{"id":"9e06b4bd-c48f-427c-b2fa-5218f79c1a9b","arxiv_id":"1908.04209","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"MixMI, a mixture of Gaussian-process and linear-regression imputers with individualized mixing weights, reports lower mean absolute scaled error than six benchmarks on all four datasets tested.","lead":"MixMI is a new imputation method for clinical test time series that mixes Gaussian process and linear regression models with per-patient weights. On MIMIC-III, NMEDW, and synthetic variants, it reports lower imputation error than six existing methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Synthetic-dataset timing leaks the true masked value into the time covariate, so the 'all 4 datasets' claim may be driven by contaminated synthetic benchmarks.","rationale":"The central claim is empirical: MixMI is the most accurate imputer on all 4 datasets. The reader's weakest assumption concerns the equal-timestamp condition on real EHR time tensors; that is a legitimate applicability concern, but it does not directly falsify the paper's benchmark on its own constructed tensors. The synthetic-time leakage is more load-bearing because it attacks the validity of two of the four datasets used to support the headline claim. If the masking were applied before generating synthetic times, the test values' time coordinates would no longer encode information about the true outcomes. This is a concrete, fixable flaw: the authors could rerun the synthetic experiments with masking-before-generation and either confirm or refute the leakage. The real-world results and the availability of code are independent evidence, so the verdict remains conditional rather than a rejection. The reader's time-tensor concern should still be addressed in revision, but it is not the single most damaging issue.","tokens_in":19244,"tokens_out":13909,"duration_ms":156378,"concrete_test":"Regenerate the two synthetic MIMIC datasets by applying the 20% masking before computing synthetic times: treat masked values as missing in the time-generation rule of Section IV.B and interpolate their synthetic times exactly as originally missing values are handled. Rerun MixMI, MixMI-LL, 3D-MICE, GP, MICE, GMM, M-RNN, and the 1000-replicate permutation test on these corrected datasets. If MixMI's synthetic-dataset MASE is no longer the best or the p-value is no longer .001, the original synthetic superiority is an artifact of label leakage into the time tensor.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing flaw is in the synthetic-data benchmark construction in Section IV.B combined with the masking procedure in Section IV.C. The synthetic time for an observed measurement is t'_i = t_i + S * sum_{j=2..i} d(Delta_x_j - Delta_t_j), where Delta_x_j is defined using the actual measurement value x_j. The 20% masking of observed values is applied only after these synthetic times are computed. Consequently, for every masked test value, the time covariate fed to all imputation models is a function of the true value being predicted. A GP or any smooth temporal model can exploit this: the adjusted time already encodes how close the missing value is to neighboring measurements. This label leakage contaminates the two 'Synthetic MIMIC' rows of Table II and the statement that MixMI is significantly better than the second-best model on all 4 datasets at p=.001. The real-world MIMIC and NMEDW results are not affected in this way, but the headline 'all 4 datasets' claim is.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MixMI, a mixture-based multiple imputation model for multivariable clinical time series. It combines two linear-regression views (cross-sectional and temporal) with a Gaussian-process component and introduces individualized per-patient mixing weights, trained via the EM algorithm. The model is evaluated against six benchmarks on two real EHR datasets (MIMIC-III and NMEDW) and on two synthetic MIMIC variants with different temporal-correlation strengths, using MASE for imputation accuracy and AUC for a downstream mortality-prediction task.","tokens_in":19485,"tokens_out":5058,"duration_ms":50631,"significance":"The model formulation is coherent, the EM updates are standard, and an open-source implementation is provided. If the reported results are sound, MixMI is a useful contribution to clinical time-series imputation, offering a statistically grounded alternative to heuristic weighting schemes such as 3D-MICE. The real-world results suggest the model is competitive, and the study includes a downstream classification comparison with GRU-D. However, the synthetic-data evaluation is contaminated by label leakage, and the significance testing rests on a single random mask, so the paper's headline 'all 4 datasets' claim is not currently supported.","major_comments":[{"comment":"The synthetic time generation in Section IV.B is a function of the true measurement values: t'_i = t_i + sum_{j=2..i} d(Dx_j - Dt_j) S, with Dx_j = |x_j - x_{j-1}| / sum_{k=2..B} |x_k - x_{k-1}|. Since the 20% masking in Section IV.C is applied only after these times are computed, every masked value has its time covariate (and hence the GP input) depending on the value the imputation model is asked to predict. This is a label-leakage flaw that contaminates the two Synthetic MIMIC rows of Table II and invalidates the Section V.A.1 statement that MixMI is significantly better than the second-best model on all 4 datasets. The real-world MIMIC and NMEDW rows are not