PaperlayerView exact revision on arXiv

Source-linked AI summary

Foundation Agents Meet Agentic Deep Research: Evidence-Grounded Clinical Code Forecasting

Junda Wang, Meysam Ghaffari, Akshat Choube, Mohsen Sharifi Renani, Hong Yu, Carlos Morato

arXiv:2608.17075v1cs.CLcs.AI

TL;DR

Forecasting the complete set of diagnosis codes at a patient's next encounter is prospective and multi-label, with candidate hypotheses requiring patient grounding and verification. ICD-Deepresearch combines EHR and language-model forecasts with bounded medical research and joint fixed-budget selection, outperforming the evaluated standalone and isolated alternatives on MIMIC-III and MIMIC-IV. Physicians also rated its retrieved evidence as more useful than evidence from the standalone research systems.

  • Problem

    Next-encounter ICD forecasting must project a patient's longitudinal trajectory into a future multi-label code set while selecting grounded candidates under a fixed top-K budget.

  • Method

    ICD-Deepresearch combines SparseEHR and GPT-5 candidate generation with bounded medical search, ICD semantic verification, and joint fixed-budget selection.

  • Results

    ICD-Deepresearch outperforms isolated candidate paths and standalone research systems on MIMIC-III and MIMIC-IV while retrieving fewer documents with a higher proportion of useful evidence.

  • Takeaways & Limitations

    DeepResearch benefits next-encounter ICD forecasting when it controls candidate expansion, evidence acquisition, and fixed-budget selection rather than treating search as future-answer retrieval.

  • Takeaways & Limitations

    The evaluation spans MIMIC-III and MIMIC-IV from one health-system setting, limiting demonstrated generalizability across settings.

Abstract

from arXiv · show

Next-encounter ICD forecasting predicts which standardized diagnosis codes will be documented at a future visit from the longitudinal record available beforehand. The task is prospective and multi-label: the target note does not yet exist, and several codes may be correct. Structured EHR foundation models capture recurrence and temporal progression, whereas language foundation models generate flexible diagnostic hypotheses. We introduce ICD-Deepresearch, a DeepResearch workflow that composes these predictive foundation models with medical search and ICD dictionaries. Because no source reveals the future code set, research evaluates candidate transitions by linking patient evidence, external clinical relations, and exact code semantics under a fixed top-K budget. Candidate Generation uses SparseEHR to produce an EHR Prior that initializes two bounded Research Expansion rounds; an independent GPT-5 Direct Forecast supplies complementary candidates. Final Selection validates, deduplicates, and jointly ranks both paths, after which a separate module writes rationales without changing predictions. Finally ICD-Deepresearch achieves patient-averaged precision/recall of 24.60/35.09% on MIMIC-III and 25.14/48.32% on MIMIC-IV. Physicians rate 51% and 68% of its retrieved documents useful, compared with 22% and 39% for standalone GPT-5 web search and 32% and 41% for Medical Deep Research. ICD-Deepresearch therefore improves over the registered local comparators while retrieving evidence with higher physician-rated usefulness than the standalone research systems

1 Introduction · 2 Related Work · 3 Method

ICD-Deepresearch forecasts the complete multi-label ICD code set at a future encounter by combining longitudinal EHR priors with independent language-model candidates and evidence-grounded medical research. Its two-path pipeline researches only patient-anchored EHR candidates, jointly selects validated codes under a fixed top-K budget, and explains them only after prediction.

