accuracy, required-field completion, and dependency completion in both suites, with zero final validation errors.The result applies to the synthetic schemas, rules, and interface changes evaluated.Source: Table 3 · p. 8, §4.1 · p. 8
  • Removing the Validator reduced validation-stress episode success to 0.500 and left 0.900 errors per episode, while clean-suite success remained 1.000.Removing schema hints reduced episode success to 0.200 in both suites; Labels only achieved 0.150 in both.Source: §4.1 · p. 8, Table 3 · p. 8

4.2. Operational simulation

  • All 2,000 clean and 2,000 validation-stress episodes reached reviewable drafts, while 68.6% of escalation-challenge episodes reached a draft and 31.4% were blocked.In the escalation challenge, unsafe-case capture was 100.0% and false reassurance was 0.0%.Source: §4.2 · pp. 8–9, Table 4 · p. 9
  • Estimated savings under the pre-specified timing scenario were 149.7 seconds per clean case, 165.5 seconds per validation-stress case, and 152.2 seconds per escalation-challenge case.These are calculated scenario outputs, not observed staff times.Source: §4.2 · p. 9, Table 4 · p. 9

4.3. Strict-output LLM audit

  • In validation stress, Qwen2.5-1.5B success rose from 0.230 with labels only to 0.480 with schema loop and 0.900 with schema plus correction; Qwen2.5-7B reached 0.840 with schema conditioning.Neither direct-model condition matched the guarded executor across both suites.Source: §4.3 · p. 9

5. Discussion

The discussion frames validation, escalation, and traceability as distinct safeguards for healthcare form completion. It presents CLAIRE as a staff-review accelerator, not an independent submitter, and keeps its findings within the evaluated settings.

5. Discussion

  • 0.500 stress-suite success after validation removal, versus 1.000 with full CLAIRE; clean performance remained unchanged.The benchmark figure describes complete success in clean and validation-stress suites for full CLAIRE, with degradation under dirty source values when validation is removed.Source: §5 · p. 9, Figure 2 · p. 10
  • Direct LLM mapping improved with schema context but remained less dependable than guarded execution across both suites.The focused strict-output audit averages results across four variants and 100 episodes per suite, model, and condition.Source: Table 5 · p. 10, Figure 3 · p. 11
  • The operational simulation separates routine drafts, focused review, and blocked escalation, while time savings remain scenario estimates until measured with staff.Perfect internal scores cover only failure types represented in the deterministic synthetic benchmark, not arbitrary portals or their authentication, CAPTCHA, uploads, scanned documents, or asynchronous behavior.Source: §5 · p. 12
  • Form-automation errors can delay access, contribute to denial, or create rework, so evaluation should include validation, escalation, traceability, and workflow fit.Field accuracy alone does not capture these operational concerns.Source: §5 · p. 12

5.1. Workflow integration and governance implications

  • CLAIRE is proposed as a staff-review accelerator that attaches source evidence and validation traces, blocks unresolved cases, and returns status to a work queue.FHIR Questionnaire and QuestionnaireResponse can represent questions, answers, and constraints; SMART on FHIR supports EHR integration, while the architecture complements APIs where portals and exception workflows remain.Source: §5.1 · p. 12

6. Limitations

The evaluation uses deterministic synthetic records and computable schemas, and its perfect benchmark score is limited to represented schemas and perturbations. The operational time model uses assumed rather than observed staff behavior.

6. Limitations

  • The deterministic synthetic records and computable schemas do not capture the diversity, missingness, ambiguity, coding practices, or distribution shifts of real organizations.Two form families were unchanged controls, and represented errors were selected to match implemented rules; authentication, uploads, scanned documents, CAPTCHA, production latency, and policy changes were untested.Source: §6 · p. 12
  • The primary benchmark does not establish autonomous LLM-agent competence; its perfect score evaluates the deterministic guarded executor under represented schemas and perturbations.Direct LLM mapping performance was evaluated separately in the strict-output audit.Source: §6 · p. 12
  • The operational time model used assumed task durations rather than observed staff behavior, so it cannot establish real productivity, workload, trust, or adoption.The proposed next step is a prospective staff-facing simulation comparing manual completion with trace-assisted review and measuring observed time, decisions, unsafe-case blocking, workload, and perceived trust.Source: §6 · p. 12
  • A later non-production portal study should test browser failures and workflow integration before any live submission is considered.The suggested sequence also includes staff-facing tests and EHR-integrated work queues before live deployment.Source: §6 · p. 13

7. Future work

Future work should progress through public interactive benchmarks, staff-facing tests, non-production portals, and EHR-integrated work queues before live deployment.

