{"id":"840c62cc-0294-4c6b-9f32-fa092a38bec6","arxiv_id":"2605.30962","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"Sequence embeddings from diagnosis histories improve prediction of 93 of 131 incident disease blocks and event-free survival beyond age, sex, and comorbidity burden in large-scale hospital data.","lead":"This paper trains a contrastive transformer on sequences of hospital diagnoses and admission timings from 7.4 million Austrian patients to create patient embeddings. These embeddings modestly improve prediction of new diseases and event-free survival compared to standard comorbidity scores like Elixhauser.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.3","headline":"Modest AUC gains over Elixhauser do not isolate sequence/timing effects from richer comorbidity encoding","rationale":"The reader's weakest assumption directly matches the missing control; the small effect sizes make the distinction between 'better comorbidity encoding' and 'longitudinal structure' decisive for the headline interpretation.","tokens_in":1776,"tokens_out":321,"duration_ms":13361,"concrete_test":"Retrain the downstream predictors using a non-sequential baseline: replace the contrastive transformer with an MLP or mean-pool over the same diagnosis embeddings but with all admission order and time deltas removed; recompute the 131 AUCs and the event-free survival AUC. If the median lift falls below 0.002 or loses significance on >50 blocks, the sequence/timing contribution is not established.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim requires that the visit-level contrastive transformer extracts signal from diagnosis order and inter-admission intervals that is not already present in a cross-sectional comorbidity summary. The reported median AUC lift of 0.006 (93/131 blocks) and survival AUC lift of 0.004 are consistent with either (a) genuine longitudinal information or (b) the transformer simply learning a higher-capacity, data-driven encoding of the same past diagnoses that Elixhauser compresses into 30 binary indicators. No ablation that replaces the sequential encoder with a permutation-invariant aggregator (e.g., mean-pooled diagnosis embeddings or a set transformer) is described, so the source of the residual predictive power remains unidentified.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper trains a visit-level contrastive transformer on 13 years of Austrian inpatient records (7.4M patients) to produce patient-history embeddings from diagnosis sequences and inter-admission intervals. These embeddings are evaluated in a 1.7M-patient downstream cohort and reported to improve prediction of 93/131 incident ICD-10 disease-block outcomes over an Elixhauser comorbidity baseline (median AUC gain 0.006), with additional gains in event-free survival (AUC 0.726 vs 0.722) that are linked to the breadth, recency, and pace of prior morbidity accumulation.","tokens_in":1933,"tokens_out":513,"duration_ms":15210,"significance":"If the residual predictive signal is shown to originate from temporal ordering and timing rather than richer cross-sectional encoding of the same diagnoses, the result would demonstrate that sequence models can extract prognostic information beyond standard comorbidity indices in large-scale hospital data. The modest effect sizes and concentration in specific disease categories (mental, musculoskeletal, nervous, metabolic) limit immediate clinical translation but could inform targeted longitudinal risk modeling.","major_comments":[{"comment":"Abstract and implied Methods: the central claim that embeddings capture information 'beyond' the Elixhauser index requires an ablation that replaces the sequential contrastive transformer with a permutation-invariant aggregator (e.g., mean-pooled diagnosis embeddings or set transformer). Without this control, the reported median AUC lift of 0.006 cannot be attributed to sequence or timing rather than higher-capacity encoding of the identical past diagnoses.","section":"Abstract"},{"comment":"Abstract/Results: the modest median AUC gain (0.006) and survival AUC lift (0.004) are presented without reported confidence intervals, statistical tests for improvement, or assessment of calibration; given the sample size of 1.7M, even small gains may be statistically detectable yet clinically marginal, weakening the link to 'breadth, recency, and pace'.","section":"Abstract"}],"minor_comments":[{"comment":"Clarify how inter-admission timing is tokenized and whether the contrastive objective explicitly penalizes or rewards temporal order.","section":null},{"comment":"Specify the exact train/validation split between the embedding pre-training cohort and the 1.7M downstream cohort to rule out leakage.","section":null}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for their thoughtful review and constructive feedback. We address the major comments point-by-point below.","responses":[{"response":"We agree that demonstrating the specific contribution of sequential information requires an ablation against a permutation-invariant baseline. In the revised version, we will add this control experiment using mean-pooled embeddings of the same diagnosis representations, allowing direct comparison to isolate the effect of ordering and timing.","revision_made":"yes","referee_comment":"[Abstract] Abstract and implied Methods: the central claim that embeddings capture information 'beyond' the Elixhauser index requires an ablation that replaces the sequential contrastive transformer with a permutation-invariant aggregator (e.g., mean-pooled diagnosis embeddings or set transformer). Without this control, the reported median AUC lift of 0.006 cannot be attributed to sequence or timing rather than higher-capacity encoding of the identical past diagnoses."},{"response":"We will include bootstrap-derived confidence intervals for the AUC values and differences, along with p-values from appropriate statistical tests (e.g., DeLong's test for AUC comparison). We will also add calibration metrics and plots to the revised manuscript to provide a more complete evaluation of the model's performance.","revision_made":"yes","referee_comment":"[Abstract] Abstract/Results: the modest median AUC gain (0.006) and survival AUC lift (0.004) are presented without reported confidence intervals, statistical tests for improvement, or assessment of calibration; given the sample size of 1.7M, even small gains may be statistically detectable yet clinically marginal, weakening the link to 'breadth, recency, and pace'."