  • 1 Introduction: Next-encounter ICD forecasting is prospective and multi-label: the target code set lies beyond the observation window and may contain multiple diagnoses.The model projects the observed longitudinal trajectory into the next documented state.
  • 1 Introduction: Longitudinal EHR models provide patient-specific recurrence and progression priors, while language foundation models propose complementary diagnoses from broader medical knowledge.This motivates combining structured trajectory modeling with flexible language-model hypotheses.
  • 2 Related Work: Prior work covers longitudinal multi-label forecasting, adaptive retrieval, clinical research agents, and retrospective coding, but next-encounter forecasting begins without the target encounter or an explicit query.Medical Deep Research is evaluated as a standalone domain-specific research comparator, whereas Research Expansion pairs patient trajectories with research.
  • 3 Method: ICD-Deepresearch has parallel Candidate Generation and Direct Forecast paths, with SparseEHR producing the EHR Prior and GPT-5 producing an independent code list.Only the EHR Prior initializes Research Expansion, preserving Direct Forecast as an independent source of candidates until Final Selection.
  • 3 Method: Two bounded Research Expansion iterations apply Plan →ToolAct →ReadBind →Verify →UpdateRank to connect patient clues, external clinical transitions, and exact ICD semantics.Tools include online Google, fixed PMC/arXiv corpora, and the dataset-specific ICD dictionary.
  • 3 Method: Final Selection canonicalizes, validates, deduplicates, and jointly ranks the research and direct candidates under one fixed top-K budget.Post-Selection Explanation then writes rationales and uncertainty statements without changing the selected codes or their order.

4 Experimental Design

The experiments evaluate forecasting performance, system components and selectors, retrieval quality and efficiency, and post-selection explanations on held-out MIMIC-III/IV admissions. They compare the full workflow with local and published baselines using dataset- and table-specific evaluation protocols.

  • Evaluation logic: Evaluation covers forecasting, component and selector controls, retrieval quality and efficiency, and post-selection explanations across RQ1–RQ4.The analysis uses held-out final admissions from MIMIC-III and MIMIC-IV, with ICD-9-CM or ICD-10-CM targets.
  • Evaluation logic: Comparisons use held-out final admissions from MIMIC-III/IV, with ICD-9-CM or ICD-10-CM targets and separate scale and diagnostic cohorts.Comparisons remain within each table and dataset.
  • Comparators and controls: Local controls compare Candidate Generation Only and the full system with standalone GPT-5 Web Search, while isolating Direct and Research paths and varying research context under a fixed candidate pool.Medical Deep Research is included as a second standalone comparator.
  • Metrics and audits: Table 1 reports patient-averaged P@K/R@K, whereas Table 2 reports pooled micro-P/R/F1 and patient-macro F1.The two tables use distinct evaluation metrics and cohort constructions.
  • Metrics and audits: Retrieval audits measure source composition, metadata support and forecast usefulness, and blinded physician usefulness; explanations use five reason dimensions and coverage-adjusted Reason5.Supplementary sections define metric, inference, source-sampling, efficiency-accounting, and explanation-judge procedures.

5 Results and Discussion

ICD-Deepresearch improves next-encounter ICD forecasting over registered local comparators while combining candidate-generation paths, research expansion, and joint selection. Its retrieved evidence receives higher physician-rated usefulness, although expansion benefits are uneven and immediate-horizon relevance remains the weakest explanation dimension.

