{"id":"0cff6df3-79cb-4285-bccc-87225b367ad9","arxiv_id":"2505.07702","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"4TaStiC, a dissimilarity that shifts time and tilts trends for time series, outperforms seven baselines on synthetic data and separates 1,989 type 2 diabetes patients into seven HbA1c trajectory groups.","lead":"Researchers built a new way to compare patient lab-result histories: slide the time points and slightly rotate the trend lines to find matches, then cluster on both levels and patterns. It beat seven existing methods on artificial datasets and split 1,989 diabetes patients into seven groups with distinct HbA1c trajectories.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unpenalized trend traveling can merge a stable HbA1c series with a steadily increasing one, so the seven Siriraj groups may reflect tilt artifacts rather than true clinical trajectories.","rationale":"The reader's weakest assumption identifies the unpenalized trend traveling as the key risk, and I agree. The concern is not merely that hyperparameters were tuned on test data—that is a standard methodological flaw that can be fixed by held-out evaluation. The deeper issue is that the dissimilarity measure itself can equate clinically different trajectories. The mathematical mechanism is clear from Eq. (7) with C=0: the correlation term is 1-Corr of the tilted series, and the Euclidean term is un-tilted but down-weighted when alpha is near 1. With a cumulative tilt of 0.825, a flat series can be rotated to match a series with a 0.07 per-visit increase. Since such trend differences are common and clinically relevant in diabetes, the method may over-merge. The suggested simulation directly tests whether the method separates stable from increasing patients under the exact parameters used in Section 5.2. If it does not, the Siriraj seven-group solution is suspect. The robustness re-run with penalized or smaller tilt would reveal whether the real-data clustering is an artifact. These concerns do not change the reader's conditional verdict; they sharpen the conditions needed for acceptance.","tokens_in":16664,"tokens_out":14866,"duration_ms":141536,"concrete_test":"Using the released FourTaStiC R package, simulate 20 series: 10 stable at HbA1c approximately 7.0 with small noise and 10 increasing linearly from 7.0 to 7.8 (slope about 0.07) with small noise. Run 4TaStiC with L=3, E={-0.075,0,0.075}, C=0, alpha from Eq. (8) with p=0.09, and average-linkage hierarchical clustering. If the ARI between predicted and true labels is below 0.8, the tilt merges clinically distinct trajectories. As a robustness check, re-run the same simulation with C=0.5 and E={-0.025,0,0.025}; if the ARI improves substantially, the original parameters are over-tilting.","verdict_should_be":"UNCHANGED","load_bearing_attack":"With C=0, the trend traveling term in Eq. (7) imposes no penalty on tilting. The chosen E={-0.075,0,0.075} and T=12 allow a cumulative rotation of up to 0.825 HbA1c percentage points across the series. Because alpha from Eq. (8) is data-dependent and likely close to 1 (the max Euclidean distance dominates the denominator), the dissimilarity is governed almost entirely by 1-Corr after tilt; the un-tilted Euclidean term in Eq. (7) has little weight. Consequently, a stable series at 7.0 tilted by +0.075 per step becomes nearly perfectly correlated with a series linearly increasing from 7.0 to 7.8 (slope 0.07). The paper's Section 4.1 justifies the tilt range by a maximum acceptable difference of 1.5, but a cumulative tilt of 0.825 is not a level shift—it changes the trajectory shape. Two patients with the same mean but different slopes (stable vs. deteriorating) can be merged. This directly threatens the claim in Section 5.2 that the seven groups reflect clinically meaningful HbA1c trajectories, because the clusters may be artifacts of the rotation. Section 6.1 itself warns that trend traveling is 'sensitive to small changes' and requires that the rotation be clinically meaningful.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes 4TaStiC, a dissimilarity measure for time series clustering that aligns series by allowing small time shifts (time traveling) and small rotations of the trend (trend traveling), then combines a weighted Euclidean distance with a Pearson-correlation dissimilarity (Eqs. (1)-(8)). The authors prove three basic properties, evaluate the method on nine synthetic datasets against seven existing methods, and apply it to cluster 1,989 type 2 diabetes patients from Siriraj Hospital into seven groups summarized in Table 7 and Figure 6. An R package is provided on GitHub.","tokens_in":17026,"tokens_out":8474,"duration_ms":78949,"significance":"If the claims are validated, 4TaStiC is a useful and interpretable dissimilarity for longitudinal laboratory data with irregular visit timing, and the paper contributes small proofs, open-source code, and a real clinical case study. The central idea of combining time shifts with a controlled trend tilt is clearly presented and can be plugged into any distance-based clustering algorithm. However, the current evaluation does not yet establish the headline performance or the clinical validity of the seven groups, because of the in-sample parameter