{"id":"1309c94c-af4c-454c-9084-397370793db3","arxiv_id":"2506.07092","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"A patient-similarity pipeline using Spark-based DTW and target-aware aWOE feature transformation is claimed to improve CAD and CHF prediction, but the experimental design leaks label information into the features.","lead":"This paper combines dynamic time warping on vital-sign time series with clustering on demographic data to find similar patients, and runs the computation on Spark to make it faster. The authors report big gains in predicting coronary artery disease and congestive heart failure, but the gains appear to come partly from using the disease labels themselves to transform and cluster the data.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The most load-bearing risk is target leakage through aWOE: Eq. (1) requires a binary 'positive/negative events' outcome, and the paper never states this outcome differs from the CAD/CHF label being predicted; if it is the same label, clustering on aWOE-transformed static data leaks the target into…","rationale":"The central claim is empirical: the proposed DT-based DPSC substantially improves patient similarity prediction for CAD and CHF. For that claim to hold, the comparison must attribute the gains to the method rather than to label information entering feature construction. The reader's weakest assumption identifies exactly this risk, and the manuscript text does not close it. Eq. (1) computes aWOE from positive and negative event distributions, which requires a binary outcome; Section 2.3 describes clustering on static data that includes historical clinical information while only 'disregarding' the target attribute. Nothing states that the binary outcome used for aWOE is different from the disease being predicted. If the target is used, the transformed features are label-derived, and cluster membership becomes a target-propensity grouping. The nearest-neighbor step then selects within groups that already encode the outcome, which can inflate every reported metric and explain the large gap between aWOE-based and No-DT methods. The GitHub repository is a real opportunity: a direct inspection of the aWOE call can settle whether the target is used. The Holm table's mislabeled third row and the tuning of cluster count on test data are additional problems, but they are secondary; they do not carry the same explanatory weight as the possible leakage. If the code shows the aWOE outcome is the target, the reported performance should not be trusted as evidence for the method. If the code shows a different outcome or a nested fitting procedure, the concern would be resolved and the remaining issues would be reporting and evaluation rigor rather than invalidation of the core idea. Since the reader already reached REJECT on essentially this basis, my stress-test does not change the verdict.","tokens_in":20429,"tokens_out":4709,"duration_ms":52343,"concrete_test":"Inspect the code at github.com/joysana1/PSC to identify the label vector passed to the aWOE routine. Then rerun the CAD and CHF pipelines with aWOE fit only on the training split's labels, and also with a shuffled placebo label. If AUC/F-measure drops from the reported 0.858/0.896 and 0.878/0.816 toward the No-DT rows of Table 3 (about 0.6 AUC), the reported improvements are target-leakage artifacts. If performance remains stable under both conditions, the leakage concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.2.1 defines aWOE = ln((Distribution of positive events + epsilon) / (Distribution of negative events + epsilon)) without specifying which binary outcome defines 'positive events.' Section 2.2 says transformations are applied to static data before clustering, and Section 2.3 says historical clinical attributes (CAD/CHF) are used in clustering, with only the target attribute itself 'disregarded.' If the disease being predicted is the outcome used to compute aWOE, then each static feature is replaced by a value summarizing the conditional distribution of the target, and clusters become groups of patients with similar target propensity. Nearest neighbors are then drawn from clusters already separated by label information, so the reported gains (AUC 0.858/0.878, F-measure 0.896/0.816 vs. No-DT ~0.6 AUC) could be produced without genuine similarity learning. The cross-disease use of CHF for CAD prediction and vice versa is a weaker, possibly legitimate comorbidity channel, but the aWOE route is decisive because it directly encodes the outcome. The SHAP finding that medical records dominate prediction is consistent with such label-encoded features. The paper must specify the outcome used in Eq. (1) and ensure it is not the predicted disease; the current text does not rule out direct leakage.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a distributed patient similarity computation (DPSC) method that combines dynamic time warping (DTW) on 18 time series variables with clustering on static patient data. Before clustering, the static data are transformed with either adaptive Weight-of-Evidence (aWOE) or Z-score. For a target patient, the nearest neighbor is selected from the same cluster using a fusion of per-variate DTW distance matrices, and the neighbor's label is used for prediction. The method is evaluated on MIMIC-III for predicting Coronary Artery Disease (CAD) and Congestive Heart Failure (CHF), with claims of large improvements over non-transformed baselines and a 40% runtime reduction from Spark-based distributed DTW. The central claim is that the aWOE transformation and the clustering step improve patient similarity prediction.","tokens_in":20618,"tokens_out":3977,"duration_ms":38221,"significance":"If the reported results are valid, the paper would