{"id":"98ff42a2-4c4b-4a81-9885-44c09d3f0b56","arxiv_id":"2608.08244","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"Parameter-efficient adaptation of a frozen wearable foundation model gives a shared women's-health representation with positive but non-significant transfer to menstrual onset and cramps, and no uniform advantage over matched temporal baselines.","lead":"FemWear is a lightweight model that adapts a general wearable AI backbone to women's health tasks by training only 1.11% of the encoder. It shows positive trends for menstrual onset and cramp estimation under a strict participant-level audit, but it does not beat all equally sized baselines and is not clinically validated.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central transfer claim rests on a frozen daily embedding cache; the most load-bearing risk is that the +2.19–6.35% nested-audit gains reflect participant-level covariates already encoded in the frozen representation, not menstrual-specific signal.","rationale":"The reader's weakest_assumption is exactly the frozen daily cache / end-to-end adaptation limitation, and the most load-bearing consequence is that the claimed positive transfer could be driven by participant-level covariates or device information in the frozen embeddings rather than by menstrual-specific signal. The paper explicitly acknowledges this in Section 6. The concern is not that the authors are dishonest; they are unusually transparent. The concern is that the central claim—parameter-efficient specialization yields positive transfer for selected menstrual endpoints—is not yet secured against the simplest confound: the frozen encoder may already encode stable participant characteristics, and the adapter/head may be learning to exploit those. The nested audit is internal, single-seed, and no endpoint survives multiplicity correction, so the empirical margin for the positive claims is thin. I agree with the reader's conditional verdict: the architecture and evaluation are sound enough to warrant a conditional accept, but the evidential claim for 'selected menstrual endpoints' should be strengthened with a scrambling/permutation control, repeated nested seeds, and an external cohort. The concrete test above would directly settle the leakage question.","tokens_in":10276,"tokens_out":2085,"duration_ms":17134,"concrete_test":"Run the same protocol on a scrambled-control: permute the menstrual-phase and onset labels across participants (or across days within participant) while keeping the frozen daily embeddings and the full FemWear training procedure identical. If FemWear still shows positive relative changes comparable to the reported +2.19–6.35% on the nested audit, then the frozen daily cache is carrying participant- or device-level leakage and the transfer claim collapses. A cleaner version: retrain the adapters and heads on raw minute-level patches (unfreezing the projection or using a small trainable projection) and compare with the cached-embedding result; if raw-level adaptation does not materially change the endpoint, then the daily cache is not the bottleneck and the claim is strengthened; if it improves endpoints substantially, the frozen daily cache is the weak point.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central claim is that a frozen general wearable representation can be repurposed with 1.11% trainable parameters to produce a shared women's-health representation that yields positive, if limited, transfer for 24-hour onset, 72-hour onset, and cramps. The weakest link is the causal pathway from raw signals to the claimed transfer. FemWear never adapts raw minute-level features: it caches daily OpenMHC embeddings from aligned daily patches and freezes the encoder (Section 3.2; Section 6 self-notes: 'Daily cached representations prevent the joint objective from adapting raw minute-level sensor features end to end'). Consequently the trained components (adapters + GRU + heads) can only recombine whatever the frozen daily embeddings already preserve. The paper does not show that the frozen OpenMHC daily embeddings retain sufficient sub-day structure—temperature trajectory shape, HRV dynamics, motion artifact patterns—for menstrual phase or onset. If they do not, no amount of adapter training can recover that information. The nested audit is internal leave-one-participant-out on 42 women, single-seed, and no endpoint has a strictly positive multiplicity-corrected interval (Section 5.2). The positive relative changes (+2.87%, +6.35%, +2.19%) are tiny relative to cross-participant variance and could reflect leakage of participant-level covariates (e.g., average heart rate or temperature level) embedded in the frozen daily representation, rather than a genuinely transferable menstrual-onset signal. The capacity-matched comparison shows FemWear does not beat a shared GRU or MMoE (Section 5.3), so the temporal architecture is not the source of advantage. The few-label experiment is a formal negative result (Section 5.7), which weakens the claim that the shared representation provides data-efficient transfer.