{"id":"da4e8fa7-1479-4059-a57b-4921bfadd3ed","arxiv_id":"2505.10856","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"ImputeINR uses implicit neural representations to impute missing time series values, reporting better MSE and MAE than nine baselines on eight datasets, with the largest gains at high missing rates.","lead":"ImputeINR is a machine learning method that fills in missing values in medical time series by learning a continuous function of time, and it reports better accuracy than existing imputation tools on eight benchmark datasets. It matters because hospital monitoring data is often heavily incomplete, and better imputation could improve downstream disease diagnosis models.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Evaluation protocol does not state where variable clustering and standardization are fit; if fit on the full dataset as Algorithm 2 suggests, test-set statistics leak into ImputeINR's group structure and can inflate the reported SOTA.","rationale":"The reader's weakest assumption identifies the same load-bearing concern I find. The central claim is state-of-the-art imputation, especially at high missing ratios, and that claim requires a fair holdout evaluation. Algorithm 2 is the only explicit description of the ImputeINR pipeline, and it operates on the input X without distinguishing training from test data. The paper also states that datasets are pre-split, but it never says whether the clustering and standardization steps are fit on the training split and then frozen for test evaluation. If they are fit on the full dataset or on the masked test set, the model's group structure and normalization are informed by test-set statistics. This is a form of transductive leakage: it does not use labels, but it still violates the standard training-only preprocessing assumption and can inflate the reported advantage. I am not claiming the authors intentionally leaked information; I am claiming the text does not rule it out, and the described algorithm points in that direction. Other concerns, such as the omission of HyperTime and TimeFlow as INR baselines and the absence of error bars, are real but secondary: they affect the strength of the SOTA claim and its statistical support, whereas the preprocessing-split issue directly affects the validity of the reported numbers. The proposed concrete check would settle the matter. If the result is unchanged under training-only preprocessing, the central claim stands and the conditional verdict is appropriate; if the result degrades, the paper needs protocol corrections and possibly weaker conclusions.","tokens_in":18741,"tokens_out":6263,"duration_ms":70857,"concrete_test":"Recompute Table 1 with one protocol change: fit the agglomerative variable clustering and the per-variable standardization statistics on the training windows only, then apply the resulting permutation and normalization to the test windows without refitting. Compare ImputeINR's average MSE and per-dataset ranks against the published table. If performance degrades materially or the ranking against BRITS/CSDI shifts, the published result depends on test-set preprocessing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"To support the headline claim, the comparison must be a fair holdout evaluation. The paper never states where variable clustering or standardization are fit. Section 4.1 says datasets are pre-split and masked based on the mask rate r, but Algorithm 2 in Appendix A.2 takes 'Time series data X with missing values' and performs feature clustering on the N features of X (line 3) before reordering and standardizing. If X is the full dataset or includes test windows, the agglomerative clustering determines the adaptive group partition from test-set statistics, and the standardization uses test-channel moments. Because the group-based INR architecture allocates different capacity to different variable groups, this is not a neutral preprocessing choice: it tailors the model structure to the evaluation distribution. The inflation can be larger for ImputeINR than for baselines that use fixed architectures and training-only normalization, so the reported 62.0% average MSE reduction over the second-best method is not yet established. Since no protocol sentence rules out this reading, the central claim rests on an unverified assumption about where preprocessing statistics are computed.