acceptodds
Desk rejected at ICLR 2027

Sleep as a Window: Auditable LLM Decoding of Aging Signatures from Integrated Sleep Profiles

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Biological aging progresses at different rates across organs, creating patterns of vulnerability that chronological age alone cannot capture. Sleep offers an information-rich window into this heterogeneity, with each participant's full clinical record providing context for its interpretation. Existing sleep-age models usually yield a single estimate, whereas large language model (LLM) outputs can be difficult to trace to patient evidence. We introduce an evidence-auditable LLM framework that decodes integrated sleep profiles into an overall sleep age and ten organ-resolved age representations. A frozen registry curated by clinical sleep experts and trace-validated reports preserve input provenance. The framework also audits the accompanying physiological chains for major contradictions. We evaluated the framework in 9,123 participants from SHHS, MrOS, and the real-world XH longitudinal cohort. After adjustment for chronological age and conventional risk factors, mortality HRs per year of sleep gap were 1.054 (95% CI 1.031–1.079) in SHHS and 1.030 (1.008–1.052) in MrOS. Ten-year mortality was higher in the top than in the bottom gap decile (SHHS 25.4% versus 8.9%, MrOS 25.5% versus 13.0%). Models combining chronological age with endpoint-matched organ gaps outperformed those using non-matched gaps in C-index (mean proportion outperformed: SHHS 92%, MrOS 88%, XH 84%). This pattern supports LLM decoding of organ-specific aging signatures. In XH, overall sleep gaps tracked incident disease and cumulative disease burden, and endpoint-matched gaps added predictive value beyond multiple clinical baseline models. Together, these findings position integrated sleep profiles as an auditable window onto systemic aging heterogeneity and its future health consequences.

Primary area:
foundation or frontier models, including LLMs
Keywords:
large language models, medical AI agents, aging signatures, sleep medicine, evidence traceability, organ-resolved aging, longitudinal health outcomes

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_fdjCOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2(+1)
Soundness
2
Read review
Paper summary

This paper proposes an evidence-auditable LLM framework that produces an overall sleep-age estimate and ten organ-specific age representations from integrated sleep and clinical profiles. The framework comprises three agents. The Evidence Orchestration Agent organizes expert-registry-defined evidence into a section-wise Evidence-ID plan, which is rendered and checked for source fidelity. The Age Estimation Agent receives the resulting report and generates eleven age estimates with accompanying physiological explanations. The Consistency Audit Agent screens the frozen estimator output for two prespecified types of contradiction without modifying the age estimates. The authors evaluate associations between the resulting age gaps and mortality, incident disease, and disease burden across three cohorts.

Critical strengths
  1. The problem is interesting and significant. Estimating overall and organ-specific aging signatures from sleep and clinical information could support more informative health-risk assessment.
  2. Combining source-traceable inputs with inspectable physiological explanations is a promising direction for LLM-based age estimation. Separating input-fidelity checks from explanation-consistency checks is also a useful design choice.
  3. The paper provides extensive experimental results across three cohorts, including longitudinal outcome associations, organ–outcome correspondence, and comparisons with alternative LLM backbones and conventional age-prediction models.
Critical weaknesses
  1. The estimation procedure and the contribution of the agents require clearer explanation. Although the supplement provides prompts and output schemas, it remains unclear what justifies the specific numerical ages produced by the estimator. The authors should explain what validates these outputs as ages measured in years, beyond their associations with clinical outcomes. Furthermore, the benefit of LLM-based evidence orchestration over deterministic organization using the same registry is not established. A comparison with a simpler pipeline would help isolate the contribution of the proposed architecture.
  2. The LLM implementation details should be consolidated and made more explicit. The paper identifies Llama 3.3-70B as the primary backbone and evaluates six alternatives, but the model assignments for each agent and the model-based audit reference should be clearly stated. The authors should also specify whether any models were fine-tuned or otherwise adapted, provide generation settings and retry procedures, and explain how sleep-specific or medical knowledge are injected into the system beyond the prompts and expert-curated registry.
  3. The separation between representation generation and statistical evaluation should be clearer. The agents generate and audit the age representations, while Cox models subsequently evaluate their associations with outcomes. Although supplementary survival-analysis specifications are provided, the main text should make this separation explicit. Additional implementation details—such as the software used, proportional-hazards diagnostics, and pooling of estimates across multiple imputations—would improve reproducibility.
  4. The scope and methodological novelty of “evidence-auditable” need clearer defnition of range. The framework establishes input provenance and fidelity and screens generated explanations for two narrowly defined contradiction types. However, these checks do not establish that each physiological claim is supported by the participant’s evidence, that the explanation is clinically valid, or that the numerical age is justified. Input fidelity is a fundamental requirement for reliable analysis; its implementation alone does not establish a substantial methodological contribution. The authors should clarify what additional auditing capability the framework provides beyond standard provenance tracking and validation, and evaluate that capability directly.
Questions for authors
  1. What determines the magnitude of each numerical age adjustment, and what evidence supports interpreting these outputs in years of aging rather than as clinical risk scores?
  2. Which model is used for each of the three agents and for the developmental audit reference? Were these models used solely through prompting, or was any fine-tuning or other adaptation performed?
  3. Beyond the expert-curated registry and prompts, does the estimator receive external medical knowledge, retrieval results, or task-specific examples?
  4. What benefit does the Evidence Orchestration Agent in estimating the sleep age and organ specific age? Can you provide a comparison with a simpler single-estimator pipeline?
  5. Can you clarify the Cox-model implementation, including proportional-hazards checks, missing-data handling, and pooling across imputations?
  6. How much of the predictive information comes from sleep measurements versus the accompanying clinical history? Would sleep-only and clinical-context-only analyses yield similar results?
6 comments
AuthorsOpenReview ↗

Response to Reviewer fdjC (Part 1 of 3)

We sincerely appreciate the reviewer's detailed comments and helpful suggestions. They have substantially improved the clarity and methodological transparency of the manuscript. We have clarified the numerical age scale, documented the agent settings and Cox analyses, and strengthened the evaluation of consistency auditing.

