{"id":"211cb2f6-4a01-498e-965a-a3be50154cf5","arxiv_id":"2607.04414","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Dynamic incremental learning with entropy-based task clustering and proxy gradients estimates blood glucose from PPG at 0.64 mmol/L MAE under subject-independent validation on a new 183-participant longitudinal benchmark.","lead":"A deep learning system adapts over time to estimate blood glucose from wearable PPG signals despite drift and scarce labels, reaching 0.64 mmol/L MAE on a new 183-person longitudinal dataset. If robust, it points toward lower-burden glucose tracking without fingersticks or CGM sensors.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"The headline MAE/RMSE may partly reflect repeated windows sharing one capillary label rather than a robust PPG–glucose mapping.","rationale":"The paper is careful: subject-independent folds, ablations, PGBA audit, aggregation tables, and proof-of-concept framing support CONDITIONAL rather than REJECT. The reader correctly flags the PPG–glucose physiological assumption as the soft spot. The more immediate load-bearing issue for the numerical claim is evaluation unit: window-level metrics inflate sample size and can understate dependence on a single capillary label per acquisition. Aggregation analyses already exist and are somewhat reassuring, so this does not overturn the verdict, but it is the check that most directly tests whether the central numbers are robust. Agreement with the reader is partial: same underlying risk (weak/confounded mapping), different concrete failure mode (label-sharing / aggregation). No change from CONDITIONAL; external generalizability and high-glucose performance remain open.","tokens_in":39584,"tokens_out":550,"duration_ms":6627,"concrete_test":"Recompute Table 4 and the main MAE/RMSE/Clarke metrics using only one prediction per 40 s acquisition (mean of its 4 s windows) and report subject-macro means with bootstrap CIs, stratified by y<5.55, 5.55–10, and ≥10 mmol/L. If acquisition-level MAE rises above ~0.9 mmol/L overall or high-glucose MAE stays near 1.2+ with large CI overlap vs. the best baseline, the headline claim weakens.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The strongest claim rests on window-level MAE 0.64 and Clarke A+B 97.69% under subject-independent 5CV. Each 40 s acquisition is split into ten 4 s windows that share one capillary BG label (Stages 4–5; §4.2, 4.4). Adjacent windows are highly correlated short-term morphology around the same reference, so window-level error can look strong even if the model mainly tracks acquisition-level vascular state rather than glucose-specific morphology. Aggregation helps (acquisition MAE 0.60, within-band 93.8%; subject-macro 0.65/93.6%), but the primary SOTA table and headline metrics remain window-level, and high-glucose slices still degrade (RMSE 2.34, within-band 81.7%). If the physiological link is weak or confounded, the reported gains over reimplemented baselines may not survive evaluation that treats each acquisition (or subject) as the independent unit and stratifies by glucose range.","agreement_with_reader":"partial"},"referee_report":{"model":"grok-4.5","summary":"The manuscript proposes Dynamic Incremental Learning (DIL) for non-invasive blood glucose estimation from PPG, combining MERDC/ME2AC for unsupervised latent-task discovery and replay with a Proxy Gradient Bridging Agent (PGBA) that exploits unlabeled PPG via bounded proxy gradients. It also introduces a longitudinal benchmark of 183 participants (5,224 four-second windows) with synchronized PPG, capillary BG, and cuff BP over 285 days. Under subject-independent 5-fold CV with explicit fold isolation, the full system reports MAE 0.64±0.01 mmol/L, RMSE 1.29±0.10 mmol/L, and 97.69±1.63% in Clarke A+B, outperforming reimplemented baselines under a fixed-configuration protocol. Ablations, a PGBA alignment audit, aggregation-level ISO-style analyses, glucose-range/temporal slices, and a small 7-day pilot support a proof-of-concept claim for label-efficient adaptation under longitudinal drift.","tokens_in":39935,"tokens_out":1640,"duration_ms":25711,"significance":"If the results hold under the stated protocol, the work is a useful contribution on two fronts: (i) a carefully curated longitudinal PPG–capillary-BG benchmark with transparent curation and leakage controls, which the field currently lacks at this