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ImputeINR, a time series imputation method based on implicit neural representations (INR). The model learns a continuous function from timestamps to multivariate values, using a transformer encoder to predict INR parameters from observed data. A variable clustering step groups variables with similar distributions, and an adaptive group-based MLP with trend, seasonal, and residual components forms the INR function. Multi-scale convolutional feature extraction is added. Experiments on eight datasets with mask rates from 10% to 90% are reported, claiming state-of-the-art MSE/MAE in most conditions, and downstream disease diagnosis on Phy2012, Phy2019, and MIMIC3 shows improved AUROC compared with other imputation baselines. The paper includes ablation studies, efficiency analysis, and robustness checks.","tokens_in":18979,"tokens_out":7468,"duration_ms":72886,"significance":"If the results hold, ImputeINR would be a useful contribution for imputing time series at high missing ratios, which is practically relevant for healthcare data. The architectural idea of combining variable clustering with a group-based INR is interesting, and the ablation studies support the contribution of each module. The authors state that code is available, which aids reproducibility. However, the empirical support for the central claim is weakened by an ambiguous preprocessing protocol, the omission of the closest INR-based baselines, and the absence of error bars and significance tests. These issues need to be addressed before the state-of-the-art claim can be accepted.","major_comments":[{"comment":"The experimental protocol never states whether variable clustering and standardization are computed on the training set only or on the full dataset including test windows. Algorithm 2 takes 'Time series data X with missing values' as input and performs feature clustering on X (line 3) before reordering and standardizing; in the evaluation setup X is the masked test set. If clustering and normalization moments are derived from test-set statistics, the adaptive group-based INR architecture is tailored to the test distribution, which is not neutral preprocessing and could inflate ImputeINR's advantage over baselines that use fixed architectures and training-only normalization. Please specify that clustering and standardization are fit on the training set only, or rerun the experiments with train-only preprocessing and report the resulting differences.","section":"Section 4.1, Algorithm 2"},{"comment":"The related work identifies HyperTime and TimeFlow as the closest INR-based imputation methods and criticizes their limitations, but neither appears in the experimental comparison in Table 1. Since the paper's central novelty is an INR-based imputation architecture, omitting these two baselines leaves the state-of-the-art claim incomplete. Please include them in the comparison under the same masking and preprocessing protocol, or provide a specific justification for their exclusion.","section":"Section 2.3, Table 1"},{"comment":"The main results report a single run per condition with no standard deviations or statistical significance tests. Several per-condition differences are very small (e.g., Weather 90% MSE 0.065 vs 0.066; Phy2019 10% MSE 0.071 vs 0.072), and the headline '62.0% average MSE reduction' is presented without confidence intervals. Without repeated-seed variability or significance testing, the claim of state-of-the-art performance is not robustly established. Please report means and standard deviations over multiple seeds and, where feasible, paired significance tests for the main comparisons.","section":"Table 1, Section 4.2"}],"minor_comments":[{"comment":"Algorithm 1 can loop forever: if the closest pair distance d(C_i,C_j) is not less than epsilon, no merge occurs but the while condition |C|>1 remains true. Please revise the termination condition, e.g., break when the minimum distance is >= epsilon.","section":"Appendix A.2, Algorithm 1"},{"comment":"Several free parameters are not reported: the trend polynomial degree m in Eq. (10), the clustering stopping criterion epsilon in Algorithm 1, the similarity metric used in Eq. (3), and the handling of missing values when computing the similarity matrix from partially observed data. Please specify these choices for reproducibility.","section":"Section 4.1, Methodology"},{"comment":"The 'Average' row should state how the average is computed (over all dataset-mask-rate conditions) and which method is considered second-best for each metric; currently the comparison baseline for the 62.0% reduction claim is ambiguous.","section":"Table 1"},{"comment":"Please specify whether the LSTM classifier is trained on the imputed training set and evaluated on the imputed test set, and whether it is retrained separately for each imputation method; otherwise the downstream diagnosis comparison is not fully reproducible.","section":"Section 4.3"},{"comment":"The efficiency analysis reports running time and model size only as a bubble chart; please provide numeric values or a table so that readers can verify the efficiency comparison quantitatively.","section":"Section 4.4, Figure 4"}],"recommendation":"major_revision","confidential_remarks":"The