offer a practical distributed approach to patient similarity that integrates static and time series data with a privacy-oriented transformation, and it would provide a useful benchmark for CAD and CHF prediction on MIMIC-III. The use of a public dataset, the comparison against several baselines, and the availability of code are strengths. However, the significance hinges entirely on whether the aWOE transformation and clustering are free of target leakage; the current manuscript does not establish this, so the headline improvements cannot be taken at face value.","major_comments":[{"comment":"Equation (1) defines aWOE using the 'Distribution of positive events' and 'Distribution of negative events' but does not specify which binary outcome defines these events. Since Section 2.1.1 lists Coronary Artery Disease and Congestive Heart Failure among the static features, and Section 2.3 uses these disease labels as clustering features, the aWOE transformation may be target-encoding the very outcome being predicted. If the outcome in Eq. (1) is the same as the prediction target, then the clusters from which nearest neighbors are drawn are constructed with label information, and the performance gaps in Table 3 (e.g., aWOE K-means CAD AUC 0.858 vs. No-DT K-means 0.605) would be artifacts of leakage rather than evidence of genuine similarity learning. The manuscript must specify, for each prediction task, the outcome used to compute aWOE and confirm that this outcome is not the target or a direct proxy for it.","section":"Section 2.2.1, Eq. (1)"},{"comment":"The text states that for CHF prediction, the Coronary Artery Disease attribute is included as a clustering feature, and for CAD prediction, Congestive Heart Failure is included, with only the target attribute itself disregarded. Because CAD and CHF are strongly correlated cardiovascular conditions, using the other disease label as a clustering feature is a probable channel of label leakage: the cluster assignment, and hence the available pool of nearest neighbors, is influenced by a label that is highly predictive of the target. The authors should report the correlation between the two labels and evaluate a variant that excludes both disease labels from the static feature set used for clustering.","section":"Section 2.3"},{"comment":"Section 3.2 describes selecting the number of clusters K based on the average performance over five runs and reports optimal values (K=125 for CAD, K=150 for CHF), but it does not clarify whether this selection used the test split. If the test set was used to choose K, the reported peak performance is optimistically biased. Moreover, Section 2.9 describes an 80/20 train/test split but does not state whether the aWOE binning and the clustering are fitted on the training fold only; if the transformation or clustering uses the full dataset, including test patients, then test labels and feature distributions leak into neighbor selection. Please clarify the validation protocol and, if needed, re-run the evaluation with the selection and fitting steps strictly inside the training folds.","section":"Section 3.2 and Section 2.9"},{"comment":"The Holm post-hoc table lists the comparison 'aWOE based DPSC vs. Z-score based DPSC' with P-value 0.133614 and marks the hypothesis as 'Rejected', which is inconsistent because 0.133614 exceeds the adjusted alpha of 0.05. In addition, the significance level is given as α=0.5 in the narrative text of Section 4 while Section 2.12 states α=0.05. These inconsistencies undermine the statistical-significance claim and should be corrected.","section":"Section 4, Table 5"}],"minor_comments":[{"comment":"The symbol K is used both for the number of clusters and for the number of univariate time series (K=18), which is confusing; please use distinct symbols (e.g., C for clusters and V for variates).","section":"Section 2.7"},{"comment":"For CAD with No-DT OPTICS, specificity is reported as 0.81 but recall as 0.375, yielding an F-measure of 0.503; a brief explanation of this imbalance would help the reader interpret the row.","section":"Table 3"},{"comment":"The abstract claims computation time reduction 'by as high as 40%', but Figures 11–13 compare runtime only for DTW distance computation, not for the full DPSC pipeline; please specify the scope of the 40% claim.","section":"Abstract and Section 3.4"},{"comment":"The fusion function F in Eq. (5) is not defined formally; please state how the K distance matrices are combined (e.g., averaging, max, or weighted sum).","section":"Section 2.7, Eq. (5)"}],"recommendation":"reject","confidential_remarks":"The central issue is the unresolved target-leakage risk through aWOE (Eq. 1) and through the cross-disease clustering features in Section 2.3. If the authors can later demonstrate that the outcome used for aWOE is distinct from the prediction target and that clustering is performed without any target-related labels, the paper could be revisable. As written, the reported performance improvements cannot be distinguished from leakage artifacts, so I cannot recommend publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe headline here is that the reported gains for CAD and CHF prediction are probably an artifact of target leakage. Eq. (1) defines aWOE with 'positive/negative events' but never says which binary outcome is used. If it's the disease being predicted, then the static features fed to clustering are target-encoded, and the clusters separate patients by outcome propensity before any nearest-neighbor vote. Section 2.3 also uses the other cardiovascular disease as a clustering feature for each task, which is a correlated-label channel. The text explicitly says only the target attribute itself is disregarded, so nothing rules out direct leakage. That is load-bearing, because the 10-20% absolute improvements over the No-DT baselines are exactly what you'd expect from label-encoded features.