affected by this flaw, but the synthetic comparison should be redesigned (e.g., compute times from observed values only) and the affected claims re-stated.","section":"IV.B–IV.C"},{"comment":"The significance claim relies on a permutation test with 1000 replicates performed within a single random 20% mask. With only one mask, the p=.001 can only describe whether the gap between MixMI and the second-best model is unexpected under permutation of that particular test set; it does not control for variability across masks or across data splits. To support the claim that MixMI is significantly better across datasets, the authors should evaluate on multiple independent masks (e.g., 5–10) and report the distribution of MASE scores and a paired test across masks.","section":"V.A.1"},{"comment":"The model assumes a complete time tensor T with t_{p,v,b}=t_{p,u,b} for all laboratory tests v,u for a given patient p at time index b. Raw EHR laboratory results are typically drawn at different clock times for different analytes, and the paper does not describe how admissions are aligned into B common time indices or how timestamps are assigned. If the alignment uses arbitrary or approximated timestamps, the GP component's squared-exponential covariance is evaluated on artificial inputs, and the claimed temporal advantage on the real-world datasets is not established. The authors should either justify the equal-timestamp assumption for the cohorts used or provide a sensitivity analysis with per-variable times.","section":"III.A"}],"minor_comments":[{"comment":"The first author's affiliation is spelled 'Northwester University' in the header; please correct to 'Northwestern University'.","section":"Author affiliations"},{"comment":"The comparison in Section V.C reports an improvement from 0.08351 to 0.07538 without identifying the dataset or setting; these numbers appear to correspond to the Synthetic MIMIC (d=1) row in Table II, and the text should make that explicit.","section":"V.C"},{"comment":"The footnote 'MTGP cannot have multiple inputs, thus not applicable to the synthetic datasets' is unclear, since MTGP is designed for multi-task settings with multiple outputs; please explain precisely why MTGP is inapplicable to the synthetic benchmark.","section":"Table II"},{"comment":"In the MASE definition, the scaling term J_{p,v}/(J_{p,v}-1) is nonstandard; please clarify that the denominator is the mean absolute error of the naive one-step-ahead forecast on the observed part of each series, and describe how series with J_p,v = 1 are handled.","section":"IV.C"}],"recommendation":"major_revision","confidential_remarks":"The synthetic-data leakage is a significant evaluation flaw that could lead readers to overstate the model's superiority. The real-world comparisons are still valuable, and the paper can be salvaged with a redesigned synthetic experiment and multi-mask evaluation. I would encourage the editor to consider major revision rather than rejection, provided the authors rerun the affected experiments and adjust their claims accordingly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The one thing to know: this is a serious imputation paper with a genuinely new modeling idea, but the headline 'all 4 datasets' significance claim rests partly on synthetic benchmarks that leak the true missing value into the time covariate. The real-world results are still worth taking seriously.\n\nWhat is actually new: MixMI mixes two linear regressions (cross-sectional and temporal) with a GP, and learns per-patient mixing weights from the data via EM. That individualized weight mechanism is a real step beyond fixed-weight mixtures like MGP or the ad hoc weighting in 3D-MICE. The model equations are coherent, the EM updates in Appendix A are standard and correct, and the authors compare against a reasonable set of baselines (MICE, M-RNN, GMM, GP, MTGP, 3D-MICE) on two real datasets. Code is public. The MIMIC-III results in Table II show a solid improvement over 3D-MICE (0.093 vs 0.112 MASE), and the variable-wise table shows consistent gains. The classification downstream test ties with GRU-D, which is fine for an imputation paper.\n\nThe soft spots, in order of severity. First, the synthetic data construction in Section IV.B is contaminated. The synthetic time for an observed measurement depends on Δx, which includes the actual measurement value. Masking is applied after these times are computed, so for every masked test point, the time covariate fed to every model is a function of the true value being predicted. Any smooth temporal model can exploit that. This invalidates the two 'Synthetic MIMIC' rows in Table II and the claim of significance on all 4 datasets at p=.001. I checked the equations; the leak is real. Second, the evaluation uses a single random 20% mask per dataset, so we have no sense of variance across masks. The permutation test only resamples within that one mask. Third, the paper assumes that, for a given patient, all lab tests at a time index share the same timestamp, while raw EHR draws are at different times. The alignment procedure is not described, so the GP component's inputs may be partly artificial on the real datasets.