  • Overall forecasting performance: 24.60/35.09% and 25.14/48.32% patient-averaged P@20/R@20 are achieved on MIMIC-III and MIMIC-IV, respectively, exceeding both local comparators at K = 20 and K = 10.At K = 20, gains over GPT-5 + Web Search are +1.83/+2.61 points on MIMIC-III and +2.96/+16.76 on MIMIC-IV; gains over Candidate Generation Only are +3.68/+5.25 and +1.79/+15.10.
  • Expansion and selection diagnostics: +0.12 and +1.90 micro-F1 points are produced by Research Expansion under a fixed candidate pool on MIMIC-III and MIMIC-IV, while fixed-pool oracles remain 5.75 and 5.59 points higher.The control fixes candidate identities, order, GPT-5 backend, and K = 20, isolating context-assisted ranking after candidate generation.
  • Expansion and selection diagnostics: 144 additional targets are recovered on EHRSHOT as mean pool size rises from 19.8 to 51.0 and micro-F1 falls from 25.82% to 16.34%; a support-score filter raises micro-F1 from 26.02% to 26.54%.Initial expansion changes micro-F1 by +3.06 and +9.04 points on MIMIC-III and MIMIC-IV, whereas follow-up changes it by −0.98 and −0.004; 26 of 28 MIMIC-IV follow-up matches occur in 13 patients while 52 patients worsen.
  • Retrieval usefulness and efficiency: 51% and 68% physician-rated usefulness are attained by Research Expansion on MIMIC-III and MIMIC-IV, despite 1,410 and 11,124 source occurrences versus substantially larger standalone retrieval sets.Research Expansion also uses 7.3×/3.4× fewer unique queries, 3.8×/4.5× fewer source accesses, and 6.1×/2.9× fewer visible tokens, with 16.8×/21.9× higher recorded batched throughput.
  • Explanation quality and limitations: 0.07 and 0.22 point improvements in exact ICD alignment and evidential calibration occur across 441 MIMIC-III shared hits, while MIMIC-IV quality improves across several dimensions and next-encounter relevance changes by only +0.03.The row-45 case shows higher shared-hit Reason5 means for ICD-Deepresearch than GPT-5, 4.80 versus 3.47; next-encounter relevance is reported as the lowest Reason5 dimension.

6 Conclusion

Next-encounter ICD forecasting is framed as budgeted hypothesis search: research expansion improves candidate coverage but increases the burden on Final Selection. Selection retains measurable headroom, while research context yields modest gains and independent relation-source coverage remains limited.

  • Conclusion: Candidate expansion broadens coverage while enlarging the pool that Final Selection must resolve.The paper characterizes next-encounter ICD forecasting across four research questions as budgeted hypothesis search.
  • Conclusion: 8.28 micro-F1 points of joint-pool selection headroom remain under the MIMIC-III target-conditioned oracle.This quantifies remaining selection capacity in the reported MIMIC-III oracle analysis.
  • Conclusion: +0.12 and +1.90 micro-F1 changes result from attaching research context on MIMIC-III and MIMIC-IV, respectively.The reported gains are dataset-specific and correspond to MIMIC-III and MIMIC-IV in that order.
  • Conclusion: Independent relation-source coverage is only 11.62% on MIMIC-III and 1.39% on MIMIC-IV.These values indicate limited overlap with independently covered clinical relations.

S1 Guide to the Supplementary Materials

The supplement provides implementation details, diagnostics, prompt contracts, case traces, and artifact provenance while following the main paper’s authoritative definitions. It covers system components, evaluation methodology, functional results, interpretation, limitations, ethics, and executable-call provenance.

  • Relationship to the main paper: The main paper remains authoritative for the reader-facing method definition, component names, information boundary, evaluation contracts, and reported claims.The supplement follows those definitions and adds implementation-oriented material rather than redefining the method.
  • Supplement organization: Section S2 specifies the EHR Prior, Direct Forecast, Research Expansion, Final Selection, and explanation writer.These are the principal system components documented in the supplement.
  • Supplement organization: Section S3 describes cohorts, comparisons, metrics, source and efficiency accounting, uncertainty, and audit contracts.This section defines the supplement’s evaluation and auditing details.
  • Supplement organization: Sections S4–S6 report functional results, retrieval efficiency, mechanism diagnostics, interpretation, limitations, ethics, exact figure values, and executable-call provenance.Together, these sections extend the reported analysis and document the provenance of quantitative and executable artifacts.
  • Artifact provenance: Quoted system messages and natural-language instructions are transcribed from executable sources identified in Table S6.Long JSON contracts are shown verbatim or as complete required-field lists, with dynamic fields marked by angle-bracket tokens and truncation or retry suffixes recorded.

S2 Full Method Specification

ICD-Deepresearch forecasts held-out diagnosis-code sets from a target-blind pre-encounter record through three forecasting stages and a separate post-selection explanation stage. Its research branch expands an EHR-derived seed for exactly two iterations, then validates and jointly ranks researched and independently generated candidates under fixed vocabulary and selection protocols.