INR idea and the group-based architecture are worth pursuing, and the paper covers a practically important problem. The main concerns are methodological clarity and experimental completeness rather than conceptual validity of the approach. I would encourage the authors to address the preprocessing leakage issue explicitly, add the omitted INR baselines, and provide error bars or significance tests. I do not see evidence of a fundamental flaw, but the current evidence is not sufficient for acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a solid, incremental extension of INR-based time series imputation. The central idea—predicting the parameters of an INR with a transformer and letting the residual term use a group-structured MLP—is well motivated, and the ablation study suggests each component (multi-scale features, clustering, adaptive groups) does something. The efficiency analysis is also a plus: the model is light and fast. If the reported numbers hold, this would be a practical contribution for high-missing-rate clinical data.\n\nBut the numbers should not be taken at face value yet. The most serious problem is the preprocessing protocol. The paper never states whether variable clustering and standardization are fit on the training set only or on the full dataset including test windows. Algorithm 2 in Appendix A.2 reads as if feature clustering is performed on the input X, which in the evaluation is the masked test set. If clustering and normalization use test-window statistics, the group structure is tailored to the evaluation distribution, and that is leakage. It would not be a neutral preprocessing choice—it could inflate the reported gains, especially because the ablation shows clustering plus grouping is the strongest module combination. This is not a fatal flaw, but it is a load-bearing ambiguity that has to be resolved before the SOTA claim is credible.\n\nThe second issue is the baseline list. HyperTime and TimeFlow are cited in Section 2.3 as prior INR imputation methods, but neither appears in Table 1. Claiming state-of-the-art imputation without comparing to your two closest relatives is a gap a reviewer will not let pass. The paper also lacks error bars or significance tests across the five mask rates, and the code link is not an actual URL in the text, so reproducibility is currently unverified.\n\nThe math itself is straightforward and there is no circularity concern: the method is evaluated on independent benchmarks with a standard reconstruction loss. The clustering idea is sensible, and the synthetic experiment in Appendix A.3 gives some intuitive support for why grouping helps. The disease diagnosis results are a nice extra, though they also lack error bars.\n\nWho is this for? Researchers working on INR-based time series modeling or on imputation for high-missing-rate healthcare data. It is worth a serious referee, but the referee should send it back for a revised protocol, the two missing baselines, and at least basic variance reporting. If the authors clarify that clustering and standardization are train-only and the results still hold, this becomes a useful paper.","headline":"Reasonable INR-imputation extension with a plausible grouped-MLP design, but the SOTA claim rests on an under-specified preprocessing protocol and missing closest baselines.","tokens_in":19486,"tokens_out":1870,"would_cite":false,"duration_ms":21969,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"ImputeINR claims that time series imputation, even with 90% of values missing, is best done by learning a continuous implicit neural function of time rather than imputing discrete points.","keywords":["time series imputation","implicit neural representations","missing data","disease diagnosis","continuous function modeling","variable clustering","healthcare time series"],"falsifier":"Re-run the eight-dataset comparison with variable clustering and standardization fit on training windows only and frozen before any test window is processed. If ImputeINR's average MSE advantage over the second-best method drops from the reported 62.0%, or from 68.2% at 90% masking, to a small or negative margin, the claim that continuous INR imputation is intrinsically superior at high missing rates would be falsified.","tokens_in":18568,"feed_emoji":"🩺","tokens_out":13418,"duration_ms":110006,"temperature":0.7,"pith_summary":"ImputeINR is built on a simple bet: a time series is a continuous signal, so the right way to impute it is to learn a continuous function of time, not to fill discrete slots. The paper proposes an implicit neural representation (INR), a network that maps any timestamp to the values of all recorded variables, whose parameters are predicted by a transformer from the few observed points. On eight benchmark datasets with 10% to 90% of values randomly