selection and the unpenalized trend traveling discussed below.","major_comments":[{"comment":"The hyperparameters p=0.09, epsilon=0.075, and C=0 are selected in Section 4.1 by maximizing or balancing accuracy and ARI on the three G3 datasets, and the same datasets are then used in Table 5 to report the Class 3 performance of 4TaStiC. This makes the headline claim of superior performance on Class 3 an in-sample result, not evidence of generalization. Moreover, the synthetic Class 3 datasets are constructed to contain the time-shift and trend-tilt structures that 4TaStiC is designed to exploit, so a comparison against methods not designed for those structures is inherently optimistic. Please provide a nested or held-out evaluation, for example tuning on one synthetic class and reporting Table 5 on the remaining datasets, or reporting results for all 45 parameter combinations across all nine datasets; without this, the numbers in Table 5 do not support the stated conclusion that 4TaStiC outperforms the compared methods.","section":"4.1-4.2"},{"comment":"With C=0 in Eq. (7), the penalty term e^{-C|epsilon|} equals 1, so the correlation term after tilting is unpenalized. Because Eq. (8) with p=0.09 can make alpha close to 1, the Euclidean term is downweighted by the factor (1-alpha), and the dissimilarity is then dominated by 1-Corr after tilting. A stable series at 7.0 and a linearly increasing series from 7.0 to 7.8 become nearly perfectly correlated after tilting the stable series by 0.075 per step (a cumulative tilt of 0.825 over 11 steps, within the 1.5 range stated in Section 4.1), so the two can be merged even though their trajectories are clinically distinct. This directly threatens the interpretation in Section 5.2 that the seven groups reflect meaningful HbA1c trajectories. Section 6.1 itself warns that trend traveling is 'sensitive to small changes' and must be clinically meaningful. Please report the actual alpha values for the synthetic and real datasets, the proportion of pairs for which the optimal epsilon is nonzero, and a comparison of the real-data clustering with E={0} or with C>0, including the ARI between the resulting partitions, to show that the seven groups are not created by the unpenalized tilt.","section":"3.3, 4.1, 5.2"}],"minor_comments":[{"comment":"The proof states that dCorr((a,0,0),(b,0,0))=0 for any nonzero a≠b; this is only true when a and b have the same sign, so the statement or the example should be corrected.","section":"3.5, proof of Proposition 3.2"},{"comment":"The list of predetermined parameters (alpha, L, E, C) omits p from Eq. (8) and the number of clusters K, both of which are user-chosen in the experiments; please make the parameter list complete.","section":"2.1"},{"comment":"The body refers to the weighted Euclidean distance 'defined in Table 2', but the formula is not reproduced in the text I reviewed; please include it explicitly so that Eq. (7) is self-contained.","section":"2.4"},{"comment":"The sentence 'With 12 time steps and as in (4), this rotates the last time point by 11×epsilon' is imprecise: Eq. (4) is the trend-traveling minimization, while the rotation is actually given by Eqs. (1)-(2) with l=0; please clarify.","section":"4.1"},{"comment":"The warning that trend traveling is 'sensitive to small changes' appears only in the conclusion; it should be moved to the sensitivity analysis and explicitly connected to the choice C=0.","section":"6.1"}],"recommendation":"major_revision","confidential_remarks":"The main risk is that the headline Class 3 performance is an artifact of tuning on the test sets. After re-evaluation, the methodological contribution may still be publishable, but the clinical application needs at least a stability analysis and a comparison with the unpenalized tilt case. I would not reject outright, but I would not accept without the requested re-analysis."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the one thing to know: this paper introduces a genuinely new dissimilarity measure that rotates one time series by a small angle before computing Pearson correlation—the \"trend traveling\" step—and combines it with the familiar time-shifting trick. That idea does not appear in the cited comparison methods, and the paper ships R code and applies the method to 1,989 real diabetes patients from Siriraj Hospital. The math is clean, the small propositions are correct, and the synthetic benchmark covers three difficulty classes with a sensible imitated-patient structure. The code availability is a real plus.\n\nThe soft spots are equally clear. The biggest is the evaluation protocol: the authors select p=0.09, epsilon=0.075, C=0 by maximizing accuracy and ARI on the same three G3 datasets that later appear in the performance table. That is circular. The numbers in Table 5 would be believable only with a train/test split or a nested tuning procedure that reports out-of-sample performance. As written, the claim of \"superior performance\" is overstated.