W1/Q1: Interpreting the numerical age scale. The prompt asks the LLM to weigh evidence of aging and resilience and estimate one overall sleep age and ten organ-specific ages (Figure M.2). The model determines the magnitude rather than adding a preset number of years per finding. Across the six joint surrogates, each additional year of chronological age corresponded to 0.97–1.03 decoded years with other inputs fixed (Figure E.2). A 10-mmHg higher systolic pressure corresponded to a cardiovascular age estimate higher by 0.69 years in SHHS, 0.34 in MrOS and 0.57 in XH. Held-out mean absolute errors were 1.49–2.20 years for overall sleep age and 2.72–3.32 years for cardiovascular age. Larger overall sleep gaps also tracked greater frailty in SHHS and MrOS and higher clinical-deficit burden in XH after adjustment for age and, where applicable, sex (Figure E.1).

W1/Q4: Function of evidence orchestration. The registry defines eligible evidence across the cohort. The orchestration agent selects nonredundant Evidence IDs for each patient and assigns them to six report sections. The renderer then copies the source labels, values and units. Compared with prompting an estimator directly, the extra step produces an inspectable record of which measurements reached the estimator and which were omitted (Figure 2, Section O).

W2/Q2–Q3: Models, medical knowledge, and inference. Section B.8 and Table B.2 consolidate role assignments and execution settings. Orchestration and estimation used Llama 3.3-70B Instruct. The Consistency Audit Agent used GPT-OSS 120B, as specified in Table B.2. Figure L.1 evaluates its decisions against clinical expert annotations for all 426 XH outputs. All agents use zero-shot prompting, with no fine-tuning, external retrieval, or labeled patient demonstrations. Medical knowledge comes from pretrained weights and the prompts, while the expert registry supplies clinical semantics. All agents were accessed through OpenRouter. The initial age-estimation response used greedy decoding (temperature 0.0, top-p 1.0), with unchanged settings for transport retries. Age vectors remain fixed after auditing. Section 6.2 discusses adoption of newer compatible backbones, including local open-weight deployment and closed models where governance permits.

W3/Q5: Representation generation and Cox evaluation. The LLM generated ages before outcome linkage. We used lifelines to fit Cox models to the fixed gaps (Sections 4.2 and B.7). In SHHS, we replaced a follow-up CVD indicator with baseline cardiovascular history and reconciled event dates against the event log and last contact. For SHHS/MrOS models with missing covariates, we pooled log HRs across five stochastic predictive-mean-matching imputations using Rubin's rules. The adjusted SHHS mortality HR per year of sleep gap is 1.048 (95% CI 1.025–1.073; Figure 4a).

W4: Additional capability of consistency auditing. Once trace checks verify input fidelity, the Consistency Audit Agent screens the generated explanations for contradictions. Two clinical experts jointly annotated all 426 XH outputs for the two prespecified contradictions. Against these clinical expert annotations, the Consistency Audit Agent's timing check detected 29/29 order reversals with no false positives. Its state check detected 4/10 incompatible states of the same variable, with 11 false positives. The agent had overall sensitivity of 86.5% and specificity of 97.4% (Figure L.1).

AuthorsOpenReview ↗

Response to Reviewer fdjC (Part 2 of 3)

Q6: Sleep measurements and clinical context. We appreciate this suggestion and agree that comparing sleep-only and clinical-only inputs would help distinguish their respective prognostic value. In practice, overnight findings and clinical history are available together. Our aim was to determine whether the age gap decoded from this integrated sleep profile improves prediction once routine clinical risk factors are known. Across 46 organ outcomes, adding the endpoint-matched gap to each of four clinical-only models improved mean held-out C-index (Table H.1). The gain was 0.0437 over age and sex and 0.0159 (95% CI 0.0103–0.0221) over the model that also included BMI, systolic blood pressure, smoking, alcohol use, hypertension and diabetes.

To examine how the two sources of information relate to the decoded ages, we fitted a linear surrogate in each cohort to approximate the LLM's overall sleep age (Figure E.2). Across cohorts, 30 more minutes awake after sleep onset corresponded to overall sleep age estimates 0.02–0.07 years higher. A 10-mmHg higher systolic pressure corresponded to 0.10–0.15 additional years in the same models. These coefficients describe each measurement's association with decoded age after accounting for the other inputs.

Manuscript locations

The locations below refer to the revised manuscript PDF, where revisions are highlighted in blue.

  • Section 3, lines 214–219 (pp. 4–5); Appendix E.1, lines 1441–1475 (pp. 27–28) explain numerical age estimation and the surrogate analyses in clinical units (Figure E.2, p. 29).

  • Section 5.4, lines 439–443 (p. 9); Figure E.1, lines 1434–1439 (p. 27) report age-adjusted associations with frailty and clinical-deficit burden.

  • Section 3, lines 195–212 (p. 4); Appendix O, lines 2470–2479 (p. 46) describe patient-specific evidence plans, source-preserving rendering, and trace validation.

  • Appendix B.8, Table B.2, lines 1113–1133 (p. 21) consolidates model assignments, OpenRouter access, inference settings, and the clinical annotation reference.

  • Section 4.2, lines 249–280 (pp. 5–6); Appendix B.7, lines 1028–1045 (p. 20) separate age generation from Cox evaluation and document event reconciliation, software, and imputation pooling.

  • Section 4.4, lines 303–308 (p. 6); Appendix L, lines 1856–1867 (p. 35) report clinical expert annotation and audit performance for the two prespecified contradictions (Figure L.1).

  • Section 5.4, lines 434–438 (p. 9); Appendix H, Table H.1, lines 1674–1700 (p. 32) report the added discrimination of matched gaps across four clinical models.

  • Section 6.2, lines 523–529 (p. 10) discusses input comparisons, independent validation, and compatible model backbones.

Tables cited in this response

The tables below reproduce the corresponding manuscript tables. Table numbers are retained for cross-reference.

