acceptodds
Under review as a conference paper at ICLR 2027

MOSAIC: Mixture-Of-Specialty Adaptation for Individualized Clinical-prediction

Abstract

Electronic health record (EHR) foundation models share one set of parameters across all patients: a patient's history enters only through the input context, and adapting these models to a clinical setting requires updating millions of parameters. Yet clinical personalization has natural structure: specialties reason differently about the same findings, and patients within a specialty follow different trajectories. We present MOSAIC (Mixture-Of-Specialty Adaptation for Individualized Clinical-prediction), which captures both levels through composition over a frozen LLM rather than per-patient parameter updates. At the cohort level, evolutionary prompt optimization discovers specialty-specific clinical reasoning rules in natural language. At the patient level, micro-adapters are composed through trajectory-conditioned routing, yielding adapter weights that reflect each patient's disease history. On a frozen Qwen2.5-7B backbone with only 78K trainable parameters, MOSAIC exceeds 18 baselines on MIMIC-IV next-visit diagnosis prediction at four levels of diagnostic granularity and outperforms standard LoRA at a matched parameter budget. Cohort reasoning contributes mainly at coarse granularity and patient calibration mainly at fine granularity. The approach transfers to eICU, a multi-center critical-care database. MOSAIC also predicts across a broad diagnostic vocabulary and improves monotonically with patient history length, where deep EHR baselines plateau.

Primary area:
foundation or frontier models, including LLMs
Keywords:
electronic health records, clinical prediction, patient trajectory, large language models, parameter-efficient fine-tuning

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

MOSAIC combines specialty prompts with small trainable adapters. The backbone remains frozen. GEPA discovers natural language guidance for eight specialties. TinyLoRA adapters learn additional adjustments through reinforcement learning. Patient diagnosis frequencies determine the adapter mixture. Training updates only 78.4K parameters across all specialties.

The method predicts diagnoses at the next visit. Code F1 improves from 0.099 to 0.124 on MIMIC. The paper also reports results on eICU.

Critical strengths

The adapter composition is mathematically coherent. The reported improvements over the Table 1 baselines have paired bootstrap support. The parameter-matched comparisons are useful.

Critical weaknesses

The method combines several established techniques. Its novelty mainly lies in their clinical integration.

The experiments do not isolate the central personalization mechanism. Several evaluation details also remain unclear. M1 and M2 concern the paper's central contribution. M5 concerns whether the comparisons can be reliably interpreted. M3 and M4 mainly require stronger evidence or narrower claims. These concerns lead me towards weak rejection.

M1. Missing Routing Ablations

Table 2 does not isolate patient-dependent routing. There is no global TinyLoRA baseline. Uniform mixtures and hard routing are also missing. A global adapter could learn the same dataset statistics. Specialty routing might therefore be unnecessary. Existing LoRA comparisons do not resolve this question. They change the adapter parameterization alongside the proposed structure.

The granularity claim has another possible explanation. GEPA optimizes group F1, while adapters use a graded code-matching reward that gives the highest reward to exact matches. Their different objectives could explain their different improvements. The paper should separate objective effects from architectural effects.

These comparisons should match prompts, rewards, and training budgets. An adapter-only variant would also clarify GEPA's contribution.

M2. Unclear Diagnostic Metrics

Section 4.4 defines chapters using the first ICD letter. However, first letters do not define official ICD chapters. The letter H spans two official chapters. Conversely, A and B share one chapter. This affects metric interpretation and the adapter reward.

Twelve clinical groups are presented after 22 body systems. Their definitions and nesting relationships are not provided. This weakens the interpretation of granularity-specific improvements. CCSR can also assign codes to multiple categories. The paper should specify how these assignments are handled.

Code F1 measures three-character categories, not complete codes. This distinction should remain explicit throughout the paper.

M3. Temporal Personalization Is Not Established

Section 2.2 routes patients using chapter frequencies. Different chronological histories can share identical chapter frequencies. Such histories receive identical adapter mixtures. The backbone still receives chronological context. However, the router itself ignores order and recency.