\n\nSecond, the stress-test concern is not a straw man. With C=0, the penalty term vanishes, so a series can be tilted by up to 0.075 per step over 11 steps, a cumulative 0.825 HbA1c points. A stable 7.0 series tilted upward becomes nearly perfectly correlated with a series climbing from 7.0 to 7.8. The authors themselves warn in Section 6.1 that trend traveling is \"sensitive to small changes\" and that the rotation must be clinically meaningful. That warning undercuts their own parameter choice. The seven Siriraj groups may partly reflect tilt artifacts rather than true trajectory groups. This does not invalidate the method, but it does mean the clinical story needs validation—for example, a comparison with clusters from a standard correlation distance on the same cohort, and a check on whether the groups differ on hard outcomes beyond the descriptive retinopathy percentages.\n\nThird, the real-data analysis is exploratory and descriptive. There is no baseline comparison on the Siriraj data, and the retinopathy rates are not adjusted for confounders. That is acceptable for a preliminary study, but the claims about direct clinical benefit are premature.\n\nOverall, this is a promising manuscript, not a desk reject. I would send it to peer review with a clear request to fix the tuning protocol and to test whether the trend-traveling tilt, at the chosen magnitude, actually preserves clinically meaningful group separation. The core dissimilarity is novel, the writing is readable, and the application is relevant. A serious referee could move this from conditional to solid with moderate revision.","headline":"Novel trend-traveling dissimilarity with a clean real-data application, but the headline numbers are inflated by tuning on the test set and the unpenalized tilt can merge clinically distinct trajectories.","tokens_in":17477,"tokens_out":3220,"would_cite":false,"duration_ms":30817,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H30","62M10","92C50"],"pacs":[],"model":"deepseek-v4-flash","headline":"4TaStiC defines a dissimilarity that shifts and tilts time series, and the paper reports that it outperforms seven existing methods on the hardest synthetic medical-like datasets and splits 1,989 diabetes patients into seven clinically…","keywords":["time series clustering","HbA1c","type 2 diabetes","hierarchical clustering","cross-correlation","dissimilarity measure","trend traveling","time traveling"],"falsifier":"A decisive check is to rerun 4TaStiC with the trend-traveling tilt disabled ($E=\\{0\\}$) on the Class 3 datasets and on the real cohort; if the accuracy and ARI gains, or the seven-group structure, largely remain, then trend traveling is not the source of the claim, and if they vanish, the reported clusters rest on the unpenalized rotation.","tokens_in":16491,"feed_emoji":"📈","tokens_out":11618,"duration_ms":103019,"temperature":0.7,"pith_summary":"The paper introduces a dissimilarity measure, 4TaStiC, that computes the distance between two patient time series by allowing them to be shifted in time and tilted in trend before taking the best match. The intended contribution is that this captures both the level and the trajectory of HbA1c records, which ordinary distances and correlation measures miss when patients visit at irregular times. On the hardest synthetic datasets, designed to imitate laboratory records, the paper reports that this measure outperforms seven existing methods, and that clustering a cohort of 1,989 type 2 diabetes patients produces seven groups with coherent HbA1c levels, trends, and complication rates. If true, the method offers a route to automatic patient stratification that could guide treatment and monitoring recommendations without reviewing every record.","feed_headline":"Time and trend traveling beats 7 methods on hard time-series clusters","feed_subtitle":"Aligning visits and tilting trends lets HbA1c clusters reflect both glucose level and behavior.","key_machinery":"The central object is the 4TaStiC dissimilarity $d_{L,E}(X,Y)=\\min_{l\\in\\{0,\\ldots,L\\},\\,\\epsilon\\in E}\\;\\{d(X^{(l-,0)},Y^{(l+,\\epsilon)}),d(X^{(l+,0)},Y^{(l-,\\epsilon)})\\}$, where $d$ blends the weighted Euclidean distance with the Pearson-correlation dissimilarity, the latter penalized by $e^{-C|\\epsilon|}$ for large tilts. The time shift $l$ realigns visits that fall on different calendars; the tilt $\\epsilon$ rotates one series so that nearly parallel trends can be recognized as the same pattern. Taking the best match over all allowed shifts and tilts is the mechanism that lets level and trajectory be considered in one number, and it is what changes the dendrogram in the paper's illustrative example.","core_discovery":"The paper's central claim is that replacing ordinary pairwise distances with the time-and-trend-traveling dissimilarity is enough to cluster long, irregularly sampled laboratory time series into interpretable groups. On three synthetic datasets designed to imitate the hardest features of patient records, 4TaStiC with both time and trend traveling reaches accuracies of 0.97, 0.90 and 0.98 and adjusted Rand indices of 0.93, 0.84 and 0.93, which are the best among the eight methods compared. Applied to 1,989 type 2 diabetes patients, it separates seven groups whose HbA1c levels, recent trends, and diabetic-retinopathy rates are coherent, such as a stable group, a recovering group, and an extremely high increasing group. The intended upshot is that a doctor could be shown each patient's group label and trajectory