Table B.2. Audit reference, model assignments, and inference settings

Token limits are maximum completion lengths, and transport retries recover failed API requests.

RoleModel or annotatorProcedure or settings
Evidence Orchestration AgentLlama 3.3-70B Instruct4,096 tokens, four transport retries, up to three case attempts per run.
Age Estimation AgentLlama 3.3-70B Instruct40,960 tokens, fourteen transport retries.
Consistency Audit AgentGPT-OSS 120B4,096 tokens, three transport retries, medium reasoning effort.
Audit referenceTwo clinical expertsJoint annotation of measurement-order reversal and incompatible states of the same variable in all 426 XH outputs.
AuthorsOpenReview ↗

Response to Reviewer fdjC (Part 3 of 3)

Table H.1. Clinical adjustment and organ correspondence across 46 organ outcomes

ModelClinical C-index+ Matched gap C-index+ Non-matched gap, mean C-indexMatched gain ΔC\Delta CMatching advantage ΔC\Delta C
A0.6639 (0.6471–0.6807)0.7076 (0.6926–0.7225)0.6758 (0.6615–0.6899)+0.0437 (0.0286–0.0579)+0.0318 (0.0231–0.0399)
B0.6806 (0.6649–0.6957)0.7072 (0.6928–0.7209)0.6854 (0.6713–0.6991)+0.0266 (0.0178–0.0355)+0.0218 (0.0154–0.0279)
C0.6823 (0.6667–0.6965)0.7048 (0.6900–0.7185)0.6849 (0.6700–0.6984)+0.0225 (0.0157–0.0294)+0.0199 (0.0149–0.0251)
D0.7000 (0.6855–0.7139)0.7159 (0.7023–0.7284)0.7019 (0.6881–0.7151)+0.0159 (0.0103–0.0221)+0.0140 (0.0094–0.0187)

Clinical models. A, chronological age and sex. B, A plus BMI and systolic blood pressure. C, B plus smoking and alcohol use. D, C plus hypertension and diabetes history. Sex was omitted when invariant. These covariates were identical across endpoints, without substituting organ-specific histories. Evaluation. Entries average equally over the same 46 organ outcomes as Figure 6 (10 SHHS, 16 MrOS, and 20 XH). All-cause death and the XH first-new-diagnosis composite were analyzed separately. Each endpoint used identical participants and five folds for all models. Each gap was added separately. The non-matched comparator averages the C-indices of the ten alternative gaps. Matched gain is the matched-gap model minus the clinical model. Matching advantage is the matched-gap model minus the non-matched mean. Differences were calculated before rounding. Uncertainty. Parentheses give pointwise 95% percentile intervals from 2,000 paired participant bootstrap draws, retaining correlations across outcomes within each cohort and combining cohorts independently.

Reviewer_kkA3OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes an evidence-auditable, multi-agent LLM framework for estimating overall sleep age and ten organ-specific ages from integrated sleep profiles. The framework combines an expert-curated evidence registry, an Evidence Orchestration Agent, an Age Estimation Agent, and a Consistency Audit Agent to preserve source provenance and flag prespecified contradictions. The method is evaluated on 9,123 participants across SHHS, MrOS, and a real-world clinical cohort (XH). The authors report associations between sleep-age gaps and longitudinal health outcomes, including mortality and incident diseases. They further investigate organ specificity by comparing matched and non-matched age gaps and report improvements over conventional age-regression baselines.

Critical strengths
  1. The expert-curated registry, evidence identifiers, deterministic rendering, and trace checks provide a structured mechanism for verifying the fidelity of input evidence.

  2. Evaluation across three cohorts and multiple longitudinal outcomes provides broader evidence than a single-cohort, cross-sectional study. Freezing the age representations before longitudinal outcome analysis is a positive design choice that helps limit direct outcome leakage.

  3. Comparing matched and non-matched organ-age gaps on the same participants is a useful approach to testing whether the representations capture outcome-specific information. The paper includes conventional age-regression baselines and alternative LLM backbones, providing useful reference points for assessing the proposed method.

Critical weaknesses
  1. The estimated ages are not validated against independent measures of organ-specific biological age. Associations with future outcomes establish prognostic relevance but do not necessarily demonstrate that the estimates measure biological aging or identify organ-specific aging processes.

  2. The conventional age-regression baselines use a common 18-variable subset, whereas the proposed framework exploits broader clinical information. Comparisons against conventional models using the same integrated inputs are needed to isolate the contribution of the LLM framework.

  3. The audit is evaluated against a model-based reference rather than independent clinical expert annotations. Its reported sensitivity of 64.9% also indicates that a substantial proportion of defined inconsistencies may remain undetected. Moreover, detecting contradictions does not establish that generated physiological explanations are evidence-supported.

  4. The matched-gap improvement in C-index is 0.014, compared with 0.006 for non-matched gaps, yielding a mean advantage of only 0.008. Although the advantage is positive across the reported endpoints, uncertainty estimates and formal paired comparisons are important for establishing its statistical and practical significance.

  5. The overall sleep-age estimates correlate strongly with chronological age (reported correlations of 0.934–0.978). Although the authors conduct adjusted analyses, further evidence is needed to establish how much independent, clinically meaningful information remains after accounting for age and baseline morbidity.

  6. The XH cohort contains 426 participants, limiting confidence in endpoint-specific findings, particularly where event counts are low. Independent external validation would strengthen the generalization claims.

  7. The respective contributions of evidence orchestration, registry-guided rendering, and consistency auditing are not fully isolated. Ablation studies would clarify whether the proposed multi-agent architecture provides meaningful advantages over simpler alternatives.

Questions for authors
  1. What independent evidence supports interpreting the estimated ages as biological aging representations rather than LLM-generated summaries of chronological age, comorbidities, and risk factors?

  2. Can the authors compare the proposed framework with conventional machine-learning models using the same integrated clinical inputs to isolate the contribution of the LLM?

  3. How do the authors rule out the possibility that organ-matched gaps perform better because the input profiles already contain organ-specific diagnoses, symptoms, medications, or risk factors?