masked, the authors report the best or second-best MSE or MAE in most settings, with an average MSE reduction of 62.0% against the second-best method and 68.2% at the 90% mask rate. Because real healthcare records are extremely incomplete, with MIMIC-III at 63.15% missing and the PhysioNet ICU challenges around 79.67%, the claim matters for downstream diagnosis, where using ImputeINR-imputed data raises AUROC on three clinical benchmarks.","feed_headline":"At 90% missing, continuous functions beat discrete imputation","feed_subtitle":"ImputeINR's continuous representation holds imputation error low at 90% missing and lifts diagnosis AUROC.","key_machinery":"The load-bearing object is the INR continuous function $f$ that maps a timestamp $t$ to the variable vector $X(t)$, an MLP queryable at any moment rather than only on the original sampling grid. It is decomposed into a polynomial trend, a Fourier seasonal component, and an adaptive group-based residual MLP, so the network can fit both smooth long-term motion and periodic structure. Its parameters are supplied by a transformer encoder that ingests multi-scale convolutional features of the reordered masked data and outputs the INR tokens serving as weights, which avoids a per-series optimization loop. Variable clustering, using agglomerative clustering with no preset number of clusters, fixes the grouping used by the residual layers, and the claimed benefit is that variables with similar distributions share capacity while cross-variable correlations are still handled by global layers.","core_discovery":"The paper's central claim is that treating a multivariate time series as a continuous signal makes imputation robust to extreme missingness. ImputeINR learns a function $f(t)$ that returns all variables at timestamp $t$, written as a polynomial trend $f_{tre}(t)$, a Fourier seasonal term $f_{sea}(t)$, and a residual term $f_{res}(t)$ computed by an MLP whose layers are partly global and partly grouped by variable cluster. The network weights are not optimized per series; a transformer encoder predicts them, producing the 'INR tokens', from the observed masked data after variable clustering and multi-scale convolutional feature extraction. The authors report that this design achieves the best or second-best imputation error in most of the forty dataset-mask-rate conditions, and that the imputed values improve disease-diagnosis AUROC on PhysioNet 2012, PhysioNet 2019, and MIMIC-III over all compared imputation baselines.","pith_inferences":["If clustering were restricted to training data, the reported numbers might shrink; the paper's Algorithm 2 applies clustering to the input data being imputed, so the exact 62% figure should be read as an upper bound until the protocol is clarified.","The trend, seasonal, and residual decomposition suggests ImputeINR should transfer to irregularly sampled and super-resolution time series, where querying off-grid is the point; neither setting is tested in the paper.","A direct test of which component matters is to compare against the same architecture without the Fourier seasonal term on nonstationary clinical signals; the seasonal term may contribute little there, isolating the residual group MLP as the driver.","In deployment, missingness is present in the very data used to cluster variables, so the grouping must be inferred from incomplete observations; the paper does not test whether clustering remains stable under missingness."],"forward_implications":["At 90% masked values, ImputeINR's average MSE is 68.2% lower than the second-best baseline, so extreme sparsity stops being a hard barrier for imputation.","Imputed healthcare data from ImputeINR improves disease-diagnosis AUROC on PhysioNet 2012, PhysioNet 2019, and MIMIC-III compared with data imputed by all nine baselines and by zero or mean imputation.","Ablations attribute the gain to the combination of multi-scale features, variable clustering, and the adaptive group MLP; using clustering plus the group MLP gives the largest paired improvement, so the continuous representation alone is not sufficient.","Because $f$ can be queried at arbitrary timestamps, the same trained model can impute at off-grid or irregularly spaced times without retraining, a capability grid-based imputers lack.","The runtime-versus-MSE analysis places ImputeINR near the bottom-left corner with a small parameter count, indicating the accuracy gain is not bought by a large model or slow inference."],"supporting_citations":[{"why":"Establishes INR-based time series imputation and the trend, seasonal, and residual decomposition that ImputeINR's continuous function follows.","marker":"[Fons et al., 2022]"},{"why":"Supplies the transformer-as-meta-learner strategy for predicting INR network weights, which ImputeINR uses to produce INR tokens.","marker":"[Chen