\n\nWhat the paper does well: the Spark-based distributed DTW is a real engineering contribution, the 12-hour observation window analysis is useful, the code is on GitHub, and the baseline sweep including Transformer and LSTM is thorough. If the leakage were resolved, the runtime reduction from distributed DTW (up to 40%) would be citable on its own.\n\nThe other soft spots are minor but real. The cluster count was selected by testing K on the evaluation set (Section 3.2), the results come from a single 80/20 split with no error bars, and the Holm table in Section 4 marks a non-significant comparison (p=0.134 vs adjusted alpha 0.05) as 'Rejected'. The privacy claim leans on self-cited work with no independent evidence.\n\nOverall: the engineering may be salvageable, but the central empirical claim is not credible as presented. The authors need to state the outcome used in Eq. (1), re-run with a proper validation split for any tuning, and report confidence intervals. If the aWOE outcome turns out to be the predicted disease, the main results collapse. I would not accept this version; I would send it back with these demands, and if the authors can't rule out leakage, reject. Still, the Spark DTW piece deserves a serious referee's attention, which is why I'd let it go to review rather than desk-reject without comment.\n\nFor you: probably not a cite, but a useful teaching example of target encoding sneaking into a similarity pipeline.","headline":"The reported CAD/CHF gains are likely label leakage through aWOE; the Spark DTW engineering is real but the central result is not credible as presented.","tokens_in":21271,"tokens_out":4644,"would_cite":false,"duration_ms":43715,"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":"Similarity-based heart-disease prediction improves when static data are transformed before clustering.","keywords":["patient similarity","clinical decision support","dynamic time warping","adaptive weight-of-evidence","data transformation","distributed computing","cardiovascular disease prediction"],"falsifier":"Re-run the aWOE-K-means pipeline twice, once with aWOE bins computed from the true disease label and once with bins computed from a shuffled binary label assigned randomly to the same patients; if the AUC and F-measure gains shrink or disappear in the shuffled case, the transformation's reported benefit is attributable to target leakage rather than to the transformation itself.","tokens_in":20121,"feed_emoji":"🫀","tokens_out":8580,"duration_ms":79913,"temperature":0.7,"pith_summary":"The paper proposes a distributed patient-similarity pipeline that predicts coronary artery disease and congestive heart failure by combining two information sources: static demographic and history features, used to cluster patients, and 18 physiological time series, compared within clusters using Dynamic Time Warping. The paper's central move is to apply data transformations—adaptive Weight-of-Evidence (aWOE) and Z-score—to the static features before clustering. The authors report that this transformation step raises AUC, accuracy, and F-measure by up to 11.4%, 10.20%, and 12.6% for coronary artery disease and up to 15.9%, 10.5%, and 21.9% for congestive heart failure relative to the same pipeline without transformation, while a distributed execution environment cuts computation time by up to 40%. A sympathetic reader should care because this suggests a cheap preprocessing step can make similarity-based diagnosis markedly more accurate and fast enough for real-time clinical decision support.","feed_headline":"Patient-similarity heart predictions gain up to 21.9%","feed_subtitle":"Transforming static features before clustering lifts AUC, accuracy, and F-measure on CAD and CHF.","key_machinery":"The load-bearing object is the neighborhood similarity fusion: for each time-series variate, a DTW distance matrix is computed between a target patient and patients in the same static-data cluster, the per-variate nearest neighbors are unioned into a similarity fusion, and the majority label among those neighbors is assigned. The other named component is adaptive Weight-of-Evidence (aWOE), a binning transform that replaces each static feature value with the log-ratio of positive to negative outcome counts within its bin, plus a small constant; the paper argues this transformation both improves cluster quality and obscures raw patient values, and it reports that aWOE consistently beats Z-score and no-transformation variants.","core_discovery":"The central claim is that the gain in patient-similarity prediction comes from the data transformation, not from the clustering or the time-series distance alone. With the aWOE-K-means variant, the model reaches an AUC of 0.858, accuracy of 0.870, and F-measure of 0.896 for coronary artery disease, and an AUC of 0.878, accuracy of 0.887, and F-measure of 0.816 for congestive heart failure; without any transformation the same pipeline drops to AUC values near 0.61 for both diseases. The paper further reports that a 12-hour observation window recovers almost all of the performance of the full-length time series (F-measure differences of 3.13 and 4.12 for the two diseases), and that adding computing nodes reduces the DTW computation time, making the method viable for real-time use.","pith_inferences":["Editorial inference: the aWOE scores in Eq. (1) are computed from distributions of positive and negative events, and the paper does not state that this outcome label is different from the disease being predicted; if the same label feeds the transformation and the clustering features (CHF appears as a clustering feature for CAD and vice