  4. Why was the consistency audit evaluated against a model-based reference rather than independent clinical expert annotations? Can the authors report clinician-adjudicated sensitivity and specificity?

  5. Are matched-versus-non-matched C-index differences statistically significant after accounting for correlated predictions, repeated cross-validation, and multiple comparisons?

  6. How stable are the estimated ages and organ rankings across repeated generations, prompt perturbations, and alternative evidence orderings?

  7. Can the authors provide independent clinical validation, particularly for the organ-specific outcomes evaluated in the relatively small XH cohort?

  8. What are the individual contributions of the orchestration, deterministic rendering, and consistency-audit components? Would simpler alternatives yield comparable performance?

5 comments
AuthorsOpenReview ↗

Response to Reviewer kkA3 (Part 1 of 3)

We thank the reviewer for the thoughtful and detailed assessment. Your suggestions have substantially improved the rigor of the evaluation and the presentation of our findings. We have clarified the interpretation of the age representations and strengthened the organ-matching analyses, clinical annotation, and XH validation.

W1/W5/Q1: Aging-related information beyond chronological age. Larger overall sleep gaps tracked frailty in SHHS and MrOS and clinical deficits in XH after age adjustment (standardized coefficients 0.240, 0.183 and 0.368; Figure E.1). In XH, the HR for first new diagnosis was 1.059 per additional year of gap (95% CI 1.020–1.100) after adjustment for age and baseline diagnosis burden (Figure 4a). Matched organ gaps outperformed non-matched gaps on average in each cohort (Figure I.2). Independent organ biomarkers would extend this outcome evidence to biological calibration (Section 6.1).

W2/Q2: Age-regression baselines. We retrained eight supervised age models using 117, 158 and 118 variables in SHHS, MrOS and XH, respectively (Table E.1). Across 46 organ outcomes, mean held-out C-index was 0.615 for the matched LLM gap versus 0.557 for MLP, the best baseline (paired difference 0.0578, 95% CI 0.0328–0.0815; max-t adjusted P=0.0005P = 0.0005). Each cohort used healthy-reference training and nested cross-validation (five outer, three inner folds) for tuning and residual-gap calibration (Figure 6, Table C.1).

W3/Q4: Audit reference and detection performance. Two clinical experts jointly annotated all 426 XH outputs for the two prespecified contradictions. Against these clinical expert annotations, the Consistency Audit Agent detected 32/37 positives (sensitivity 86.5%, 95% CI 72.0–94.1%; specificity 97.4%, 95% CI 95.3–98.6%). Its state check identified 4/10 state contradictions as such (Section L, Figure L.1). Clinical support for individual explanations is discussed separately in Section 6.2.

W4/Q5: Paired inference for organ correspondence. We added 2,000 participant-paired bootstrap resamples, sharing each person's weight across outcomes and candidate gaps. Mean matched advantages were 0.011 (95% CI 0.005–0.019), 0.018 (0.010–0.027), and 0.057 (0.039–0.073) in SHHS, MrOS, and XH. Max-t adjustment gave P=0.006P = 0.006, P<0.001P < 0.001, and P<0.001P < 0.001. Figure I.2 reports simultaneous intervals across the three cohort means. Section I describes the separate multiplicity adjustment across 46 endpoint tests. The revised methods specify one shared five-fold split and conditional uncertainty for its held-out predictions. The absolute advantage was modest in SHHS, but the mean favored the matched gap in all three cohorts.

Q3: Organ matching after clinical adjustment. All eleven gaps derive from the same patient report, so matched and non-matched gaps reflect the same medical history. We evaluated four clinical Cox models (A–D), with progressively more risk factors, to compare the incremental discrimination from matched versus non-matched gaps (Table H.1). Each endpoint used identical participants and folds for all comparisons. Clinical model D included age, sex, BMI, systolic pressure, smoking, alcohol, hypertension and diabetes. Across 46 outcomes, mean C-index was 0.7159 with the matched gap versus 0.7019 with the ten alternatives on average (paired difference 0.0140, 95% CI 0.0094–0.0187).

Q6: Repeated predictions and evidence order. Keeping report order improved repeatability in XH (426 participants, four runs per condition, prompt wording fixed). With original versus shuffled order, overall sleep-gap ICC was 0.687 versus 0.562 and mean organ-gap ICC was 0.560 versus 0.472. Within-patient organ rankings showed moderate agreement (mean Spearman correlation 0.573 versus 0.530; Table F.1 and Section F).

W6/Q7: Validation in XH. We added five-fold internal cross-validation in XH, keeping the LLM age estimates fixed. For each outcome, Cox models were fitted on four folds and evaluated on the held-out fifth. Across 20 organ outcomes, mean held-out C-index was 0.666, close to the in-sample value of 0.668 (Table F.1, panel B). Section 6.1 discusses future validation with independent organ-level biomarkers, and Section 6.2 addresses replication of XH organ-specific associations in independent cohorts.

AuthorsOpenReview ↗

Response to Reviewer kkA3 (Part 2 of 3)

W7/Q8: Roles of the pipeline components. We clarified the roles of the pipeline components in Section 3. The Evidence Orchestration Agent selects nonredundant evidence and assigns source-linked Evidence IDs to six report sections. Compared with directly passing measurements to an estimator, this produces an inspectable record of the evidence selected for each patient. Registry-guided rendering and trace checks then verify each reported value and unit against its source (Figures 1–2). We also re-evaluated the Consistency Audit Agent against two clinical experts' joint annotations of two prespecified contradictions in all 426 XH outputs. The agent detected 32/37 positives, with 86.5% sensitivity and 97.4% specificity, without altering age estimates (Figure L.1).

Manuscript locations

The locations below refer to the revised manuscript PDF, where revisions are highlighted in blue.

  • Section 5.4, lines 434–443 (p. 9); Appendix B.6, lines 1009–1016 (p. 19); Figure E.1, lines 1434–1439 (p. 27) report age-adjusted associations with frailty and clinical-deficit burden.