\n\nWho this is for: researchers working on EHR imputation or mixture models for time series. The model idea is worth borrowing, and the real-world results suggest MixMI is competitive. It doesn't open a new area, but it's a useful contribution. With the synthetic leakage fixed and repeated-mask variance reported, this could be a solid paper. I'd send it to referee, but the authors should be asked to address the leak and the timestamp assumption.","headline":"A serious imputation method with a genuinely new per-patient weighting idea, but the headline 'all 4 datasets' significance claim rests on synthetic benchmarks that leak the true missing value into the time covariate.","tokens_in":19965,"tokens_out":3556,"would_cite":true,"duration_ms":35737,"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":"A mixture of two linear regressions and a Gaussian process, with per-patient weights, imputes missing clinical time-series values more accurately than six baselines on all four datasets studied.","keywords":["multiple imputation","Gaussian process","mixture model","clinical time series","missing data","electronic health records","tensor imputation","EM algorithm"],"falsifier":"Run MixMI on a real clinical cohort that preserves each laboratory test's actual clock time instead of forcing one shared timestamp per time index, and check whether its MASE advantage over the second-best baseline persists; if the advantage shrinks or reverses, the reported temporal gain depends on the shared-timestamp assumption.","tokens_in":1751,"feed_emoji":"🩺","tokens_out":3293,"duration_ms":98192,"temperature":0.7,"pith_summary":"The paper argues that missing values in multivariable clinical time series are best imputed by a mixture model that uses both cross-sectional information (other tests at the same time) and temporal information (the same test over time), rather than either view alone. It introduces MixMI, which for each variable and time index fits a three-component mixture: two linear regressions and a Gaussian process over time. It also claims that individualized, per-patient mixing weights matter, because the predictive confidence of a Gaussian process varies across patients and along the time series. On two real-world and two synthetic datasets, MixMI produces the lowest mean absolute scaled error, significantly better than the second-best method (p=.001, permutation test with 1000 replicates). A sympathetic reader would care because missing laboratory values are common in clinical records, and more accurate imputation can feed any downstream predictive model.","feed_headline":"MixMI beats six imputation baselines on four datasets","feed_subtitle":"Per-patient blending of cross-sectional and temporal predictors lowers imputation error on every cohort tested.","key_machinery":"The load-bearing device is the individualized mixing weight, defined as $$\\$Pi^{{(k)}}$_{p,v,b} = \\frac{\\hat\\$pi^{{(k)}}$_{v,b} \\mathcal{N}(V_{p,v,b}|\\hat\\$mu^{{(k)}}$_{v,b}, \\hat\\$Sigma^{{(k)}}$_{v,b})}{\\sum_{j=1}^3 \\hat\\$pi^{{(j)}}$_{v,b} \\mathcal{N}(V_{p,v,b}|\\hat\\$mu^{{(j)}}$_{v,b}, \\hat\\$Sigma^{{(j)}}$_{v,b})},$$ where $V_{p,v,b}$ is the patient's observed input vector. This replaces a single mixing weight shared by all patients with a per-patient posterior weight, reflecting the fact that Gaussian-process predictions are more confident for some patients and at some time indices than others. The EM algorithm supplies the optimized responsibilities that these individualized weights are designed to approximate, and an automatic selection step chooses between the two-linear-component variant and the full three-component variant based on training error.","core_discovery":"MixMI represents aligned patient data as a patient-by-test-by-time tensor and, for each variable and time index, trains a mixture of three predictive components: a linear regression on the other tests at the same time, a linear regression on the same test at other times, and a per-patient Gaussian process over time. The components are combined not with fixed weights but with individualized mixing weights computed from each patient's observed inputs, which approximate how much responsibility each component should take for that patient. Imputation proceeds iteratively and multiply: several copies of the tensor are filled with random initial guesses, each copy undergoes repeated passes over every variable and time index, and the imputed values are averaged across copies. The paper reports that MixMI outperforms all comparison models on all four datasets and is significantly better than the second-best model, and that the individualized-weight version beats its fixed-weight counterpart by 9.73% in MASE on the synthetic real-world-derived dataset.","pith_inferences":["A testable extension the paper leaves open is replacing the squared-exponential Gaussian-process kernel with periodic or Matern kernels; longer intensive-care stays often have circadian laboratory rhythms that the current kernel may not capture.","If the shared-timestamp assumption is violated, one could run MixMI on each analyte's actual clock time and treat the change in MASE as a measure of how much of the claimed gain is real temporal signal rather than an artifact of alignment.","The individualized mixing weight could itself be learned as a function of the input, for example through a small neural gating network, instead of being fixed as a posterior ratio; that would let the model adapt across datasets with different missingness mechanisms, though the paper does not test this.","The 30-day mortality experiment ties with a no-imputation recurrent model on AUC, suggesting that the immediate practical payoff may be improved data quality for other tasks rather than a direct classification lift on this particular