scale; and (ii) a domain-adapted continual-regression recipe that jointly addresses distribution drift and label scarcity, with component ablations and a proxy-gradient reliability audit that go beyond typical PPG-glucose papers. The fixed-configuration reimplementation of baselines and aggregation-aware reporting are strengths. The clinical and translational significance remains provisional: single-center capillary references, degraded high-glucose performance, and a six-subject pilot limit claims beyond research-stage proof-of-concept, which the authors largely acknowledge.","major_comments":[{"comment":"Abstract, Table 4, and §4.3 lead with window-level MAE/RMSE/Zone A, but each 40 s acquisition is partitioned into ten 4 s windows that share one capillary BG label (§3.1.2 Stages 4–5; §4.2). Adjacent windows are highly correlated morphology around the same reference, so window-level error can overstate effective sample independence and inflate comparative gains. Tables 5–6 show acquisition- and subject-macro metrics remain favorable (e.g., acquisition MAE 0.60, within-band 93.8%), yet these are secondary. For the central SOTA claim to be load-bearing, Table 4 and the abstract should report acquisition-level (and preferably subject-macro) metrics as co-primary, with window-level retained only as a secondary analysis, and any ranking of methods should be verified at those units.","section":null},{"comment":"Table 6 high-glucose slice (y≥10.0 mmol/L): RMSE rises to 2.34 mmol/L and ISO-related within-band falls to 81.7%, versus 1.02 / 93.9% in the 5.55–10 range; Fig. 6 also shows vertical spread near ~16 mmol/L. This is clinically the most consequential regime and is under-emphasized relative to the headline 0.64 MAE and 97.69% A+B. The manuscript should (i) report baseline methods on the same glucose-range slices, (ii) state sample counts per range, and (iii) qualify the main claim so that strong overall averages are not read as uniform accuracy across the 4.2–16.3 mmol/L spectrum.","section":null},{"comment":"§3.5.2 Eqs. (12)–(15): ME2AC’s task structure rests on histogram plug-in entropy and mutual information over binned standardized raw PPG-window coordinates. In high-dimensional waveform space this is a coarse discrete morphology proxy; the paper does not show that discovered partitions correspond to interpretable physiological regimes (e.g., perfusion, autonomic state, glycemic context) beyond clustering scores (Tables 10–12). Because DIL’s curriculum, memory allocation m_b, and gradient projection all depend on these partitions, please add evidence that ME2AC tasks are more than a useful regularizer—e.g., stability of partitions under feature alternatives (learned embeddings vs raw bins), or correlation of task identity with metadata available in the benchmark (BP, diabetes status, encounter timing)—or else reframe ME2AC strictly as an operational clustering heuristic rather than latent","section":null},{"comment":"Table 4 comparison protocol (§4.3): baselines are reimplemented with original/default settings and no fold-specific tuning, while the proposed method uses a fixed but author-chosen configuration (C_mem=512, α=0.15, 3:1 labeled:unlabeled, ME2AC quantiles, etc.; §4.1). This is preferable to test leakage, but it leaves open whether large margins (e.g., vs Zhang et al. 2024 MAE 0.99) partly reflect unequal optimization effort or backbone capacity rather than DIL/MERDC/PGBA. Please add (i) a controlled backbone-matched comparison (same InceptionTime without DIL modules), already partly in Table 7, elevated into the main comparison narrative, and (ii) at least a minimal sensitivity or published-hyperparameter sweep for the strongest baselines so the ranking is not solely default-vs-tuned.","section":null}],"minor_comments":[{"comment":"§5.1 / Fig. 9: the seven-day pilot (n=6, alternate wrist device, mixed SMBG/venous references) is useful feasibility evidence but should be more clearly labeled as exploratory and not used to support robustness language that could be read as multi-day clinical validation.","section":null},{"comment":"Notation: 𝐉, 𝓁, ̄𝑔, and ̃𝑔 are dense in §3.4–3.6; a short symbol table already exists (Table 1) but does not list all gradient symbols—extend it for readability.","section":null},{"comment":"Fig. 10(b): the y-axis is a sampled-window index for visualization only; the caption states this, but the main text still risks