  • Appendix B.5, lines 943–984 (pp. 18–19); Table C.1, lines 1272–1284 (p. 24); Table E.1, lines 1407–1417 (p. 27) document the supervised benchmark, cohort inputs, and held-out discrimination.

  • Section 4.4, lines 303–308 (p. 6); Appendix L, lines 1856–1867 (p. 35) describe the clinical annotation reference and audit detection performance (Figure L.1).

  • Appendix I, lines 1703–1712 (p. 32); Figure I.2, lines 1799–1804 (p. 34) specify participant-paired resampling, multiplicity adjustment, and uncertainty in organ correspondence.

  • Appendix B.7, lines 1097–1111 (p. 21); Appendix H, Table H.1, lines 1674–1700 (p. 32) compare matched and non-matched gaps under clinical models A–D.

  • Appendix F, lines 1552–1557 (p. 29); Table F.1, lines 1567–1593 (p. 30) report evidence-order repeatability, organ-rank agreement, and five-fold internal validation in XH.

  • Section 6.1, lines 465–479 (p. 9); Section 6.2, lines 523–529 (p. 10) discuss biological interpretation, independent biomarkers, and future cohort validation.

  • Section 3, lines 195–230 (pp. 4–5) clarifies the roles of orchestration, deterministic rendering, trace checks, and consistency auditing.

Tables cited in this response

The tables below reproduce the corresponding manuscript tables. Table numbers are retained for cross-reference.

Table E.1. Cohort-specific inputs and healthy references for supervised age models

Counts describe the full cohort before endpoint-specific exclusions.

CohortParticipantsHealthy referenceInput representation
SHHS5,7932,415117 variables, including sleep measures, demographics, lifestyle, and medical history
MrOS2,904734158 variables, including sleep measures, demographics, lifestyle, medical history, and medication use
XH426124118 variables, including sleep measures, demographics, anthropometry, laboratory tests, and medical history

Chronological age was excluded from predictors. Counts precede feature filtering. Methods are described in Section B.5.

Table C.1. Summary of held-out gap-only Harrell C-index across the same 46 organ outcomes

Supervised models use residual gaps and the LLM uses its endpoint-matched gap, as defined in Section B.5.

RepresentationMean C-indexMinimumMaximum
LLM: endpoint-matched gap0.6150.4460.846
Random forest0.5240.2250.662
Extra-trees0.5320.3010.717
Histogram boosting0.5190.3950.624
MLP0.5570.3940.808
XGBoost0.5170.3500.695
LightGBM0.5280.3590.650
CatBoost0.5530.3640.709
TabM0.5430.3320.664
AuthorsOpenReview ↗

Response to Reviewer kkA3 (Part 3 of 3)

Table H.1. Clinical adjustment and organ correspondence across 46 organ outcomes

ModelClinical C-index+ Matched gap C-index+ Non-matched gap, mean C-indexMatched gain ΔC\Delta CMatching advantage ΔC\Delta C
A0.6639 (0.6471–0.6807)0.7076 (0.6926–0.7225)0.6758 (0.6615–0.6899)+0.0437 (0.0286–0.0579)+0.0318 (0.0231–0.0399)
B0.6806 (0.6649–0.6957)0.7072 (0.6928–0.7209)0.6854 (0.6713–0.6991)+0.0266 (0.0178–0.0355)+0.0218 (0.0154–0.0279)
C0.6823 (0.6667–0.6965)0.7048 (0.6900–0.7185)0.6849 (0.6700–0.6984)+0.0225 (0.0157–0.0294)+0.0199 (0.0149–0.0251)
D0.7000 (0.6855–0.7139)0.7159 (0.7023–0.7284)0.7019 (0.6881–0.7151)+0.0159 (0.0103–0.0221)+0.0140 (0.0094–0.0187)

Clinical models. A, chronological age and sex. B, A plus BMI and systolic blood pressure. C, B plus smoking and alcohol use. D, C plus hypertension and diabetes history. Sex was omitted when invariant. These covariates were identical across endpoints, without substituting organ-specific histories. Evaluation. Entries average equally over the same 46 organ outcomes as Figure 6 (10 SHHS, 16 MrOS, and 20 XH). All-cause death and the XH first-new-diagnosis composite were analyzed separately. Each endpoint used identical participants and five folds for all models. Each gap was added separately. The non-matched comparator averages the C-indices of the ten alternative gaps. Matched gain is the matched-gap model minus the clinical model. Matching advantage is the matched-gap model minus the non-matched mean. Differences were calculated before rounding. Uncertainty. Parentheses give pointwise 95% percentile intervals from 2,000 paired participant bootstrap draws, retaining correlations across outcomes within each cohort and combining cohorts independently.

Table F.1. Age-gap repeatability and outcome discrimination in XH (n=426)

A. Pairwise agreement of age gaps across four runs per condition

Age gapOriginal orderShuffled order
Overall sleep0.687±0.0240.687 \pm 0.0240.562±0.0430.562 \pm 0.043
Cardiovascular0.699±0.0140.699 \pm 0.0140.609±0.0270.609 \pm 0.027
Neurological0.571±0.0270.571 \pm 0.0270.469±0.0270.469 \pm 0.027
Respiratory0.613±0.0350.613 \pm 0.0350.558±0.0580.558 \pm 0.058
Metabolic0.658±0.0230.658 \pm 0.0230.571±0.0180.571 \pm 0.018
Skeletal0.483±0.0240.483 \pm 0.0240.381±0.0510.381 \pm 0.051
Hepatic0.594±0.0240.594 \pm 0.0240.518±0.0390.518 \pm 0.039
Renal0.594±0.0200.594 \pm 0.0200.545±0.0240.545 \pm 0.024
Immune0.452±0.0270.452 \pm 0.0270.324±0.0610.324 \pm 0.061
Urinary0.498±0.0470.498 \pm 0.0470.395±0.0260.395 \pm 0.026
Digestive0.435±0.0430.435 \pm 0.0430.353±0.0260.353 \pm 0.026
Mean across 10 organs0.560±0.0150.560 \pm 0.0150.472±0.0310.472 \pm 0.031