Figure 2 also compares patients with different history lengths. This does not isolate the benefit of additional history. A recency-aware router would provide a useful comparison.

M4. Coverage Does Not Establish Rare Disease Accuracy

Figure 2 reports 585 distinct predicted codes. This measures output diversity, not accuracy on rare diagnoses. However, the paper claims accurate prediction across the diagnostic long tail. Aggregate F1 can improve through frequent diagnoses alone.

Also, new diagnoses should be separated from previously observed diagnoses. Figure 3 should report disease prevalence and uncertainty.

M5. Incomplete Evaluation Details

Several details are necessary to interpret the reported comparisons. These include encounter construction, cutoff selection, and eligible patient counts. The requested prediction count also needs justification. This count must be selected without future target information. F1 averaging and output thresholds also need specification.

Baseline training and tuning budgets are insufficiently described. Some baselines predict only 28 to 53 distinct codes. Threshold choices could contribute to this narrow coverage. These details matter when interpreting the claimed performance advantage.

The eICU transfer claim introduces another ambiguity. Training on eICU would demonstrate replication, rather than frozen transfer. The paper should identify every component adapted using eICU. These omissions prevent reliable interpretation.

Questions for authors

The suggested fixes corresponding to the concerns above are:

M1. Can you isolate routing benefits using matched controls, including a global TinyLoRA baseline, uniform mixtures, and hard routing? These comparisons should match prompts, rewards, and training budgets. An adapter-only variant would also clarify GEPA's contribution.

M2. Can you release validated mappings and correct affected evaluations?

M3. Can you compare routers while truncating the same patients' histories?

M4. Can you report precision and recall across diagnosis-frequency groups?

M5. Can you release complete protocols, including explicit eICU adaptation details?

2 comments

We thank the reviewer for a careful and specific review. Changes in the revised manuscript are in blue; the analyses below are in Appendix D (Table 7), E (Tables 8–9), F (Table 10), G (Table 11), H (Table 12) and I. Numbers are macro F1 (Chapter / CCSR / Clinical Group / Code).

M1: routing ablations

All adapter rows match MOSAIC's 78K budget, prompts, GRPO objective and reward; the global adapter is one 78K adapter trained on all patients, no routing; hard routing uses the stored specialty label's adapter; uniform mixture averages the eight adapters equally.

VariantChapterCCSRClin. GroupCode
MOSAIC0.2580.3080.3770.124
Global adapter0.2540.3050.3680.116
Hard routing0.2540.3010.3710.121
Uniform mixture0.2510.3040.3720.117
Prompts only0.2490.2960.3610.102
LoRA 78K, evolved prompts0.2550.2920.3450.105

MOSAIC exceeds every other row of the table at every granularity. No single specialist reaches the mixture at Clinical Group or Code (0.359 to 0.371 and 0.107 to 0.120), and the best is within 0.001 at Chapter and CCSR. With the objective held fixed, the gain from composing the adapters is architectural. Table 2's LoRA rows use MOSAIC's reward with a generic instruction; with the evolved prompts (last row) the 78K LoRA gains at most 0.014 and stays below MOSAIC everywhere.

M2: metric definitions and mappings

The paper's Chapter F1 groups 3-character categories by their first letter; re-scored by the official ICD-10-CM chapters, MOSAIC reaches 0.291 against 0.278 for the best baseline, and all 18 baselines stay below under both definitions. The 12 clinical groups are our own, not nested in the 22 CCSR body systems: unions of whole ICD-10-CM blocks by the clinical service that acts on the prediction, and categories outside them score as a residual Other label (31% of targets). A category whose codes span CCSR body systems takes the AHRQ default system of the majority of its codes. Code F1 scores the 3-character category, as Section 4.4 and the revised caption of Table 1 now state. Both mapping tables will be released.

M3: temporal information

The router uses only the chapter composition of the history and is order-free. We permuted each patient's event order, keeping events, routing weights and prompt unchanged.