being read as a physiological axis—add an explicit one-line caveat in §5.2.","section":null},{"comment":"Data availability: controlled-access release is acceptable, but for a paper whose contribution includes establishing a community benchmark, state the access criteria, expected timeline, and what metadata (comorbidities, encounter intervals) will ship with the windows.","section":null},{"comment":"Minor consistency: abstract reports 97.69±1.63% A+B while Table 13 mean A+B is 97.70±1.63—align rounding throughout.","section":null},{"comment":"§4.5.3 signal-length ablation: only durations near 4 s are discussed in detail; if 1–2 s and >7 s were tested, report the full curve or table values, not only the narrative summary.","section":null}],"recommendation":"major_revision","confidential_remarks":"Solid engineering paper with above-average experimental hygiene for PPG-glucose work; the main risk is over-reading window-level SOTA numbers as a solved sensing problem. Fit is appropriate for a biomedical signal-processing venue if reframed firmly as benchmark + proof-of-concept continual learning. I would not reject on physiological skepticism alone—the authors already hedge—but I would require acquisition-level primary metrics and high-glucose stratification before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful part of this paper is not a miracle non-invasive glucometer. It is a carefully built 183-person, 285-day PPG + capillary BG + cuff BP benchmark, plus a domain-specific continual-learning stack (MERDC/ME2AC task discovery, GEM-style replay, PGBA proxy gradients without pseudo-labels) evaluated under subject-independent 5CV with explicit fold isolation.\n\nWhat is actually new is the packaging for continuous regression under unlabeled PPG streams and latent physiological regimes, not class-incremental learning. They reimplement baselines under fixed configs, run component ablations, audit PGBA alignment on labeled folds, and report acquisition- and subject-macro ISO-style bands. That is more disciplined than most PPG-glucose papers. The 7-day pilot is small but at least tries longitudinal behavior. Citation pattern is normal for the area; math is standard GEM + histogram MI/entropy clustering, not hand-wavy.\n\nThe stress-test on shared labels is real but overstated as a collapse. Ten 4 s windows share one capillary label, so window MAE 0.64 and Clarke A+B 97.7% can look better than independent clinical units. The authors already show acquisition MAE ~0.60 and subject-macro ~0.65 with within-band still high, and they flag high-glucose degradation (RMSE 2.34). Primary SOTA table is still window-level, which is the soft spot to push in review, not a reason to dismiss the work. Bigger limits: single-center capillary (not venous) reference, controlled-access data without shipped code, fixed free parameters (C_mem, alpha, ME2AC quantiles, 3:1 labeled/unlabeled), and the usual unproven PPG–glucose physiological link under meds/comorbidities.\n\nThis is for wearable sensing and digital diabetes methods people who care about drift and label scarcity, not for clinicians ready to deploy. Internal experimental claims look solid enough; external generalizability is open and the paper mostly says so.\n\nI would send it to peer review. Engage if you work on continual physiological sensing or need a longitudinal PPG-BG corpus; treat the headline numbers as proof-of-concept under their protocol, not as device accuracy.","headline":"Solid methods-and-benchmark paper: real longitudinal PPG–BG corpus plus a carefully isolated continual-learning pipeline; headline window metrics are a bit soft, but aggregation and ablations keep the claim honest enough for referees.","tokens_in":40518,"tokens_out":562,"would_cite":true,"duration_ms":7177,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A continual-learning pipeline estimates blood glucose from PPG with 0.64 mmol/L MAE on a 183-person longitudinal wearable benchmark.","keywords":["non-invasive blood glucose","photoplethysmography","dynamic incremental learning","continual learning","proxy gradients","mutual entropy clustering","wearable physiological signals","longitudinal benchmark"],"falsifier":"On held-out people or multi-center wearable streams, if the full DIL+MERDC+PGBA system fails to stay near 0.64 mmol/L MAE with high Clarke A+B under subject-independent evaluation—especially if ablating MERDC or PGBA