summary instead of being asked to review the full sequence of records.","pith_inferences":["Editorial inference: with C=0, trend traveling is equivalent to correlating after an unpenalized linear detrending of one series; an ablation that clusters the detrended series under the same time shifting would reveal whether the tilt parameter adds anything beyond that.","Editorial inference: the retinopathy rates in the paper imply a prospective test: if trajectory groups are medically meaningful, the increasing-trend groups should show higher new complication rates in later follow-up data, which the paper does not provide.","Editorial inference: the method should transfer to other intermittently measured biomarkers such as blood pressure or kidney function whenever both level and trajectory matter; the paper asserts this applicability but does not demonstrate it outside HbA1c."],"forward_implications":["If the synthetic results carry over, 4TaStiC becomes a candidate default for clustering irregularly sampled medical time series, because it handles level and trajectory in one dissimilarity rather than in two separate steps.","The seven-group partition would let clinicians act on trajectory information: patients in increasing-trend groups can be flagged for closer monitoring, while stable and recovering groups can be given maintenance advice.","Because the method only produces a dissimilarity matrix, the same measure can be inserted into any distance-based clustering algorithm, so its usefulness is not tied to the particular hierarchical clustering used in the paper.","The paper's own sensitivity table shows the method is stable only for a limited range of the tilt parameter, so users of unpenalized trend traveling need to check that the chosen epsilon matches their clinical tolerance for what counts as the same trajectory."],"supporting_citations":[{"why":"Supplies the lag-penalized weighted-correlation baseline, the closest prior method to 4TaStiC, which the comparison must beat.","marker":"[29]"},{"why":"Supplies the global-alignment-kernel baseline used in the algorithm comparison.","marker":"[27]"},{"why":"Supplies the dynamic-time-warping baseline for handling shifted time points.","marker":"[26]"},{"why":"Supplies the cross-correlation dissimilarity baseline for correlation-based clustering.","marker":"[28]"},{"why":"Supplies the adjusted Rand index used to score every clustering result, including the reported accuracy and ARI values.","marker":"[42]"},{"why":"Supports the idea of combining correlation-based and distance-based dissimilarities into one base measure.","marker":"[34]"}],"fun_headline_variants":["4TaStiC: time-travel clustering beats 7 on tricky time series","Trend-traveling clustering wins on irregular lab time series","New clustering aligns visits and trends to group diabetes patients","4TaStiC: clustering that time-travels and trend-matches","Cluster long-term diabetes data with time and trend travel"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that letting one patient's HbA1c curve be tilted up or down by up to 0.825 percentage points, with no penalty, to force a better correlation only creates matches that are clinically meaningful, not spurious.","fun_headline_variants_meta":{"raw":{"variants":["4TaStiC: time-travel clustering beats 7 on tricky time series","Trend-traveling clustering wins on irregular lab time series","New clustering aligns visits and trends to group diabetes patients","4TaStiC: clustering that time-travels and trend-matches","Cluster long-term diabetes data with time and trend travel"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000535,"raw_usage":{"total_tokens":2605,"prompt_tokens":1010,"completion_tokens":1595,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":626,"completion_tokens_details":{"reasoning_tokens":1515}},"tokens_in":626,"tokens_out":1595,"duration_ms":10986,"temperature":1.0,"reasoning_tokens":1515,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:09:46.660919+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive check is to rerun 4TaStiC with the trend-traveling tilt disabled ($E=\\{0\\}$) on the Class 3 datasets and on the real cohort; if the accuracy and ARI gains, or the seven-group structure, largely remain, then trend traveling is not the source of the claim, and if they vanish, the reported clusters rest on the unpenalized rotation.","supporting_citations":[{"cited_title":"Chandereng, A","cited_arxiv_id":null,"evidence_quote":"Supplies the lag-penalized weighted-correlation baseline, the closest prior method to 4TaStiC, which the comparison must beat."},{"cited_title":"Cuturi,Fast global alignment kernels(Omnipress, Madison, WI, USA, 2011), ICML’11, p","cited_arxiv_id":null,"evidence_quote":"Supplies the global-alignment-kernel baseline used in the algorithm comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the cross-correlation dissimilarity baseline for correlation-based clustering."},{"cited_title":"Ranked differences Pearson correlation dissimilarity with an application to electricity users time series clustering","cited_arxiv_id":"2505.02173","evidence_quote":"Supports the idea of combining correlation-based and distance-based dissimilarities into one base measure."}],"review_version":1}