  • Information boundary: Inference uses only the pre-target record and independently fixed resources, with target codes and target-encounter notes withheld until predictions are fixed.The held-out target is accessed only for metric computation and by the explicitly gold-aware explanation-evaluation judge.
  • Research Expansion: Research Expansion starts only from the unordered EHR seed, maintains candidate hypotheses and evidence state, and applies the registered state transition exactly twice.The available tools include online Google, fixed PMC/arXiv corpora, and the dataset-specific ICD dictionary; the second iteration conditions on the first iteration’s state and verification outcomes.
  • Final Selection: Final Selection validates and deduplicates the researched pool before combining it with Direct Forecast under either Union+RRF or the Evidence-Aware Reranker, while explanation runs afterward.The two selectors compare selection rules while fixing candidate availability, and the downstream explanation audit uses finalized RRF-selected outputs.
  • Forecasting stages: Candidate Generation produces a registered 20-code EHR Prior and seed set from structured EHR, while Direct Forecast independently returns a bounded target-vocabulary ranking.SparseEHR supplies the structured-EHR backbone; Direct Forecast receives no prior, research candidate, retrieved source, or other-system output before final selection.
  • Candidate formulation: Candidate records bind each proposed code to a clinical term, concrete patient clue, temporal class, ongoing-evidence flag, falsifiable transition hypothesis, and search phrase.Allowed temporal classes include persistent status, chronic active disease, recurrent condition, planned follow-up, treatment consequence, and acute or laboratory speculation.

S3 Full Experimental Protocol

The experimental protocol separates scale-cohort primary forecasting from subset component and selector analyses, using fixed patient targets, configurations, and exact-K evaluation contracts. It compares ICD-Deepresearch components, standalone research systems, and retrieval usefulness under target-blind, reproducible procedures.

  • Datasets and targets: MIMIC-III targets each patient’s final-admission unique ICD-9-CM diagnoses, while MIMIC-IV uses the corresponding ICD-10-CM construction.SparseEHR supplies Candidate Generation’s structured-EHR backbone for both ICD tasks.
  • Primary analysis: Scale-cohort primary analysis covers 7,496 eligible MIMIC-III trajectories and 62,537 completed MIMIC-IV trajectories under a patient-averaged exact-K contract.These scale-cohort values are distinct from the subset results in the component table and are not numerically compared with them.
  • Evaluation cohorts: The component comparison uses one development-disjoint MIMIC-III cohort and one frozen-order MIMIC-IV cohort, each with n = 100 patients.The same identifiers and target sets are used across configurations, which are independently executed three times under K = 20.
  • Comparators and ablations: Panel A compares five ICD-Deepresearch configurations, while Panel B compares GPT-5 + Web Search (Standalone) and Medical Deep Research (Standalone) on the same diagnostic cohorts.Research Expansion starts from the Candidate Generation seed set, and Final Selection combines its finalized state with Direct Forecast.
  • Prompt and retrieval controls: Direct Forecast and GPT-5 + Web Search use target-blind prompts specifying the next-encounter horizon, exact-code specificity, uncertainty handling, structured output, and validation retries.Web Search additionally requires tool use, allows an uncapped self-selected search count, and requests an exhaustive ranking.
  • Evidence evaluation: Physician screening samples 50 source occurrences per system–dataset cell, with two physicians independently assigning binary usefulness labels to every sampled document.The reported usefulness rate is the arithmetic mean of the two physician-specific 50-document rates.

S4 Additional Results and Diagnostics

Additional diagnostics show that ICD-Deepresearch’s joint selection improves micro-F1 over isolated and comparator paths, while research expansion introduces a coverage–cardinality trade-off. Supplementary audits also indicate efficiency gains and stronger judged alignment, alongside limitations in automated validation and patient anchoring.