outcome."],"forward_implications":["Every one of the 13 laboratory variables in the real-world cohort is imputed with lower MASE by MixMI or MixMI-LL than by any comparison model, so the advantage is not driven by a single test.","Because MixMI is a pure imputer, any downstream classifier or regressor can be trained on the completed data; the paper demonstrates parity with a no-imputation recurrent model on a 30-day mortality classification task.","The individualized-weight mechanism generalizes beyond Gaussian processes: any mixture component whose predictive confidence varies per input can be weighted this way.","The automatic choice between the linear-only and full Gaussian-process variants makes the method usable without knowing in advance whether temporal structure is strong enough to justify Gaussian processes.","Multiple copies and iterative passes follow the multiple-imputation paradigm, so the spread across copies can serve as a measure of imputation uncertainty."],"supporting_citations":[{"why":"Supplies the chained-equations multiple-imputation baseline that MixMI is compared against.","marker":"[8]"},{"why":"The prior GP-plus-chained-equations imputer whose static weighting strategy MixMI extends with optimized per-patient weights.","marker":"[41]"},{"why":"The recurrent neural-network imputer that serves as the strongest neural benchmark in the comparison.","marker":"[60]"},{"why":"Multi-task Gaussian process prediction, the multi-output GP baseline on the real-world dataset.","marker":"[35]"},{"why":"Gaussian mixture model imputation, the cross-sectional-only baseline that lacks temporal structure.","marker":"[43]"},{"why":"The no-imputation recurrent classification model used as the downstream comparison for MixMI-imputed data.","marker":"[63]"},{"why":"Supplies the real-world intensive-care data used for the primary evaluation cohort.","marker":"[66]"},{"why":"Defines the mean absolute scaled error used to score all imputation results.","marker":"[69]"}],"fun_headline_variants":["MixMI imputes missing clinical time series with per-patient Gaussian mixtures","Patient-specific mixing weights boost temporal imputation accuracy","Mixture of cross-sectional and temporal models wins on all four datasets","Tensor-based imputation outperforms six baselines on clinical time series","Individualized weights beat fixed weights by 9.73% in MASE"],"cache_read_input_tokens":22144,"weakest_assumption_plain":"The load-bearing premise is that the time tensor is completely known and that, for a given patient, all tests at a time index share the same timestamp; if real laboratory draws occur at different times per analyte, the Gaussian-process component is trained on artificial spacings.","fun_headline_variants_meta":{"raw":{"variants":["MixMI imputes missing clinical time series with per-patient Gaussian mixtures","Patient-specific mixing weights boost temporal imputation accuracy","Mixture of cross-sectional and temporal models wins on all four datasets","Tensor-based imputation outperforms six baselines on clinical time series","Individualized weights beat fixed weights by 9.73% in MASE"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000857,"raw_usage":{"total_tokens":3666,"prompt_tokens":835,"completion_tokens":2831,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":451,"completion_tokens_details":{"reasoning_tokens":2741}},"tokens_in":451,"tokens_out":2831,"duration_ms":19643,"temperature":1.0,"reasoning_tokens":2741,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:47:39.565847+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run MixMI on a real clinical cohort that preserves each laboratory test's actual clock time instead of forcing one shared timestamp per time index, and check whether its MASE advantage over the second-best baseline persists; if the advantage shrinks or reverses, the reported temporal gain depends on the shared-timestamp assumption.","supporting_citations":[{"cited_title":"mice: Multivariate imputation by chained equations in r,","cited_arxiv_id":null,"evidence_quote":"Supplies the chained-equations multiple-imputation baseline that MixMI is compared against."},{"cited_title":"3d-mice: inte- gration of cross-sectional and longitudinal imputation for multi-analyte longitudinal clinical data,","cited_arxiv_id":null,"evidence_quote":"The prior GP-plus-chained-equations imputer whose static weighting strategy MixMI extends with optimized per-patient weights."},{"cited_title":"Multi-task gaussian process prediction,","cited_arxiv_id":null,"evidence_quote":"Multi-task Gaussian process prediction, the multi-output GP baseline on the real-world dataset."},{"cited_title":"Imputation through finite gaussian mixture models,","cited_arxiv_id":null,"evidence_quote":"Gaussian mixture model imputation, the cross-sectional-only baseline that lacks temporal structure."},{"cited_title":"Recurrent neural networks for multivariate time series with missing values,","cited_arxiv_id":null,"evidence_quote":"The no-imputation recurrent classification model used as the downstream comparison for MixMI-imputed data."},{"cited_title":"Mimic-iii, a freely accessible critical care database,","cited_arxiv_id":null,"evidence_quote":"Supplies the real-world intensive-care data used for the primary evaluation cohort."},{"cited_title":"Another look at measures of forecast accuracy,","cited_arxiv_id":null,"evidence_quote":"Defines the mean absolute scaled error used to score all imputation results."}],"review_version":1}