7. Future work

  • DECIDE-AI can guide early live evaluation, while CONSORT-AI and SPIRIT-AI would become relevant only for a later trial testing whether CLAIRE changes a process of care.The proposed progression places staff-facing tests, non-production portals, and EHR-integrated work queues before live deployment.Source: §7 · p. 13

8. Conclusion

CLAIRE places language-model reasoning inside explicit medical-informatics controls for healthcare administrative form completion. Its benchmark and operational simulation support validation, traceability, human review, and escalation, while the time-savings estimate remains dependent on timing assumptions.

8. Conclusion

  • In a 24,000-episode benchmark, schema grounding and validation were necessary for robust performance under interface perturbations and validation-stress conditions.The 6,000-episode operational simulation produced reviewable drafts, focused review, and escalation capture; estimated savings were 149.7–165.5 seconds per case under explicit timing assumptions.Source: §8 · p. 13
  • The supported design principle embeds language-assisted mapping within structured information-system controls and subjects it to validation, traceability, human review, and escalation.Source: §8 · p. 13

Ethics statement

The ethics statement limits the study to synthetic administrative data and identifies review required before real-data deployment.

Ethics statement

  • No real patient records, protected health information, or human-subject data were used.The study used synthetic healthcare administrative records and public benchmark-derived answer-key data.Source: Ethics statement · p. 13
  • The experiments did not evaluate diagnosis, treatment recommendation, or patient-facing clinical decision-making.Source: Ethics statement · p. 13
  • Deployment on real healthcare data would require institutional privacy, security, compliance, and workflow review.Source: Ethics statement · p. 13

CRediT authorship contribution statement

The contribution statement credits the authors’ roles, reports no specific grant funding, and makes synthetic benchmark materials available on request.

CRediT authorship contribution statement

  • Garapati Keerthana is credited with conceptualization, methodology, software, investigation, data curation, formal analysis, drafting, and visualization.Manik Gupta is credited with supervision, conceptualization, methodology, and review and editing.Source: CRediT authorship contribution statement · p. 13

Funding

  • The research received no specific grant from public, commercial, or not-for-profit funding agencies.Source: Funding · p. 13

Data availability statement

  • No patient data or protected health information were used; synthetic records and derived benchmark materials are available from the corresponding author on reasonable request.Source: Data availability statement · p. 14

Appendix A. AI, reproducibility, and implementation transparency

The appendix documents CLAIRE’s model-use boundaries and evaluation materials, while an external answer-key check tests field–value alignment rather than full interactive reproduction.

appendix

  • The 24,000-episode executor prompted Qwen2.5-0.5B, but model output did not authorize actions, which relied on schemas, deterministic matching, validation, and controlled correction.The separate strict-output audit evaluated executable mappings from Qwen2.5-1.5B and Qwen2.5-7B.Source: appendix · p. 14
  • The evaluation covered synthetic structured records, computable schemas, clean and validation-stress suites, four interface variants, a strict-output audit, and a 6,000-episode operational simulation.A staff-facing pilot protocol and trust survey template were prepared for follow-on evaluation; the work was evaluated in benchmark and simulation settings before live clinical or payer deployment.Source: appendix · pp. 14–15
  • Staff retain authority to approve, correct, escalate, or stop submission to a payer or vendor interface.Source: Figure A1 · p. 15

Appendix B. External answer-key alignment check

  • Exact ID automation reached 0.750 success, while Labels only, Monolithic, CLAIRE without schema, and full CLAIRE each reached 1.000 success.The secondary check covered 230 healthcare-like examples; its flattened answer-key mappings support semantic field–value alignment, not full interactive-coordinate reproduction or external clinical validation.Source: Appendix B · p. 15
  • Figure A2 shows the shared schemas, records, and interface variants used for all six comparators; Figure A3 contrasts a schema-defined target with a stress record containing invalid values and explicit corrections.Figure A3 states that its displayed names and identifiers are synthetic.Source: Figure A2 · p. 16, Figure A3 · p. 16
Loading 2609.32787v1…