  • Joint-selection ablations: ICD-Deepresearch exceeds Candidate Generation Only, Direct Forecast Only, Research Path Only, and RRF selection by 4.02, 3.69, 4.98, and 2.52 MIMIC-III micro-F1 points, respectively.On MIMIC-IV, it reaches 35.44 micro-F1: +4.08 over Direct Forecast Only, +2.54 over Research Path Only, and +0.59 over the RRF selector.
  • Expansion and retention diagnostics: Research rounds recover 144 additional EHRSHOT held-out matches, but mean set size more than doubles as micro-F1 falls from 25.82% to 16.34%.Post-hoc accounting finds initial expansion clears the micro-F1/2 boundary on both MIMIC diagnostics, whereas Research Round 2 falls below it on all five.
  • Expansion and retention diagnostics: The support-score rule retains 21 of 28 Research Round 2 matches while reducing candidate growth, outperforming source-availability rules on retained matches.These findings motivate development and prospective evaluation of a state-aware retention policy.
  • Research-context controls: On MIMIC-III, adding research context changes mean micro-F1 by +0.12 points, with a 95% CI of −1.12 to +1.33, under a fixed target-blind pool and K = 20.On MIMIC-IV, the corresponding change is +1.90 points, reported as a descriptive point estimate without an archived interval.
  • Evidence audits: On 441 MIMIC-III shared hits, ICD-Deepresearch scores 0.07 points higher in ICD alignment and 0.22 points higher in evidential calibration than the RRF-selected comparator.The audit is automated and gold-aware rather than independent clinical validation; GPT-5 generates several compared outputs, leaving judge–generator dependence as a limitation.

S5 Extended Interpretation, Limitations, and Ethics

The section clarifies that next-encounter code forecasting measures recovery of future documentation rather than clinical benefit, and documents important evaluation, reproducibility, and deployment limitations. Clinical use therefore requires stronger validation, safeguards, and oversight.

  • Interpretation: Next-encounter metrics assess recovery of the future documentation set, not clinical decision benefit, which requires severity grading, workflow simulation, and prospective evaluation.At K = 20, the output budget may exceed the mean target size and expose multiple false positives per patient.
  • Reproducibility and scope: The primary analysis uses an Evidence-Aware Reranker on saved, target-blind proposal states, whereas end-to-end replay additionally requires the full Candidate Generation implementation and live Web results.The explanation audit uses an archived RRF-selected prediction prefix and does not audit the full Evidence-Aware output.
  • Evaluation limitations: Source-usefulness percentages are descriptive: physicians independently label sampled documents, with rates averaged across their 50-document assessments and no adjudicated consensus or uncertainty interval.The source audit samples 50 retrieved occurrences per system–dataset cell.
  • Evaluation limitations: GPT-5 reason-quality judging is not independent clinical validation because GPT-5 also generates several compared outputs; dual-clinician annotation and adjudication would strengthen evaluation.Saved traces also reveal unsupported patient anchors that external sources cannot validate on their own.
  • Ethics and deployment: Clinical deployment requires de-identification, controlled source retention, prompt-injection defenses, deterministic anchor and relation verification, severity-aware false-positive analysis, and clinician or certified-coder oversight.The archive contains partial token, latency, cost, prompt, checkpoint, and runtime records, while live search results change over time.

S6 Prompt, Figure, and Case Reproducibility Archive

The reproducibility archive records the exact values and implementation provenance underlying the explanation figures, while reporting comparative CA-RQ5 intervals and independent-source coverage for ICD-Deepresearch. It also maps archived LLM/API calls and legacy interfaces to the reader-facing forecasting functions.