versa), part of the reported gain could reflect target leakage","Editorial inference: the privacy claim rests on the aWOE binning being a generalization-based anonymity mechanism; a concrete re-identification attack or a k-anonymity audit on the transformed static matrix would test that guarantee directly.","Editorial inference: the 40% runtime reduction is measured on one cluster configuration; the released code would allow a reader to check whether the speedup transfers to longer time series or larger cohorts."],"forward_implications":["If the transformation result holds, a simple preprocessing change—aWOE on static features—can lift the accuracy of similarity-based CAD and CHF prediction above all tested baselines, including LSTM and Transformer models.","The 12-hour result implies that intensive-care monitoring windows of half a day are enough for near-optimal patient matching, which matters for emergency settings where full histories are unavailable.","The reported runtime scaling, with the full pipeline taking roughly 53 seconds per target patient on a 100-patient cohort, implies distributed DTW can meet real-time decision-support demands as node counts grow.","Because the neighborhood fusion ignores a variate when a patient's record is missing it, the method has a built-in tolerance for missing time-series data."],"supporting_citations":[{"why":"supplies the 4,418 cardiovascular patient records and the 18 physiological time series used in every experiment.","marker":"[39]"},{"why":"defines the neighborhood-population fusion and the static-plus-time-series similarity setup that the proposed pipeline extends.","marker":"[34]"},{"why":"introduces the adaptive Weight-of-Evidence transformation and its claimed privacy-preserving property.","marker":"[43]"},{"why":"provides prior evidence that aWOE and Z-score data transformations improve prediction performance.","marker":"[25]"},{"why":"is the source of the baseline feature extraction (mean, standard deviation, median, skewness, median absolute deviation) used to build the comparison classifiers.","marker":"[16]"},{"why":"defines Dynamic Time Warping, the distance used for all time-series similarity computations.","marker":"[4]"}],"fun_headline_variants":["Heart predictions gain up to 21.9% via patient similarity","Data transformation boosts patient similarity by up to 22%","Transform patient data to improve heart disease predictions","Patient similarity method cuts runtime 40% and boosts accuracy","Data transformation is key to patient similarity gains"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the binary outcome used to compute the aWOE transformation and the disease labels used as clustering features are not the same as, or strongly correlated with, the disease being predicted; if they are, the reported improvements could come from target information leaking into the features rather than from better similarity learning.","fun_headline_variants_meta":{"raw":{"variants":["Heart predictions gain up to 21.9% via patient similarity","Data transformation boosts patient similarity by up to 22%","Transform patient data to improve heart disease predictions","Patient similarity method cuts runtime 40% and boosts accuracy","Data transformation is key to patient similarity gains"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000601,"raw_usage":{"total_tokens":2873,"prompt_tokens":1077,"completion_tokens":1796,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":693,"completion_tokens_details":{"reasoning_tokens":1719}},"tokens_in":693,"tokens_out":1796,"duration_ms":16048,"temperature":1.0,"reasoning_tokens":1719,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:42:08.919543+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the aWOE-K-means pipeline twice, once with aWOE bins computed from the true disease label and once with bins computed from a shuffled binary label assigned randomly to the same patients; if the AUC and F-measure gains shrink or disappear in the shuffled case, the transformation's reported benefit is attributable to target leakage rather than to the transformation itself.","supporting_citations":[{"cited_title":"Physionet-MIMICIII","cited_arxiv_id":null,"evidence_quote":"supplies the 4,418 cardiovascular patient records and the 18 physiological time series used in every experiment."},{"cited_title":"Effective patient similarity computa- tion for clinical decision support using time series and static data, pp","cited_arxiv_id":null,"evidence_quote":"defines the neighborhood-population fusion and the static-plus-time-series similarity setup that the proposed pipeline extends."},{"cited_title":"Privacy-Preserving Customer Churn Prediction Model in the Context of Telecommunication Industry","cited_arxiv_id":"2411.01447","evidence_quote":"introduces the adaptive Weight-of-Evidence transformation and its claimed privacy-preserving property."},{"cited_title":"A novel customer churn prediction model for the telecommunication industry us- ing data transformation methods and feature selection","cited_arxiv_id":null,"evidence_quote":"provides prior evidence that aWOE and Z-score data transformations improve prediction performance."},{"cited_title":"Tsiklidis, Talid Sinno, S.L.D., 2022","cited_arxiv_id":null,"evidence_quote":"is the source of the baseline feature extraction (mean, standard deviation, median, skewness, median absolute deviation) used to build the comparison classifiers."},{"cited_title":"Using dynamic time warping to find patterns in time series, in: In Pro- ceedings of the 3rd International Conference on Knowledge Discovery and Data Mining (AAAIWS’94), p","cited_arxiv_id":null,"evidence_quote":"defines Dynamic Time Warping, the distance used for all time-series similarity computations."}],"review_version":1}