B. Outcome discrimination across 20 incident organ outcomes

Prediction setsNumber of setsMean C-index
Reference predictions, in-sample10.668
Reference predictions, five-fold held-out10.666
Reference group + four original-order runs50.674±0.0150.674 \pm 0.015
Reference group + four shuffled runs50.646±0.0140.646 \pm 0.014

A. The reference group is excluded. Each value is the mean ± sample SD of six pairwise intraclass correlations for single-prediction absolute agreement [ICC(A,1)]. The ten organs are weighted equally within each pair; overall sleep is reported separately. B. Both reference rows use the same frozen Llama 3.3-70B predictions. In the first row, each gap-only Cox model was fitted and scored on all eligible participants for its outcome (Figure 7c). All other rows use the same five folds: models were fitted on four folds and scored on the fifth, pooling comparable pairs within test folds. Each outcome contributes equally. Five-set summaries report the mean ± sample SD across sets, including the shared reference.

Reviewer_B8hhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a framework for estimating overall and organ specific biological ages from integrated sleep profiles. An expert-curated evidence registry, an orchestration agent, and deterministic source-traceable reports. A pretrained LLM then generates one overall sleep age and ten organ-specific ages, while a separate audit agent screens the explanations for contradictions. The authors evaluate the resulting age gaps across SHHS, MrOS, and XH, reporting associations with mortality and incident disease, organ–outcome correspondence, and improvements over selected age-regression and clinical baselines.

Critical strengths
  • The study evaluates the generated representations across three cohorts with different populations and available clinical information. Associations with future outcomes provide more meaningful validation than agreement with chronological age alone.
  • Comparing matched and non-matched age gaps on identical participants and cross-validation folds is a useful way to assess whether the organ-specific outputs contain organ-specific prognostic information.
  • The paper provides complete agent prompts and distinguishes input traceability from output consistency. Predictions are reportedly frozen before outcome analysis, and the evaluation includes covariate adjustment and multiple-testing correction.
Critical weaknesses
  • Limited methodological novelty. Prior work already estimates overall and organ-specific ages using LLMs. The benefit of the orchestration agent is unclear because the registry defines routing rules and deterministic software renders and validates inputs. No ablation isolates its contribution.
  • Unequal input across methods Conventional regressors receive only 18 shared variables, while the LLM uses richer profiles. The strongest XH clinical baseline includes only age, sex, and diagnosis count.
  • Missing evidence robustness For a new cohort in practice, many of these evidence currently using might be missing. The paper does not systematically examine how missing variables affect organ-age estimates, explanations, or downstream performance.
  • Uncertain results Several XH outcomes have few events and wide C-index confidence intervals, limiting confidence in individual improvements and aggregate performance claims.
Questions for authors
  1. How does the authors see the orchestration agent compared to deterministic routing or direct prompting?

  2. Can the authors compare against stronger baseline models using the same available features, and potentially compare or discuss the implication of the disease risk prediction using biological age compare to PSG foundation models [1, 2]

  3. Can sleep-only, clinical-only, and combined-input analyses establish the contribution of sleep? Can clinicians independently validate the explanations and audit decisions?

[1] Shuai, Zitao, et al. "OSF: On pre-training and scaling of sleep foundation models." arXiv preprint arXiv:2603.00190 (2026). [2] Thapa, Rahul, et al. "SleepFM: multi-modal representation learning for sleep across brain activity, ECG and respiratory signals." arXiv preprint arXiv:2405.17766 (2024).

6 comments
AuthorsOpenReview ↗

Response to Reviewer B8hh (Part 1 of 3)

We sincerely thank the reviewer for the constructive suggestions, which have substantially improved the manuscript's quality and clarity. We strengthened the baseline comparisons, clarified our relationship to PSG foundation models, and expanded the statistical evaluation in XH.

W1/Q1: Evidence orchestration. Building on prior LLM age estimation, we trace each estimator input to its source and test whether prespecified organ gaps preferentially discriminate corresponding outcomes across three cohorts (Figures 2 and 5). The registry defines eligible evidence. The orchestration agent assigns each patient's nonredundant evidence IDs to six report sections before rendering. Rules could also produce a traceable plan. Direct prompting could bypass it, but would need a separate record to show which evidence reached the estimator. With this plan, a reader can trace each reported measurement back to its source field and verify its value and unit.

W2/Q2a: Supervised and clinical baselines. Eight supervised regressors use 117, 158 and 118 variables in SHHS, MrOS and XH, respectively, with chronological age as the target (Table E.1). We also added TabM (ICLR 2025), a recent tabular deep-learning model, and CatBoost, a strong gradient-boosted tree baseline (Section B.5). Healthy-reference training, tuning and residual calibration use nested cross-validation (five outer folds, three inner folds). Across all 46 eligible organ outcomes in SHHS, MrOS and XH, mean held-out gap-only C-index was 0.615 for the LLM and 0.557 for MLP, the strongest baseline. The paired difference was 0.0578 (95% CI 0.0328–0.0815, adjusted P=0.0005P = 0.0005; Figure 6, Table C.1). Table H.1 evaluates four clinical baselines on the same outcomes. With age, sex, BMI, systolic pressure, smoking, alcohol, hypertension and diabetes, adding the matched gap raised mean C-index from 0.7000 to 0.7159 (gain 0.0159, 95% CI 0.0103–0.0221).

Q2b: Relation to PSG foundation models. Following your suggestion, we now cite and discuss OSF and SleepFM in Section 2, highlighting their strengths and clarifying the scope of our report-based approach. OSF and SleepFM learn transferable representations directly from raw PSG. Our pretrained LLM estimates ages by zero-shot inference from reported clinical measurements, without cohort-specific fine-tuning or labeled examples. We then fit outcome models to evaluate the resulting gaps. Using reported measurements avoids requiring identical waveform channels across sites (Section 2).