  • Exact Values for the Explanation Figures: Table S5 supplies displayed values for main-paper Figure 3a–c, including pooled true-positive Reason5 means and CA-RQ5 at K = 20.Means are rounded to two decimals, while figure differences and intervals use unrounded item-level values.
  • Exact Values for the Explanation Figures: The ICD-Deepresearch arm in the saved implementation records is RRF-selected and labeled AF/Union+RRF, distinct from the full Evidence-Aware output.This identifies the exact arm associated with the archived figure values.
  • Exact Values for the Explanation Figures: −0.02 to +0.09 and +0.15 to +0.27 are the 95% confidence intervals for ICD-Deepresearch minus GPT-5 CA-RQ5 differences on MIMIC-III and MIMIC-IV, respectively.The comparison uses complete explanation configurations with different supplied input states and does not isolate any single evidence field.
  • Exact Values for the Explanation Figures: 58/499 (11.62%) and 8/577 (1.39%) independent clinical-relation sources cover ICD-Deepresearch true-positive explanations.The reported coverage values correspond to MIMIC-III and MIMIC-IV, respectively.
  • Prompt and Artifact Provenance: Historical Stage 1/Stage 2 source identifiers correspond to Iterations 1/2 of Equation 12.The mapping distinguishes archived source identifiers from reader-facing system names.
  • Prompt and Artifact Provenance: Internal FCG/DCG/DR-CG/JCS and legacy AF/FR labels identify archived interfaces and runners rather than reader-facing system names.The reader-facing names are EHR Prior, Direct Forecast, Research Expansion, Final Selection, and ICD-Deepresearch.
  • Prompt and Artifact Provenance: FR is the Evidence-Aware selector for the full ICD-Deepresearch system, whereas AF/Union+RRF denotes the RRF-selected arm.The distinction prevents conflating the full Evidence-Aware output with the saved RRF-selected implementation record.
  • Prompt and Artifact Provenance: Table S6 inventories executable calls by forecasting function, with deterministic and API rows defining complete execution boundaries.The inventory also identifies executed call sites through line ranges and archival source descriptions.

Post-selection explanation

Post-selection explanation uses candidate-bound search and reader observations, then applies deterministic ICD dictionary processing to normalize and validate target-code identity. The documented execution boundary records provenance across explanation writers and related forecasting, selection, and audit functions.

  • Explanation controls: The workflow uses SearchCans/Google, PubMed or fixed PMC/arXiv workers, and a local ICD dictionary.These are the specified search, reading, and dictionary components.
  • Explanation controls: Search engines and readers return observations tied to candidates, while deterministic dictionary processing normalizes and validates target-code identity.This separates evidence retrieval from code-identity control.
  • Execution boundary: The execution boundary covers LLM calls, search and page-reading APIs, local ICD lookup, validation gates, RRF, explanation writers, and audit judges.These components provide call-level provenance for the EHR Prior, Direct Forecast, Research Expansion, Final Selection, explanation, and audit functions.

S7 Prompt Design and Internal Call Contracts

The section specifies executable prompt contracts and deterministic output requirements for ICD-Deepresearch’s candidate ranking, research calls, direct branch, and selector variants. Its ranking policy prioritizes patient-specific temporal evidence, exact-code specificity, and bounded confidence while preserving plausible acute and direct candidates.

  • Ranking rules: Patient-specific pre-target evidence dominates population prevalence and generic literature, while direct and research ranks remain imperfect priors rather than labels.Direct top-20 membership remains an independent forecast signal, and direct candidates with plausible patient anchors are preserved unless concretely contradicted or outcompeted.
  • Ranking rules: Research-only candidates require both a concrete patient anchor and an exact temporal transition; frequency-only and unsupported candidates are ranked near the bottom.The rules also prohibit converting risk factors into diagnoses and require exact specificity rather than family-level inference.
  • Ranking rules: Acute diagnoses remain competitive when supported by prior occurrence, current physiology, relevant procedures, or recurrent pathways, even without web evidence.The acute-preservation rule specifically warns against suppressing AKI, heart failure, respiratory failure, infection, or acute metabolic disorders merely because chronic codes are easier to justify.
  • Output contracts: The reranker returns every supplied candidate exactly once with confidence in [0, 100], patient-anchor and exact-code-support categories, and a patient-specific eov_threshold.Candidates with confidence ≥ eov_threshold form the variable-length prediction set, and no new codes may be introduced.
  • Selector contracts: RRF combines research rank with independent Direct GPT-5 rank, rewarding cross-branch agreement and potentially promoting complementary mid-ranked codes without yielding probabilities or proving retrieval caused rankings.The direct branch receives no retrieval tool, candidate pool, baseline prediction, dictionary content, or other system output.