ChapterCCSRClin. GroupCode
MOSAIC0.2580.3080.3770.124
Event order shuffled0.2290.2770.3370.096

The loss (0.028 to 0.039, paired difference in Appendix G) exceeds that from removing the adapters entirely, and its 95% interval excludes zero everywhere: the predictions depend on the ordered trajectory. Two routers fed truncated histories of the same patients, each applying the single adapter it weights highest:

RouterChapterCCSRClin. GroupCode
MOSAIC0.2580.3080.3770.124
Last ten diagnoses0.2580.3050.3770.122
Most recent encounter0.2580.3050.3750.120

Neither exceeds MOSAIC at any granularity: it already reads the ordered trajectory, so an explicit recency-aware router is unnecessary.

M4: long tail, prior versus new diagnoses

In three equal groups of the 357 categories by target frequency (74% / 18% / 8% of targets), MOSAIC's macro F1 is 0.138 / 0.120 / 0.113 on frequent / medium / rare codes, highest in every group against its two ablation points, the frozen backbone (0.116 / 0.095 / 0.086) and the 78K LoRA of Table 2 (0.109 / 0.107 / 0.094), and against TransformEHR (0.127 / 0.046 / 0.016), the strongest deep EHR baseline in Table 1; MOSAIC's widest margins are on medium and rare codes. Pooled precision / recall (Table 9): MOSAIC 0.179 / 0.175, 0.151 / 0.123, 0.118 / 0.117, balanced in each group; TransformEHR 0.199 / 0.188, 0.179 / 0.055, 0.148 / 0.012, precise but recovering only 5.5% and 1.2% of medium and rare targets. On targets split by whether they already appear in the record (43% prior, 57% new; Table 12), MOSAIC has the highest Clinical Group F1 on both (0.385, 0.224), above every baseline on new diagnoses.

M5: evaluation protocol

Section 4.2 and Appendix C state the full protocol. Every system is scored on its top three predictions, with no tuned threshold, macro F1 over a fixed 357-category vocabulary. The deep EHR baselines were retrained from scratch, following the original configuration where one is specified at a comparable scale. On eICU (Appendix C), prompts and adapters are refit with backbone, router (MIMIC-IV prototypes) and hyperparameters unchanged, ICU stays in unit-visit order, so Table 3 is a replication; paired tests put MOSAIC significantly above 16 of 18 baselines at Chapter, CCSR and Clinical Group and 18 of 18 at Code.

We hope these results address your concerns and would be glad to answer any further questions.

Dear authors,

Thank you for the efforts you have made in this rebuttal. However, the evidence for the paper’s central personalization claim remains modest. On balance, I am maintaining my weak-reject rating.

Wishing you well in the last stretch of reviews.

Reviewer_2BqvOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The authors propose MOSAIC, a electronic health record (EHR) model built for personalization across clinical specialties and patients. Experiments were conducted on MIMIC-IV and eICU on a next-visit diagnosis task. The manuscript includes several leading EHR foundation models as comparator models including MOTOR, ETHOS, and Delphi as well as several LLM based models. The core issue with this paper is on task validity: the authors take a patient history and prompt their model to predict the set of ICD-10 diagnosis codes that will be recorded in the next clinical encounter. Generally, these codes are discharge billing codes and are not necessarily new diagnoses, and a model can simply be learning to copy forward previous codes in the encounter. The authors report diagnoses of conditions like rectal cancer and soft tissue sarcoma in their results (Figure 3): these are conditions that are rarely, if ever established during acute care. The questionable task validity has downstream effects in metric choice which makes it seem like MOSAIC is more performant than it may actually be in clinically relevant tasks. Given this, I recommend clear rejection for this manuscript.