W3: Handling unavailable evidence. The registry records unavailable variables separately from observed findings (Section O). Available variable sets differ across cohorts (Table E.1), as do populations and outcome ascertainment. This leaves the effect of missing inputs unresolved. Section 6.2 proposes comparing sleep-only, clinical-only and combined profiles in the same patients, using a common measurement set across centers. This would assess changes in decoded ages, explanations and outcome discrimination.

W4: Uncertainty in XH. Figure C.2 shows the C-index interval for each XH outcome. We also evaluated aggregate performance across all 20 eligible XH organ outcomes using 2,000 participant-paired bootstrap resamples. Resampling preserved each person's contributions across outcomes while keeping held-out predictions fixed. The matched gap exceeded the mean of the ten alternatives by 0.057 in mean held-out C-index (95% CI 0.039–0.073, max-t adjusted P<0.001P < 0.001; Figure I.2). These results support the prognostic value of organ matching across XH outcomes.

AuthorsOpenReview ↗

Response to Reviewer B8hh (Part 2 of 3)

Q3a: Contribution of sleep measurements. We appreciate this suggestion and agree that comparing sleep-only, clinical-only and combined inputs in the same patients would clarify what the overnight measurements add. Our study asks what aging signatures can be decoded from each integrated sleep profile as a whole. To test their prognostic value, we compared four clinical-only models with and without the endpoint-matched age gap, using the same participants and folds for each comparison. Across 46 organ outcomes, adding the gap raised mean held-out C-index by 0.0437, 0.0266, 0.0225 and 0.0159 for models A–D, respectively (Table H.1).

To see how individual sleep measurements related to the LLM's overall sleep age, we fitted a linear surrogate in each cohort (Figure E.2). With other inputs held fixed, one additional hour of sleep corresponded to estimates lower by 0.31 years in SHHS, 0.13 in MrOS and 0.09 in XH. The XH interval included zero. For an apnea–hypopnea index higher by 10 events per hour, the corresponding age differences were +0.16, +0.15 and +0.23 years.

Q3b: Clinical review of explanations and audit decisions. Two clinical experts jointly annotated two prespecified contradictions in all 426 XH outputs. Against these clinical expert annotations, the Consistency Audit Agent referred 32 of 37 reference-positive outputs and 10 of 389 negatives (Figure L.1). Reviewing each physiological claim against the patient record is a separate clinical task (Section 6.2).

Manuscript locations

The locations below refer to the revised manuscript PDF, where revisions are highlighted in blue.

  • Section 2, lines 125–137 (p. 3) discusses OSF and SleepFM and the scope of report-based age estimation.

  • Section 3, lines 195–212 (p. 4); Appendix O, lines 2470–2479 (p. 46) explain evidence selection, trace validation, and the handling of unavailable variables.

  • Appendix B.5, lines 943–984 (pp. 18–19) details the eight supervised baselines, healthy-reference training, nested cross-validation, and paired comparisons.

  • Table C.1, lines 1272–1284 (p. 24); Table E.1, lines 1407–1417 (p. 27) report baseline discrimination, variable counts, and healthy-reference sample sizes.

  • Section 5.4, lines 434–443 (p. 9); Appendix H, Table H.1, lines 1674–1700 (p. 32) report incremental discrimination beyond four clinical models and matched versus non-matched gap comparisons.

  • Appendix E.1, lines 1441–1475 (pp. 27–28) describes the surrogate models and clinical-unit associations with decoded ages (Figure E.2, p. 29).

  • Figure C.2, lines 1267–1269 (p. 24); Appendix I, lines 1703–1712 (p. 32) report XH endpoint intervals and paired statistical analysis across outcomes (Figure I.2, p. 34).

  • Section 4.4, lines 303–308 (p. 6); Appendix L, lines 1856–1867 (p. 35) describe joint clinical expert annotations and the evaluation of audit decisions (Figure L.1).

  • Section 6.2, lines 523–529 (p. 10) outlines comparisons of sleep-only, clinical-only, and combined inputs and further cohort validation.

Tables cited in this response

The tables below reproduce the corresponding manuscript tables. Table numbers are retained for cross-reference.

Table E.1. Cohort-specific inputs and healthy references for supervised age models

Counts describe the full cohort before endpoint-specific exclusions.

CohortParticipantsHealthy referenceInput representation
SHHS5,7932,415117 variables, including sleep measures, demographics, lifestyle, and medical history
MrOS2,904734158 variables, including sleep measures, demographics, lifestyle, medical history, and medication use
XH426124118 variables, including sleep measures, demographics, anthropometry, laboratory tests, and medical history

Chronological age was excluded from predictors. Counts precede feature filtering. Methods are described in Section B.5.

Table C.1. Summary of held-out gap-only Harrell C-index across the same 46 organ outcomes

Supervised models use residual gaps and the LLM uses its endpoint-matched gap, as defined in Section B.5.

RepresentationMean C-indexMinimumMaximum
LLM: endpoint-matched gap0.6150.4460.846
Random forest0.5240.2250.662
Extra-trees0.5320.3010.717
Histogram boosting0.5190.3950.624
MLP0.5570.3940.808
XGBoost0.5170.3500.695
LightGBM0.5280.3590.650
CatBoost0.5530.3640.709
TabM0.5430.3320.664
AuthorsOpenReview ↗

Response to Reviewer B8hh (Part 3 of 3)

Table H.1. Clinical adjustment and organ correspondence across 46 organ outcomes

ModelClinical C-index+ Matched gap C-index+ Non-matched gap, mean C-indexMatched gain ΔC\Delta CMatching advantage ΔC\Delta C
A0.6639 (0.6471–0.6807)0.7076 (0.6926–0.7225)0.6758 (0.6615–0.6899)+0.0437 (0.0286–0.0579)+0.0318 (0.0231–0.0399)
B0.6806 (0.6649–0.6957)0.7072 (0.6928–0.7209)0.6854 (0.6713–0.6991)+0.0266 (0.0178–0.0355)+0.0218 (0.0154–0.0279)
C0.6823 (0.6667–0.6965)0.7048 (0.6900–0.7185)0.6849 (0.6700–0.6984)+0.0225 (0.0157–0.0294)+0.0199 (0.0149–0.0251)
D0.7000 (0.6855–0.7139)0.7159 (0.7023–0.7284)0.7019 (0.6881–0.7151)+0.0159 (0.0103–0.0221)+0.0140 (0.0094–0.0187)