S8 Case Studies

The case studies show how ICD-Deepresearch combines direct forecasting with research expansion and joint ranking, while exposing failures caused by unsupported patient anchors. They also motivate pre-search anchor resolution and separate entailment and horizon checks.

  • Safety failures: Failed expansion attached credible GERD and recurrent-back-pain relations to invented patient anchors, producing false additions including K21.9 and M54.5.Six further false additions were omitted, and some candidates relied only on dictionary definitions or reversed immediate-horizon direction.
  • Safety failures: Every candidate anchor should be resolved to an immutable event identifier or exact pre-target span before search, with source entailment and immediate-horizon direction checked separately.This safeguard directly addresses the failed-expansion trace’s unresolved patient claims and directional errors.
  • Complementary forecasting paths: 5 true positives and F1@20=.286 were obtained by Direct Forecast Only, with E27.40 ranked 19th despite uncertain imminent adrenal decompensation.Research Path Only ranked E27.40 25th, illustrating different candidate priorities across forecasting paths.
  • Complementary forecasting paths: 6 true positives and F1@20=.343 were obtained by Research Path Only, ranking F32.9 fifth and Z87.891 tenth.Direct Forecast Only ranked these codes 28th and 27th, respectively.

D2 ICD-Deepresearch w/o Evidence-Aware Reranker

Without the Evidence-Aware Reranker, ICD-Deepresearch combines complementary direct and research candidates, achieving 8 true positives with P/R/F1=.400/.533/.457 in the evaluated case. Its research trace also exposes inconsistencies in direct-forecast code interpretations while linking selected code titles to evidence.

  • Candidate complementarity: E27.40 entered through the direct rank, whereas F32.9 and Z87.891 entered through research.The two paths contributed distinct candidates to the final set.
  • Candidate complementarity: E87.0 was promoted to rank 20 from moderate component ranks 25/23 by agreement.Agreement between component rankings supported the promotion.
  • Evaluation: 8 true positives yielded P/R/F1=.400/.533/.457, illustrating candidate and rank complementarity in the case.The evaluator result is reported for ICD-Deepresearch w/o Evidence-Aware Reranker.
  • Interpretation: GPT-5 Direct Forecasting produced contradictory descriptions for O34.33, O34.13, B95.61, and O90.81.The cited interpretation identifies errors involving uterine-scar codes, an E. coli organism description, and a generic puerperal-complication description.
  • Interpretation: The ICD-Deepresearch writer exposed selected code titles through resolvable citations and connected the observed fibroid trajectory to O34.13.This contrasts with the internally contradictory entity descriptions attributed to GPT-5 Direct Forecasting.

S9 Cross-Case Error Analysis

Cross-case traces identify distinct Research Expansion failure modes and motivate a three-link validation contract connecting pre-target anchors, source-entailed directional relations, and exact immediate-horizon ICD candidates. They also show that complementary branch rankings can preserve true positives, though the final observation is only a single-case illustration.

  • Validation contract: Research Expansion is evaluated against a three-link contract: a resolvable pre-target anchor, a source-entailed directional relation, and an exact immediate-horizon ICD candidate.The contract links patient evidence, external relations, and precise code selection.
  • Failure mechanisms: Row 52 recovers candidates but reveals mapping errors and reliance on definition-only support.Candidate recovery alone does not ensure correct code mapping or directionally relevant evidence.
  • Failure mechanisms: Row 59 breaks both the anchor and horizon links required by the validation contract.This case fails at identifying a suitable pre-target anchor and at matching the immediate forecast horizon.
  • Selector behavior: In row 62, the RRF selector retains true positives available at complementary ranks across the two branches.The case illustrates the benefit of combining candidate rankings from the two research paths.
Loading 2608.17075v1…