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces FemWear, a parameter-efficient specialization of the frozen OpenMHC wearable foundation model for women's-health tasks. FemWear trains 239,236 encoder parameters (1.11% of the encoder) via low-rank residual adapters, a three-expert physiological-regime adapter bank, and causal task-family heads, producing a shared longitudinal representation for menstrual, symptom, affective, sleep/HRV, autonomic, activity, and pregnancy outcomes. The evaluation spans six cohorts with 63 comparable primary metrics, including 33 from women's-health cohorts, and includes a fixed participant split, a nested 42-participant leave-one-participant-out audit, capacity-matched baselines, calibration analysis, missing-history audits, and a label-efficiency audit. The fixed split shows gains for cycle-phase macro-F1, cramps, mood, and sleep, while the nested audit retains positive point estimates for onset and cramps but no endpoint has a strictly positive multiplicity-corrected confidence interval. The paper is carefully hedged and explicitly disclaims universal dominance and clinical validity.","tokens_in":10515,"tokens_out":4470,"duration_ms":42077,"significance":"If the claims hold, FemWear provides a reproducible template for specializing general wearable foundation models to women's-health research with a very small trainable parameter budget. The strengths of the manuscript are real and should be credited: participant-disjoint splits, participant-cluster bootstrap with 2,000 replicates, Holm correction within task families, train-only temperature scaling, capacity-matched baselines under fixed seeds and budgets, and a public code/data-interface release. The honest reporting of negative results, including the failure to outperform shared-GRU and MMoE baselines and the negative label-efficiency result, is a notable strength. However, the central positive transfer claim rests on statistically fragile evidence: the fixed split has only six validation participants, and the nested audit yields no endpoint with a positive corrected confidence interval. The paper is therefore a useful methodological contribution and a clear evidence-boundary statement, but its headline 'targeted transfer' claim should be treated as exploratory rather than established.","major_comments":[{"comment":"The nested leave-one-participant-out audit is the paper's stated evidential boundary, but no endpoint has a strictly positive multiplicity-corrected confidence interval, so the headline 'retained positive changes' (+2.87%, +6.35%, +2.19%) are point estimates that do not survive the paper's own prespecified inference. The abstract and Section 5.2 should state explicitly that the positive-transfer claim is an exploratory trend, not a statistically supported effect; as written, the central claim is not supported by the reported evidence.","section":"§5.2, Table 3"},{"comment":"Because the OpenMHC encoder is frozen and only daily cached embeddings are adapted, the trained adapters cannot recover any sub-day or raw minute-level structure that the frozen daily representations discard. The paper does not test whether the observed gains reflect menstrual-specific signal rather than reweighting of participant-level covariates (e.g., average temperature or heart-rate level) already present in the frozen embeddings. A concrete control—for example, a linear probe on the same frozen daily embeddings, regressing out participant-level means before adaptation, or an adapter trained on participant-identity-shuffled labels—would be needed to support the 'women's-health specialization' interpretation.","section":"§3.2, §6"},{"comment":"The fixed participant split has only six validation participants, so the four endpoints that 'improved in every seed' on that split are not strong evidence of generalization. Since the abstract's first quantitative claims come from this split, the paper should either report the nested audit as the primary result or add an explicit warning in the abstract that the fixed-split gains are development-set results.","section":"§5.1, Table 2"},{"comment":"FemWear failed the preregistered majority-task criteria against both the shared GRU and the MMoE, and no architecture pair produced a Holm-significant advantage. This is honestly reported, but it means the paper does not establish that the proposed dual-path temporal model is superior to equally sized temporal baselines; the architectural contribution should be presented as a negative result or restricted to a post-hoc task-family analysis rather than as a demonstrated advantage.","section":"§5.3, Table 4"}],"minor_comments":[{"comment":"The label 'FemMHC' in Figure 3a should be 'FemWear' to match the rest of the text.","section":"Figure 3a"},{"comment":"The paper reports 69 task outputs and 63 comparable primary metrics; the six excluded outputs should be identified so readers can assess what is not comparable.","section":"§4.1"},{"comment":"The ethics statement says the final manuscript 'must list' source-study