Critical strengths
  • Comparator models. The authors include a mix of leading EHR foundation models (ETHOS, Delphi, and MOTOR) as well as LLMs (Qwen2.5 variants, Gemma 4 etc.)
Critical weaknesses
  • Task validity. The main results of the paper are on next-visit diagnosis prediction (Table 1 and Table 3). ICD codes are discharge billing codes and do not necessarily reflect new diagnoses in the visit, but rather could be carried forward from prior clinical history. Several conditions that the authors report (Figure 3) including rectal cancer and soft tissue sarcoma require longitudinal workup and would almost never be diagnosed in an acute care setting like MIMIC-IV or eICU.
  • Lack of details on code harmonization. MIMIC-IV contains both ICD-9 and -10 codes, and the only mention of ICD-9 in the manuscript is under Section 4.2 in “(ICD-9 mapped to ICD-10 via AHRQ 3-character prefix mapping, 97.8% patient-level coverage).” The standard approach is to use a cross-walk (CMS’s General Equivalence Mappings), as prefix matching alone can cause collisions (e.g. diabetes and sepsis are both split across several prefixes in ICD-10). Given the task definition relies on good codes, the authors should ensure that this section is well specified.
Questions for authors

Please see the weakness section above.

2 comments

We thank the reviewer for pointing out the two concerns. We answer each below. Every change in the revised manuscript is marked in blue.

Task validity (first weakness)

Why ICD codes. ICD codes record both conditions diagnosed at an encounter and conditions carried from earlier care, so we report the two separately below. We use them for two reasons. First, availability: ICD codes are the only structured record of each patient's diagnoses over time that both MIMIC-IV and eICU provide. Second, comparability: every foundation model we compare against (MOTOR, ETHOS, Delphi, CLMBR and the others) is trained to predict these same codes, so using them lets Table 1 compare all 19 systems on one task. The target is the set of ICD-10 diagnoses recorded after the cutoff, in practice the patient's final recorded hospitalization, which begins a median of 27 days after the cutoff (Appendix C); the first three in source order are scored.

We hope these results address your concerns and would be glad to answer any further questions.

Prior diagnoses versus new diagnoses. We split the 10,025 scored target codes by whether the same category was already in the patient's record: 43% were prior diagnoses recurring in the target window and 57% were new diagnoses. We then computed the paper's metric, macro F1 at the Clinical Group level, separately on each part:

SystemPrior diagnosesNew diagnosesAll targets (Table 1)
MOSAIC0.3850.2240.377
Qwen2.5-7B, frozen0.3610.2070.349
TransformEHR0.3720.1940.340
CPLLM0.3620.1850.334
Copy forward (three most recurrent prior codes)0.3030.1610.290

The table (Table 12 in Appendix H of the revised manuscript) compares MOSAIC with the strongest baseline of each Table 1 block at this granularity: the frozen Qwen2.5-7B, which is also MOSAIC's backbone without training, TransformEHR among the deep EHR models and CPLLM among the fine-tuned LLMs, plus a copy-forward reference. MOSAIC has the highest F1 on both parts, and on new diagnoses it is above every one of the 18 baselines in Table 1. To test the concern that a model could simply carry prior codes forward, we added a copy-forward reference that outputs each patient's three prior codes that recur most often in the population (last row). It scores below MOSAIC on both parts, by 0.082 on prior and 0.063 on new diagnoses. MOSAIC's advantage therefore does not come from carrying conditions forward from the patient's history: for diagnoses that are new to the patient, it predicts the correct clinical group more accurately than every baseline.

Conditions that need longitudinal workup. In our test set, the rectal cancer and soft tissue sarcoma targets recur in patients whose records already carry a cancer or cancer-care code (all nine cases). The task is to predict every coded diagnosis of the next encounter, recurring or new, and Table 12 scores the two separately.

Code harmonization (second weakness)

We thank the reviewer for the suggestion to use the CMS General Equivalence Mappings (GEM) crosswalk. We ran it, and it gives essentially the same result as our mapping. The submission's phrase "AHRQ 3-character prefix mapping" described our lookup table by its keys and its use: it maps each ICD-9 code prefix to one ICD-10 3-character category and to that category's AHRQ CCSR body system. The ICD-10 entries of that table were themselves taken from the CMS GEM files, which Section 4.2 of the revised manuscript now states; the crosswalk the reviewer suggests applies the full GEM code by code instead of through our table. The mapping touches only the ICD-9 codes in patient histories; every target is a native ICD-10 code.