}],"tokens_in":1469,"tokens_out":364,"duration_ms":28237,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The paper finds modest predictive lifts from visit-level contrastive transformer embeddings of longitudinal hospital diagnoses compared with the Elixhauser index. On 1.7 million patients the embeddings improve AUC for 93 of 131 incident disease-block outcomes with a median gain of 0.006, concentrated in mental, musculoskeletal, nervous-system, and metabolic blocks, and they produce a small survival AUC edge plus a residual-risk split that tracks 132-183 fewer event-free days.\n\nThe scale of the Austrian inpatient records (7.4 million patients, 13 years) and the concrete breakdown by disease category are the clearest contributions. The residual-risk survival analysis also gives a direct link between the embedding signal and breadth, recency, and pace of prior accumulation.\n\nThe central weakness is the missing ablation. Nothing in the reported work tests whether a permutation-invariant aggregator over the same diagnoses would recover similar gains; without that comparison the source of the extra signal stays ambiguous. The effect sizes themselves are small enough that even statistically reliable differences may not shift practice.\n\nThis work is for health-informatics groups that already use administrative data for risk modeling and want category-level diagnostics on where sequence information adds anything. A reader building or refining comorbidity-adjusted predictors will find the numbers and the residual-risk framing useful.\n\nThe paper engages the existing comorbidity literature directly and reports falsifiable counts, so it deserves a serious referee. I would send it for review and specifically request the sequence-versus-set ablation plus confidence intervals on the AUC differences.","headline":"Small AUC gains from a contrastive transformer on diagnosis sequences over Elixhauser, but no ablation isolates timing or order from richer comorbidity encoding.","tokens_in":2416,"tokens_out":379,"would_cite":false,"duration_ms":15299,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Longitudinal hospital diagnosis sequences contain predictive information beyond age, sex, and comorbidity burden.","keywords":["diagnosis sequences","contrastive transformer","comorbidity burden","longitudinal hospital data","disease prediction","event-free survival","disease accumulation"],"falsifier":"A model that adds the embeddings to a baseline already containing the Elixhauser index, age, and sex shows no AUC improvement, or randomizing the order of diagnoses within patient histories removes the observed gains.","tokens_in":2676,"feed_emoji":"📊","tokens_out":668,"duration_ms":22496,"temperature":0.7,"pith_summary":"The paper asks whether the timing, sequence, and pace of diagnoses in hospital records hold information not captured by standard cross-sectional comorbidity measures such as the Elixhauser index. It trains a visit-level contrastive transformer on 13 years of Austrian inpatient data covering millions of patients to produce embeddings that incorporate diagnosis order and inter-admission intervals. These embeddings yield modest AUC gains over comorbidity-only models for 93 of 131 incident disease-block outcomes, concentrated in mental, musculoskeletal, nervous system, and metabolic disorders. The embeddings also identify patients with shorter event-free survival, linking the added signal to the breadth, recency, and pace of prior disease accumulation.","feed_headline":"Diagnosis sequences add signal beyond comorbidity scores","feed_subtitle":"Transformer embeddings from Austrian hospital histories improve prediction for 93 of 131 outcomes and link to pace of accumulation.","key_machinery":"visit-level contrastive transformer that encodes diagnosis sequences and inter-admission timing into patient-history embeddings","core_discovery":"A visit-level contrastive transformer encodes diagnosis sequences and inter-admission timing into patient-history embeddings that improve prediction of 93 of 131 incident ICD-10 disease blocks over Elixhauser-based models, with the added signal concentrated in the breadth, recency, and pace of prior disease accumulation as measured by reduced event-free survival.","pith_inferences":["The approach could be applied to outpatient or claims data to test whether sequence effects persist outside inpatient settings.","Residual risk scores derived from embeddings might support targeted monitoring for patients showing rapid accumulation patterns.","Shuffling diagnosis order in retraining experiments would isolate the contribution of sequence versus simple count of conditions.","Similar embeddings could be compared across countries to examine whether accumulation pace varies by healthcare system."],"forward_implications":["Embeddings improve prediction for 93 of 131 incident disease blocks with a median AUC gain of 0.006.","Gains concentrate in mental, musculoskeletal, nervous system, and metabolic disorders.","Patients with high residual risk have 132-183 fewer event-free days over five years.","Event rates for high-residual-risk patients match those of low-residual-risk patients more than a decade older.","The embedding signal tracks the breadth, recency, and pace of prior disease accumulation."],"fun_headline_variants":["Diagnosis sequences improve prediction for 93 of 131 disease outcomes","Visit embeddings encode timing to improve over Elixhauser scores","Diagnosis pace signals lower event-free survival beyond burden","Embeddings from sequences link to breadth and recency of accumulation","Sequence models show pathways beyond comorbidity in hospital data"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The embeddings from the contrastive transformer capture information about diagnosis sequences and timing that is not already contained in age, sex, and the Elixhauser comorbidity index.","fun_headline_variants_meta":{"raw":{"variants":["Diagnosis sequences improve prediction for 93 of 131 disease outcomes","Visit embeddings encode timing to improve over Elixhauser scores","Diagnosis pace signals lower event-free survival beyond burden","Embeddings from sequences link to breadth and recency of accumulation","Sequence models show pathways beyond comorbidity in hospital data"]},"model":"grok-4.3","cost_usd":0.007321,"raw_usage":{"total_tokens":3386,"prompt_tokens":700,"num_sources_used":0,"completion_tokens":77,"cost_in_usd_ticks":73212000,"prompt_tokens_details":{"text_tokens":700,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2609,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":700,"tokens_out":77,"duration_ms":16832,"temperature":1.0,"reasoning_tokens":2609,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-28T20:13:11.211516+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A model that adds the embeddings to a baseline already containing the Elixhauser index, age, and sex shows no AUC improvement, or randomizing the order of diagnoses within patient histories removes the observed gains.","supporting_citations":[],"review_version":1}