ethics approvals and data-use agreements after author verification, which is not a completed disclosure; please add the actual approvals or state clearly that they are pending.","section":"Ethics Statement"},{"comment":"The 'Utility' column in Table 4 is not defined in the text or appendix; please provide the formula or a reference for this metric.","section":"Table 4"},{"comment":"The calibration results report calibrated Brier scores and ECE values but not the uncalibrated baselines; please report both so the 19.8%/30.7% and 88.2%/84.2% reductions are interpretable.","section":"§5.5"},{"comment":"The references for the DEPRESS Fitbit subset, inPHRsym subset, and the wearable HRV/sleep release are missing; the reproducibility checklist itself notes these must be finalized, but a submitted manuscript should include the dataset citations.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is unusually transparent about its limitations, which is a strength, but the central positive result is statistically fragile and the mechanism of transfer is not identified. I would encourage the editor to treat the 'targeted transfer' claim as exploratory until the frozen-embedding confound is addressed with a concrete control and the nested audit is repeated with multiple seeds and a prespecified positive threshold."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nRead FemWear. The one thing to know: this is not a breakthrough model paper; it is a careful empirical study of parameter-efficient specialization of a frozen wearable foundation model for women's-health tasks, and its main value is the transparent evaluation suite plus the explicit reporting of where specialization does and does not help. The second thing: the paper's abstract is appropriately modest, but the strongest positive numbers come from a fixed split with only six validation participants, and the stricter 42-participant nested audit—the one that should carry the claim—is single-seed and produces no endpoint with a strictly positive corrected confidence interval.\n\nWhat is genuinely good: the experimental hygiene is unusually high for this area. Participant-disjoint splits; fixed checkpoints; capacity-matched baselines (shared GRU, MMoE, MLP) with matched parameters, seeds, and budgets; train-only calibration; Holm correction; bootstrap CIs clustered by participant; and formal negative results (few-label, architecture comparison, domain-dependent continual specialization). The three-bin onset head enforcing P24 ≤ P72 is a coherent design choice. The paper ships code, configs, and aggregate artifacts. It also repeatedly states its boundary: no universal dominance, no clinical validity, no from-scratch women-only pretraining. That honesty earns credit.\n\nSoft spots, in proportion. The load-bearing risk is the frozen daily embedding cache. FemWear adapts only 1.11% of encoder parameters via adapters on the final two blocks and caches daily OpenMHC embeddings; it never adapts raw minute-level features. So the trainable components can only recombine what the frozen daily representation already preserves. The paper acknowledges this in Section 6. That means the positive transfer for onset and cramps could partly reflect participant-level covariates (average temperature, heart rate level) already encoded in the frozen embeddings, rather than a genuinely menstrual-specific sub-day signal. The stress-test note pushes this harder than the paper does, but the paper's own admission supports the concern.\n\nSecond, the statistical foundation for the central claim. Six validation participants in the fixed split means the +8.15% phase F1 and the ~9% MAE reductions are suggestive, not confirmatory. The nested audit is better by construction but single-seed, internal leave-one-out, and no corrected CI excludes zero. The capacity-matched result shows FemWear does not beat an equally sized shared GRU or MMoE; the architecture is not the source of any advantage. And the few-label experiment is a clean negative, undercutting the hope that the shared representation provides data-efficient transfer.\n\nNone of this kills the paper. The central modest claim—parameter-efficient specialization can repurpose a frozen general wearable encoder for selected women's-health endpoints, with a reproducible benchmark—is defensible, even if not fully proven. The gaps are ones a serious referee can push on: repeated nested evaluation, an external cohort, and an analysis of what information the frozen daily embeddings actually retain.\n\nWho this is for: wearable ML researchers and anyone building on OpenMHC or similar foundation models. It deserves real peer review; it is exactly the kind of honest, reproducible work that should be in the literature, with revisions to tighten evidential language and finalize ethics/citation items.