and Wang, 2022]"},{"why":"CSDI is the conditional diffusion baseline that represents the multi-step stochastic approach the paper argues is slow and weak under extreme sparsity.","marker":"[Tashiro et al., 2021]"},{"why":"SSSD is the second diffusion baseline, providing the comparison against structured state-space diffusion imputation.","marker":"[Alcaraz and Strodthoff, 2022]"},{"why":"TimesNet is one of the nine imputation baselines and, with the companion protocol, supplies the random-masking evaluation used at five mask rates.","marker":"[Wu et al., 2023]"},{"why":"FPT is another baseline and the second source of the masking strategy adopted in the experiments.","marker":"[Zhou et al., 2023]"},{"why":"Provides the official LSTM benchmark and evaluation protocol used for the downstream disease-diagnosis AUROC results.","marker":"[Harutyunyan et al., 2019]"},{"why":"PhysioNet 2012 is one of the three ICU healthcare datasets, with a 79.67% real missing rate, used to test high-sparsity imputation and diagnosis.","marker":"[Silva et al., 2012]"},{"why":"PhysioNet 2019 is the second ICU sepsis-prediction dataset, also carrying a 79.67% missing rate.","marker":"[Reyna et al., 2019]"},{"why":"MIMIC-III is the largest healthcare benchmark, with 63.15% missing values, used for both imputation and diagnosis.","marker":"[Johnson et al., 2018]"}],"fun_headline_variants":["ImputeINR: continuous functions win at 90% missing","Continuous signals beat discrete imputation at high missingness","ImputeINR: robust imputation for extreme missing data","Implicit neural representations fix extreme missingness","Continuous representation boosts imputation and diagnosis"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported gains assume the variable clustering and standardization are computed from the training split only; the paper never states this, and its Algorithm 2 runs clustering on the data being imputed, so if test windows shape the variable groups, part of the measured advantage could come from test-set information.","fun_headline_variants_meta":{"raw":{"variants":["ImputeINR: continuous functions win at 90% missing","Continuous signals beat discrete imputation at high missingness","ImputeINR: robust imputation for extreme missing data","Implicit neural representations fix extreme missingness","Continuous representation boosts imputation and diagnosis"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000173,"raw_usage":{"total_tokens":1269,"prompt_tokens":924,"completion_tokens":345,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":540,"completion_tokens_details":{"reasoning_tokens":269}},"tokens_in":540,"tokens_out":345,"duration_ms":2812,"temperature":1.0,"reasoning_tokens":269,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:01:44.477447+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the eight-dataset comparison with variable clustering and standardization fit on training windows only and frozen before any test window is processed. If ImputeINR's average MSE advantage over the second-best method drops from the reported 62.0%, or from 68.2% at 90% masking, to a small or negative margin, the claim that continuous INR imputation is intrinsically superior at high missing rates would be falsified.","supporting_citations":[{"cited_title":"Transformers as meta-learners for implicit neural repre- sentations","cited_arxiv_id":null,"evidence_quote":"Supplies the transformer-as-meta-learner strategy for predicting INR network weights, which ImputeINR uses to produce INR tokens."},{"cited_title":"One fits all: Power general time series analysis by pretrained lm","cited_arxiv_id":null,"evidence_quote":"FPT is another baseline and the second source of the masking strategy adopted in the experiments."},{"cited_title":"Kale, Greg Ver Steeg, and Aram Galstyan","cited_arxiv_id":null,"evidence_quote":"Provides the official LSTM benchmark and evaluation protocol used for the downstream disease-diagnosis AUROC results."},{"cited_title":"Predict- ing in-hospital mortality of icu patients: The phys- ionet/computing in cardiology challenge 2012.Computing in cardiology, 39:245,","cited_arxiv_id":null,"evidence_quote":"PhysioNet 2012 is one of the three ICU healthcare datasets, with a 79.67% real missing rate, used to test high-sparsity imputation and diagnosis."},{"cited_title":"Early prediction of sepsis from clinical data: the physionetcomputing in cardiology challenge","cited_arxiv_id":null,"evidence_quote":"PhysioNet 2019 is the second ICU sepsis-prediction dataset, also carrying a 79.67% missing rate."},{"cited_title":"The mimic code repos- itory: enabling reproducibility in critical care research","cited_arxiv_id":null,"evidence_quote":"MIMIC-III is the largest healthcare benchmark, with 63.15% missing values, used for both imputation and diagnosis."}],"review_version":1}