First, the two mappings largely agree. Across the 24,443 ICD-9 code occurrences in the test patients' histories, the ICD-10 category we assign is one of the GEM targets for that code in 99.7% of cases, and GEM's most frequent target in 66%; in almost all of the remaining cases it is another GEM target for the same code.

Second, model performance does not change. We re-ran MOSAIC with the GEM crosswalk:

Mapping of ICD-9 history codesChapterCCSRClin. GroupCode
Our mapping (paper)0.2580.3080.3770.124
CMS GEM crosswalk0.2560.3070.3770.126

No difference exceeds 0.002 at any granularity. The numbers reported in the paper, for MOSAIC and all 18 baselines, use our table, so every system is compared on one input; Appendix H of the revised manuscript reports both mappings.

I thank the authors for their response, updated manuscript, and additional experiments. Please see my comments below:

Task Validity. Thank you for your comments. While it is true other EHR foundation models (e.g. MOTOR) do use ICD diagnosis codes, they also 1. utilize non-acute setting, longitudinal data (such as MERIATIVE for MOTOR and the UK Biobank and longitudinal data for Delphi) and 2. Look at better designed tasks than simple next-visit diagnosis across several disease types (e.g. time-to-event for carefully selected tasks in MOTOR).

This task design and validity question is not a trivial one: EHRSHOT [1] showed that “CLMBR-T-base (blue) consistently outperforms the count-based GBM (red) at k=64k =64, but lags in higher label settings for the Assignment of New Diagnoses tasks” showing that these models can be beaten by simple baselines under slight shifts in settings. From the manuscript and the authors’ rebuttal, it does not seem like there has been sufficient care in dealing with this validity issue.

Prior vs New Diagnoses. Thank you for running the new experiments. While the results in the table show some promise, it shows a gap between prior diagnoses and new diagnosis across all methods (and is missing other FM-based methods)--I find the claim “MOSAIC's advantage therefore does not come from carrying conditions forward from the patient's history” too strong given the results (particularly given the gap between a frozen Qwen2.5 model and MOSAIC).

Code Harmonization. I thank the authors for running the GEM cross walk. The results look promising and I encourage the authors to use the GEM mappings in future submissions.

My main concern around task validity remains. Given the main thrust of the authors’ manuscript is in next-visit diagnosis, the concerns around validity I raised, the evaluation / metrics raised by other reviewers, and the potential fragility of EHR foundation models (see my comments on EHRSHOT), I am maintaining my initial review.

[1] Wornow, Michael, et al. "Ehrshot: An ehr benchmark for few-shot evaluation of foundation models." Advances in Neural Information Processing Systems 36 (2023): 67125-67137.

Reviewer_s5dYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces MOSAIC, a framework for predicting next-visit diagnoses from longitudinal electronic health records using a frozen large language model. The method combines specialty-specific prompts optimized through GEPA with eight TinyLoRA adapters trained using reinforcement learning. At inference, the adapters are combined according to the similarity between the patient’s diagnosis distribution and specialty profiles, while the dominant specialty’s prompt guides prediction. The framework uses approximately 78K trainable parameters on a frozen Qwen2.5-7B backbone. Experiments on MIMIC-IV and eICU show improvements over a broad range of baselines, with paired bootstrap tests supporting the main MIMIC-IV comparisons. Ablation results suggest that optimized prompts primarily improve predictions of broad diagnostic categories, while adapters provide larger gains at the diagnosis code level. However, further controlled comparisons are needed to establish the specific contribution of patient-dependent adapter routing.

Critical strengths
  1. The design is practical and easy to understand. Pairing inspectable, natural-language specialty guidance with very small adapters is an appealing recipe, and the matched-budget LoRA comparison (Table 2) is the right control for the parameter-efficiency claim.

  2. The evaluation is reasonably broad: two datasets, baselines spanning end-to-end EHR models, fine-tuned LLMs, and frozen LLMs at several scales, bootstrap confidence intervals, and paired tests on MIMIC-IV (Table 5).