\n\nRecommendation: send to referees. I would read it again.","headline":"A transparent, well-controlled parameter-efficient wearable specialization study whose honest negative results are as valuable as its modest positive transfer; the central claim is plausible but the statistical evidence is thinner than the fixed-split headline numbers suggest.","tokens_in":11173,"tokens_out":3078,"would_cite":true,"duration_ms":25136,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"FemWear repurposes a frozen general wearable encoder with only 1.11% trainable parameters and shows positive transfer for menstrual onset and cramps under strict participant-level validation.","keywords":["wearable foundation models","women's health","parameter-efficient adaptation","menstrual health","multitask learning","low-rank adapters","calibration","longitudinal health modeling"],"falsifier":"Running the identical nested leave-one-participant-out audit on a new participant-disjoint menstrual cohort collected under a prespecified protocol would settle the claim: if the bootstrap confidence interval for any of the positive relative changes (24-hour onset, 72-hour onset, cramps) includes zero or is negative, the claimed transfer does not generalize.","tokens_in":10002,"feed_emoji":"🩺","tokens_out":9528,"duration_ms":75344,"temperature":0.7,"pith_summary":"FemWear is a parameter-efficient specialization of a general wearable foundation model for women's-health tasks. The paper's central claim is that freezing the pretrained multimodal encoder and training only 239,236 adapter parameters (1.11% of the encoder) yields a shared longitudinal representation that improves selected menstrual and symptom endpoints. On a fixed participant split, cycle-phase macro-F1 rose by 8.15% and cramps, mood-symptom, and sleep-problem mean absolute error fell by 9.32%, 5.80%, and 9.43%. Under a stricter nested leave-one-participant-out audit across all 42 participants, 24-hour onset, 72-hour onset, and cramps retained positive changes of 2.87%, 6.35%, and 2.19%, while phase, mood, and sleep were flat or negative and no endpoint had a strictly positive corrected confidence interval. The paper proposes this as a reproducible specialized-model formulation and a precise boundary of where transfer occurs, not as proof of universal dominance or clinical validity.","feed_headline":"Frozen model + 1.11% adapter improves menstrual onset prediction","feed_subtitle":"Onset and cramps improve under a strict 42-woman audit, while other endpoints do not.","key_machinery":"The load-bearing mechanism is the rank-32 bottleneck adapter inserted into the final two Transformer blocks of a frozen general wearable encoder, so only 239,236 parameters (1.11%) are trainable while the patch projection and the rest of the encoder stay frozen. The adapter update is a zero-initialized residual of the form $\\mathrm{Adapter}(H)=H+W_{\\mathrm{up}}\\,\\mathrm{GELU}(W_{\\mathrm{down}}\\,\\mathrm{LN}(H))$, and a three-expert physiological-regime adapter bank computes a soft mixture of similar updates from the pooled daily embedding. On top of the encoder, a gated recurrent unit summarizes the causal daily history, and task-family heads map the participant state to outputs. The menstrual-onset head outputs a three-bin distribution over onset within 24 hours, 24–72 hours, and later, so $P_{24}=p_0$ and $P_{72}=p_0+p_1$ enforce $P_{24}\\le P_{72}$ by construction. The whole specialization is trained with a partial-label multitask loss that averages within task domains, so cohorts with more labels do not dominate through task count, and a train-only temperature scaling improves probability calibration.","core_discovery":"On the paper's own terms, the discovery is that a frozen general wearable representation can be repurposed into a women's-health-specific representation with a 1.11% trainable-parameter budget, and that the repurposing transfers to a subset of menstrual outcomes rather than across the board. FemWear adds rank-32 residual adapters to the final two Transformer blocks, a causal GRU over daily embeddings, and task-family heads, then trains the specialization on five women's-health cohorts using partial-label multitask learning. The fixed-split gains are substantial for cycle phase (+8.15% macro-F1), cramps (9.32% lower MAE), mood symptoms (5.80% lower MAE), and sleep problems (9.43% lower MAE), but 24-hour onset AUPRC fell by 3.40%. In the stricter 42-participant nested audit, only 24-hour onset, 72-hour onset, and cramps kept positive relative changes, and the paper explicitly reads this as the current evidential boundary: the model enables targeted transfer and coherent probability outputs, but it does not establish universal performance dominance or clinical validity.","pith_inferences":["If the frozen daily embedding bottleneck is real, then larger gains on menstrual phase and symptoms would likely require adapting at the sub-daily or raw-signal level, or a wearable foundation model with finer temporal resolution.","The absence of a few-label advantage suggests that adding independent participants, not simply adding more label days, is the primary path to improving specialization for women's-health tasks.","The capacity-matched comparison points to a testable hypothesis: the positive transfer