Clinical models. A, chronological age and sex. B, A plus BMI and systolic blood pressure. C, B plus smoking and alcohol use. D, C plus hypertension and diabetes history. Sex was omitted when invariant. These covariates were identical across endpoints, without substituting organ-specific histories. Evaluation. Entries average equally over the same 46 organ outcomes as Figure 6 (10 SHHS, 16 MrOS, and 20 XH). All-cause death and the XH first-new-diagnosis composite were analyzed separately. Each endpoint used identical participants and five folds for all models. Each gap was added separately. The non-matched comparator averages the C-indices of the ten alternative gaps. Matched gain is the matched-gap model minus the clinical model. Matching advantage is the matched-gap model minus the non-matched mean. Differences were calculated before rounding. Uncertainty. Parentheses give pointwise 95% percentile intervals from 2,000 paired participant bootstrap draws, retaining correlations across outcomes within each cohort and combining cohorts independently.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Comments

We sincerely thank all three reviewers for their careful reading and constructive suggestions. Their comments have substantially improved the clarity, rigor, and presentation of the manuscript. Revisions are highlighted in blue in the manuscript.

Summary of revisions

  • Stronger supervised baselines. We compared eight age regressors, including TabM (ICLR 2025) and CatBoost, using variable sets aligned with the LLM inputs. Across 46 organ outcomes, mean held-out C-index was 0.615 for the LLM gap and 0.557 for MLP, the strongest supervised baseline (Figure 6, Table C.1).

  • Organ matching and incremental prognostic value. Participant-paired bootstrap comparisons supported the matched-gap advantage in all three cohorts after multiple-comparison correction. Adding the matched gap improved mean held-out discrimination for all four clinical models across 46 organ outcomes (Figures 5 and I.2, Table H.1).

  • Interpretation and repeatability of age estimates. We added analyses relating the overall gap to frailty and clinical deficits, and clinical and sleep measurements to decoded ages. Repeated XH estimates quantify age and organ-ranking stability under evidence reordering (Figures E.1 and E.2, Table F.1).

  • Five-fold validation in XH. We added five-fold internal cross-validation of the outcome models in XH. Mean held-out C-index across 20 organ outcomes was 0.666, compared with 0.668 in-sample (Table F.1).

  • Longitudinal analysis. We corrected SHHS baseline cardiovascular history and event dating, and pooled adjusted hazard ratios across five imputations in SHHS and MrOS (Section B.7, Figures 4 and C.1).

  • Audit evaluation against clinical expert annotations. Two clinical experts jointly annotated all 426 XH outputs for two prespecified contradictions. Against these annotations, the Consistency Audit Agent achieved 86.5% sensitivity and 97.4% specificity (Figure L.1). We also documented the models and inference settings for each agent (Table B.2).

  • Related work, agent roles, and future validation. We expanded the discussion of OSF and SleepFM and clarified the scope of report-based age estimation (Section 2). We clarified how evidence orchestration organizes inputs and enables each reported measurement to be traced to its source (Section 3). Sections 6.1–6.2 outline further validation with independent cohorts and organ biomarkers.

AuthorsOpenReview ↗

Final summary of the revised evidence

Dear Area Chair,

Thank you for considering our revised paper. The reviewers' questions led us to strengthen the experiments and explain the method more clearly. As the discussion closes, we would like to draw your attention to the results we consider most important for the decision.

We set out to make heterogeneous clinical and sleep records useful for organ-resolved aging analysis while keeping the evidence supplied to the estimator inspectable. The framework produces one overall sleep age and ten organ-resolved ages, with a separate check of the generated explanations. We evaluated it in 9,123 participants across SHHS, MrOS, and XH. Three results make the strongest case for the paper:

  • Clinically adjusted prognostic value and organ correspondence. Across 46 organ outcomes, adding the matched gap to clinical model D increased mean held-out C-index from 0.7000 to 0.7159. The gain was 0.0159 (95% CI 0.0103–0.0221). Matched gaps also exceeded the non-matched mean by 0.0140 (0.0094–0.0187) under the same adjustment, participants, and folds (Table H.1, p. 32). These aggregate results support both the added information in the gaps and the value of matching their organ identity to the outcome.
  • Prospective associations across heterogeneous cohorts. After conventional-risk and medical-history adjustment, overall sleep gaps predicted mortality in SHHS and MrOS, with HRs of 1.048 (1.025–1.073) and 1.030 (1.008–1.052) per year. The XH first-new-diagnosis association remained after adjustment for chronological age and baseline diagnosis burden: HR 1.059 (1.020–1.100) (Sections 5.1–5.2, p. 7).
  • Traceable evidence and expert-validated explanation screening. Source-linked rendering makes estimator inputs inspectable across cohort field sets (Figure O.1, p. 47). The consistency audit captured 32/37 expert-annotated positive outputs while referring 42/426 outputs (9.9%), with 97.4% specificity for the two specified contradiction types (Figure L.1, p. 35). It concentrates problematic explanations into a small review set.

In response to the reviews, we also added eight supervised age baselines, participant-paired inference, five-fold XH evaluation, repeated predictions, and demographic ablation of age/gap surrogates. Removing demographics changed gap-reconstruction MAE by less than 0.007 years across all six cohort–target comparisons (Figure E.2, p. 28). This also helps us interpret the gaps: the clinical and sleep measurements reconstruct them with nearly unchanged fidelity.

We believe the paper makes a useful contribution by combining a reusable evidence workflow with organ-resolved predictions that add value after clinical adjustment. The clinical annotation study also gives a concrete measure of what the explanation audit achieves. We hope these results help you weigh the contribution of the revised paper, and we appreciate your time and consideration.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.