  3. The granularity-specific decomposition is the most interesting empirical finding. Showing that prompts and adapters contribute at different levels of the ICD hierarchy, and that scaling LoRA to 128 to 256 times the budget recovers coarse but not fine performance, gives the paper a claim beyond “we beat the baselines.”

Critical weaknesses
  1. The novelty is in the composition, but the experiments do not isolate what the composition adds. GEPA, TinyLoRA, and GRPO are existing methods; the paper’s contribution is patient-specific routing over specialty adapters. Table 2 contains no (a) single shared TinyLoRA adapter at the same 78K budget, (b) hard routing to the dominant specialty’s adapter only, (c) uniform mixing of all eight adapters, or (d) adapters trained without the evolved prompts. Without these, the gains cannot be attributed to routing rather than to specialty-partitioned training or to TinyLoRA itself. It is also not stated whether the LoRA rows use the same evolved prompts, the same GRPO objective, and the same reward, so the matched-budget comparison may be confounded.

  2. There is a train/inference mismatch that is acknowledged but not tested. Each adapter is trained with its own specialty prompt active (Section 2.2), but at inference several adapters are mixed under a single dominant-specialty prompt (Section 2.3). The shared projection basis makes the mixture well-defined algebraically, but that does not show it is beneficial. Comparing the mixture against each individual adapter under the same prompt would address this directly.

  3. The “trajectory-conditioned” framing is stronger than the mechanism. The router is a cosine similarity over a chapter-frequency histogram of pre-cutoff diagnoses, which discards temporal order entirely. The paper’s language about evolving disease profiles and individual trajectories should be toned down, or a temporal routing variant should be evaluated.

  4. Reproducibility and protocol gaps affect the main conclusions. Missing or unclear: how encounters and “next visits” are constructed (especially in eICU, which is ICU-stay based); how many codes each model predicts and how prediction-set size or thresholds are chosen per baseline (this strongly affects set F1); how the deep EHR baselines were configured and tuned, and how their outputs were converted to diagnosis sets; adapter target modules and rank, optimizer, and learning rate; and whether eICU involves retraining prompts and adapters or direct transfer from MIMIC-IV. Results are from a single training run; bootstrap intervals over test patients do not capture seed variance in GEPA evolution or GRPO training. eICU results also lack the paired tests reported for MIMIC-IV.

  5. The long-tail claim needs stronger evidence. Predicting 585 unique codes shows breadth, not accuracy on rare codes, and Figure 3 covers 25 selected conditions with a rank-based score that is not a standard AUROC. Frequency-stratified precision and recall would make the claim testable.

Questions for authors
  1. Under matched prompts, training objective, and parameter budget, how does MOSAIC compare with a single shared TinyLoRA adapter, hard routing to the dominant specialty, uniform mixing, and each specialty adapter used alone?

  2. Are the LoRA rows in Table 2 trained with the evolved prompts and the same GRPO reward? If not, can the comparison be rerun under identical conditions?

  3. How are next visits defined and ordered in eICU? Were prompts and adapters retrained on eICU, or transferred from MIMIC-IV?

  4. How many codes does each system predict per patient, and how are prediction counts or thresholds set for the deep EHR baselines and the frozen LLMs?

  5. Can the authors report results across several independent seeds for GEPA and adapter training, and provide the full CCSR and Clinical Group mappings and implementation details?

  6. Does the long-tail advantage hold under frequency-stratified precision and recall, rather than unique-code counts?

2 comments

We thank the reviewer for a precise and constructive review. We ran the controls below; changes in the revised manuscript are in blue, and these analyses appear in Appendix D (Table 7), E (Tables 8 and 9) and G (Table 11). Numbers are macro F1 (Chapter / CCSR / Clinical Group / Code).

W1, W2, Q1, Q2: what the composition adds

All adapter rows match MOSAIC's 78K budget, prompts, GRPO objective and reward; the single shared adapter is one 78K adapter trained on all patients, no routing; hard routing uses the stored (dominant-)specialty label's adapter; uniform mixture averages all eight adapters equally.