in FemWear may be carried more by the frozen pretrained representation than by the adapter architecture, so replacing the dual-path head with a shared GRU of equal size might preserve most of the benefit.","The train-only calibration result is directly reusable: any downstream task using a three-bin nested onset head can import the same temperature-scaling protocol without touching participant-level held-out data."],"forward_implications":["A frozen general wearable representation can be specialized for a new health domain with about 1% of the encoder's parameters, making it practical for small research cohorts.","Menstrual onset probabilities produced by the three-bin head are logically coherent and can be reliably calibrated with train-only temperature scaling, so they can serve as research-grade outputs even while discrimination stays modest.","The evaluation template—participant-cluster bootstrap, Holm correction, capacity-matched baselines, and a nested leave-one-participant-out audit—gives a transparent way to report where transfer is positive, neutral, or negative.","Domain-dependent continual specialization shows that adding women's-health data to a frozen adapter does not uniformly help; pregnancy transfer actually worsened, so task-family-aware specialization is necessary."],"supporting_citations":[{"why":"Supplies the scalable wearable foundation-model pretraining approach that FemWear inherits as its frozen backbone.","marker":"Narayanswamy et al., 2024"},{"why":"Provides the open-weight general wearable encoder, the OpenMHC representation, and the 32-task ability-retention benchmark used as the transfer reference.","marker":"Schuetz et al., 2026"},{"why":"Provides the mcPHASES dataset of 42 menstruating participants with synchronized wearable, symptom, and hormone data that anchors the menstrual evaluation.","marker":"Lin et al., 2025"},{"why":"Introduces the residual bottleneck adapter mechanism that FemWear uses to specialize the frozen encoder with few parameters.","marker":"Houlsby et al., 2019"},{"why":"Supplies the low-rank adaptation perspective underlying the FemWear adapter update.","marker":"Hu et al., 2022"},{"why":"Provides the multitask negative-transfer result that motivates the paper's capacity-matched control experiments and domain-dependent findings.","marker":"Standley et al., 2020"},{"why":"Supplies the temperature-scaling calibration method used to reduce onset expected calibration error.","marker":"Guo et al., 2017"},{"why":"Supplies the gestational-age activity dataset used as the pregnancy cohort.","marker":"Ravindra et al., 2023"}],"fun_headline_variants":["1.11% adapter repurposes wearable model for women's health","Frozen backbone plus tiny adapter lifts some women's health metrics","1.11% adapter women's health model: gains on some endpoints","Tiny adapter tunes wearable model for women's health, mixed gains","Women's health wearable model: 1.11% params, not universal"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The frozen daily cached embeddings must preserve the needed information from raw minute-level sensor signals (temperature trajectory, heart-rate-variability dynamics, movement patterns), because the joint objective never adapts raw features end to end.","fun_headline_variants_meta":{"raw":{"variants":["1.11% adapter repurposes wearable model for women's health","Frozen backbone plus tiny adapter lifts some women's health metrics","1.11% adapter women's health model: gains on some endpoints","Tiny adapter tunes wearable model for women's health, mixed gains","Women's health wearable model: 1.11% params, not universal"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001254,"raw_usage":{"total_tokens":5226,"prompt_tokens":1116,"completion_tokens":4110,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":732,"completion_tokens_details":{"reasoning_tokens":4016}},"tokens_in":732,"tokens_out":4110,"duration_ms":25640,"temperature":1.0,"reasoning_tokens":4016,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T00:14:00.344567+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Running the identical nested leave-one-participant-out audit on a new participant-disjoint menstrual cohort collected under a prespecified protocol would settle the claim: if the bootstrap confidence interval for any of the positive relative changes (24-hour onset, 72-hour onset, cramps) includes zero or is negative, the claimed transfer does not generalize.","supporting_citations":[{"cited_title":"2025 , note =","cited_arxiv_id":null,"evidence_quote":"Provides the mcPHASES dataset of 42 menstruating participants with synchronized wearable, symptom, and hormone data that anchors the menstrual evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the multitask negative-transfer result that motivates the paper's capacity-matched control experiments and domain-dependent findings."},{"cited_title":"Weinberger","cited_arxiv_id":null,"evidence_quote":"Supplies the temperature-scaling calibration method used to reduce onset expected calibration error."}],"review_version":1}