no longer hurts, or high-glucose and comorbidity subgroups collapse—the central claim fails.","tokens_in":40497,"feed_emoji":"💓","tokens_out":728,"duration_ms":7670,"temperature":0.7,"pith_summary":"Fingerstick and CGM glucose checks are invasive and costly, while wearable PPG is easy to collect but drifts over time and usually lacks matching glucose labels. This paper argues that the right learning setup can still turn short PPG windows into usable glucose estimates: discover latent physiological regimes, replay them under gradient constraints so older knowledge is not erased, and train a proxy-gradient agent so unlabeled PPG streams help the model adapt without inventing fake glucose labels. The authors also release a 183-participant, 285-day fingertip PPG, capillary glucose, and cuff blood-pressure benchmark for the community. On subject-independent five-fold validation the full system reaches 0.64 mmol/L mean absolute error, 1.29 mmol/L RMSE, and about 98% of estimates in Clarke zones A+B, remaining stable when windows are averaged by acquisition or by subject. The practical claim is a proof-of-concept that adaptive, label-efficient wearable glucose estimation is possible under real longitudinal drift, not only on static snapshots.","feed_headline":"PPG glucose estimates hit 0.64 mmol/L MAE with continual learning","feed_subtitle":"A 183-person longitudinal wearable benchmark and label-efficient adaptation under drift","key_machinery":"Dynamic Incremental Learning (DIL) with MERDC and PGBA: ME2AC partitions streaming PPG into latent tasks by balancing local signal entropy and inter-cluster mutual information for task-balanced replay and gradient projection, while PGBA maps model outputs to bounded proxy gradients so unlabeled PPG windows update the model without pseudo-glucose labels.","core_discovery":"Dynamic incremental learning that pairs mutual-entropy task discovery and replay (MERDC/ME2AC) with an uncertainty-quantified Proxy Gradient Bridging Agent can continually adapt a PPG-to-glucose regressor under distribution drift and sparse labels, reaching 0.64 ± 0.01 mmol/L MAE and 97.69 ± 1.63% Clarke A+B on a new 183-participant longitudinal benchmark under subject-independent validation.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Dynamic incremental learning hits 0.64 mmol/L MAE on wearable PPG glucose","MERDC + PGBA adapt PPG-to-glucose under drift: 0.64 MAE, 97.7% Clarke A+B","183-person longitudinal PPG benchmark enables 0.64 mmol/L non-invasive glucose MAE","Label-efficient continual adaptation yields 0.64 mmol/L MAE for PPG glucose estimates","Subject-independent DIL reaches 0.64 ± 0.01 mmol/L MAE on 285-day PPG glucose set"],"cache_read_input_tokens":32896,"weakest_assumption_plain":"The claim stands only if short PPG waveform windows carry a stable, learnable link to capillary glucose that can be split into a few unsupervised latent tasks whose replay and bounded proxy gradients are enough to keep adapting without erasing older knowledge.","fun_headline_variants_meta":{"raw":{"variants":["Dynamic incremental learning hits 0.64 mmol/L MAE on wearable PPG glucose","MERDC + PGBA adapt PPG-to-glucose under drift: 0.64 MAE, 97.7% Clarke A+B","183-person longitudinal PPG benchmark enables 0.64 mmol/L non-invasive glucose MAE","Label-efficient continual adaptation yields 0.64 mmol/L MAE for PPG glucose estimates","Subject-independent DIL reaches 0.64 ± 0.01 mmol/L MAE on 285-day PPG glucose set"]},"model":"grok-4.5","effort":"low","cost_usd":0.003158,"raw_usage":{"total_tokens":1079,"prompt_tokens":829,"num_sources_used":0,"completion_tokens":118,"cost_in_usd_ticks":31580000,"prompt_tokens_details":{"text_tokens":829,"audio_tokens":0,"image_tokens":0,"cached_tokens":0},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":132,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":829,"tokens_out":118,"duration_ms":3058,"temperature":1.0,"reasoning_tokens":132,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-11T19:20:36.829042+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On held-out people or multi-center wearable streams, if the full DIL+MERDC+PGBA system fails to stay near 0.64 mmol/L MAE with high Clarke A+B under subject-independent evaluation—especially if ablating MERDC or PGBA no longer hurts, or high-glucose and comorbidity subgroups collapse—the central claim fails.","supporting_citations":[],"review_version":1}