VariantChapterCCSRClin. GroupCode
MOSAIC0.2580.3080.3770.124
Single shared adapter0.2540.3050.3680.116
Hard routing0.2540.3010.3710.121
Uniform mixture routing0.2510.3040.3720.117
Prompts only0.2490.2960.3610.102
LoRA 78K, evolved prompts0.2550.2920.3450.105

MOSAIC exceeds every other row of the table at every granularity. Each adapter applied alone to every patient (W2, Table 11) scores 0.359 to 0.371 at Clinical Group and 0.107 to 0.120 at Code against the mixture's 0.377 and 0.124, and the best single adapter is within 0.001 of the mixture at Chapter and CCSR. Training the shared adapter 3.6 times longer (1,800 steps) leaves it below MOSAIC at CCSR, Clinical Group and Code.

Q2: Table 2's LoRA rows use MOSAIC's GRPO reward with a generic instruction. Retrained with the evolved prompts (last row), the matched LoRA gains at most 0.014 and stays below MOSAIC everywhere.

W3: trajectory conditioning

The router uses only the chapter composition of the history and is order-free. We permuted each patient's event order, keeping events, routing weights and prompt unchanged.

ChapterCCSRClin. GroupCode
MOSAIC0.2580.3080.3770.124
Event order shuffled0.2290.2770.3370.096

The loss (0.028 to 0.039, paired difference in Appendix G) exceeds that from removing the adapters entirely, and its 95% interval excludes zero everywhere: the predictions are conditioned on the trajectory. MOSAIC reads the ordered trajectory through its input context, and the router stays order-free so that its weights remain an inspectable statistic of the record.

W4, Q3 to Q5: protocol and reproducibility

  • Q3. Each record is one chronological event sequence. MIMIC-IV: cutoff 30 days before the last recorded event, target the ICD-10 diagnoses recorded after it. eICU: ICU stays concatenated in unit-visit order, cutoff 7 days before the last event, target every diagnosis in those 7 days. Table 3 refits prompts and adapters on eICU with backbone, router (MIMIC-IV prototypes) and hyperparameters unchanged; paired tests (Appendix B) put MOSAIC significantly above 16 of 18 baselines at Chapter, CCSR and Clinical Group and 18 of 18 at Code.
  • Q4. Every system is scored on its top three predictions over one fixed 357-category vocabulary with no tuned threshold (LLMs: five codes, first three scored; deep EHR baselines: three highest-scoring codes).
  • W4. The deep EHR baselines were retrained from scratch; where the original paper specifies a configuration at a comparable scale, we followed it.
  • Q5. Adapters act on all seven projection matrices in all 28 blocks (rank 2, 50 values per module, Adam 1e-4); retraining all eight adapters from a second seed changes F1 by at most 0.003, not significant (Appendix C). Appendix I defines the clinical groups and Section 4.4 the CCSR assignment rule; both mapping files will be released.

W5, Q6: long-tail accuracy

Three equal groups of the 357 categories by target frequency. Rows: MOSAIC, its two ablation points (the frozen Qwen2.5-7B backbone and the matched 78K LoRA of Table 2), and TransformEHR, the strongest deep EHR baseline in Table 1 at every granularity:

Macro F1Frequent (74% of targets)Medium (18%)Rare (8%)
MOSAIC0.1380.1200.113
Qwen2.5-7B0.1160.0950.086
LoRA0.1090.1070.094
TransformEHR0.1270.0460.016

MOSAIC has the highest F1 in every group, with its widest margins on medium and rare codes. Pooled precision / recall (Table 9):

FrequentMediumRare
MOSAIC0.179 / 0.1750.151 / 0.1230.118 / 0.117
TransformEHR0.199 / 0.1880.179 / 0.0550.148 / 0.012

TransformEHR's rare-code precision comes from predicting rare codes almost never (recall 0.012); MOSAIC recovers ten times as many rare targets at comparable precision.

We hope these results address your concerns and would be glad to answer any further questions.

Thank you for your response and for addressing my questions. After carefully reviewing your rebuttal, I maintain my initial rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.