{"id":"1b24f402-ffb0-467d-88f1-c7d76f05a51f","arxiv_id":"1908.00903","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"The paper proposes an event box visualization that encodes duration and time of occurrence for events in temporal sequences, demonstrated on hospital patient-flow data.","lead":"This paper introduces a visual analytics method for event sequences that shows both how long each event lasts and when it occurs, using compact event boxes inside a single overview. It applies the method to a year of rheumatology clinic data and reports four types of findings about patient flow, such as longer consultations when certain steps are skipped.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central 'otherwise could not be obtained' claim is unverified: no readability test, no baseline comparison, and case-study findings lack statistical validation.","rationale":"The reader's weakest assumption is exactly the load-bearing point: the central claim depends on both the truth of the reported findings and the usability of the new encoding, and neither is supported by evidence. I agree with the CONDITIONAL verdict rather than moving to REJECT, because the method is clearly described and the encoding is novel and plausible; the gap is in evaluation, not in internal coherence. The proposed concrete test would directly probe the 'otherwise could not be obtained' phrase by comparing against traditional visualizations and by statistically verifying at least one headline finding. This is the single most decisive check because it targets the value proposition itself rather than secondary implementation details.","tokens_in":7154,"tokens_out":3163,"duration_ms":38064,"concrete_test":"Run a controlled comparison using the same Rheumatology logs: give independent analysts the raw event sequences plus traditional views (chronological strip charts, per-event duration boxplots, time-of-day histograms) and ask them to list findings. Compare their findings with the paper's eight case-study findings and measure recall and time-to-discovery. As a follow-up, if needed, compute a simple statistical check for Finding 3 (e.g., Spearman correlation between time of occurrence and Late Arrival duration). If comparable findings emerge with traditional views, the 'otherwise could not be obtained' claim is refuted; if they do not emerge but also cannot be confirmed statistically, the findings themselves may be artifacts of the encoding.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline contribution (Abstract, Sec. 6) is that the Sequential and Time Patterns overview allows derivation of findings 'otherwise could not be obtained using traditional visualizations.' This requires two conditions: (i) the reported findings are genuine properties of the data, not artifacts of the display or of the unspecified frequency threshold used to select unique sequences (Sec. 4.1), and (ii) users can correctly decode the event-box encoding to discover those findings. Neither condition is tested. The case study (Sec. 5.2) lists eight findings from inspection only, with no statistical confirmation; for example, Finding 3 ('late arrival duration decreases as the day goes by') is asserted without a correlation or regression check. No comparison is made against EventFlow, Outflow, LifeLines2, or even simple boxplots and time-of-day histograms, so 'could not be obtained' is asserted rather than demonstrated. The only stated limitation (Sec. 6) concerns scalability, not interpretability or validity. Because this unsupported claim is the central value proposition, the contribution is plausible but not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a visual analytics methodology for temporal event sequences, centered on a 'Sequential and Time Patterns overview' that embeds a boxplot-inspired 'event box' encoding of event duration (horizontal) and time of occurrence (vertical). The pipeline selects a set of frequent unique sequences, sorts them by similarity using complete-link hierarchical clustering with Levenshtein edit distance, aligns them by user-selected events, and renders each event as an interactive event box with multiple levels of detail. The authors demonstrate the approach on one year of Rheumatology outpatient clinic event logs and propose a four-class taxonomy of findings (main sequential patterns, impact of an event on subsequent event durations, duration trends with respect to time of occurrence, and temporal distribution of time attributes). The central claim is that the proposed visualization allows derivation of findings 'that otherwise could not be obtained using traditional visualizations' (Abstract; Sec. 6).","tokens_in":7354,"tokens_out":1639,"duration_ms":17555,"significance":"If substantiated, the event-box encoding and the integrated sequential/time overview would be a useful addition to event-sequence visualization, particularly for healthcare process data where waiting times and durations matter. The paper's strengths are its use of a real-world dataset, the explicit taxonomy of findings, and the interactive design with multiple alignment and levels of detail. However, the central value proposition—that the method yields insights unobtainable with existing tools—rests on two untested assumptions: that the visual encoding is readable by analysts, and that the reported findings are genuine properties of the data rather than artifacts of the display or of the unstated frequency threshold. The paper provides no user study, no comparison with existing tools (EventFlow, Outflow, LifeLines2, Eventpad, or even simple histograms/boxplots), and no statistical verification of the case-study findings. The contribution is therefore plausible but not established at its current level of evidence.","major_comments":[{"comment":"The central claim that findings can be derived 'otherwise could not be obtained using traditional visualizations' is not validated. The paper provides no comparison against existing systems such as EventFlow [16], Outflow [22], LifeLines2 [21], or Eventpad [4], and no user study demonstrating that analysts can correctly decode the event-box layout. Because this claim is the headline contribution, it needs at least a baseline comparison or a controlled usability study to be supported.","section":"Abstract; Sec. 6"},{"comment":"The frequency threshold used to remove 'unique sequences with relatively low frequency' (step 1 in Sec. 4.1) is never specified, and no sensitivity analysis is reported. The case-study findings in Sec. 5.2 are derived from a display that depends on this threshold, so it is unclear whether the observed patterns (e.g., Finding 2 on consultation duration without Height and Weight, or Finding 3 on late-arrival duration decreasing through the day) are robust or artifacts of the chosen cutoff. The threshold value and a sensitivity analysis should be reported.","section":"Sec. 4.1; Sec. 5.2"},{"comment":"The reported findings are asserted from visual inspection without statistical verification. For example, Finding 3 states that 'the duration of the late arrival event decreases as the day goes by,' but no correlation, regression, or group comparison is provided to confirm this trend or its significance. Similarly, Finding 6 ('consultation times are significantly shorter on Tuesdays') uses the word 'significantly' without a statistical test. Each quantitative claim in the taxonomy should be backed by an appropriate statistical check or explicitly labeled as an exploratory observation.","section":"Sec. 5.2"},{"comment":"The event-box encoding assumes that the vertical placement of data points according to time of occurrence and their horizontal placement according to duration can be read jointly to identify trends and outliers. However, no evidence is presented that this encoding is perceptually effective—e.g., whether users can distinguish quartile ranges from outlier points, or can compare time-of-occurrence distributions across event boxes. Given that the entire methodology rests on this encoding, a perceptual or empirical evaluation is needed.","section":"Sec. 3.1; Fig. 2a"}],"minor_comments":[{"comment":"There is a typo in the phrase 'Sequential and Time patters overview' near the end of the first paragraph of Sec. 6; it should read 'Sequential and Time Patterns overview'.","section":"Sec. 6"},{"comment":"Reference [2] lists the edition as '3nd edition'; this should be corrected to '3rd edition.'","section":"References"},{"comment":"The clustering step is described as 'complete-link agglomerative hierarchical clustering [9] and the Levenshtein edit distance [12],' but details are missing: how the edit distance is computed for sequences of differing lengths, whether costs are unweighted, and how the dendrogram is cut to produce the displayed ordering. A short description or pseudocode would improve reproducibility.","section":"Sec. 4.1"},{"comment":"The figure is dense and the numbered findings (1–8) are referenced in the text but are not all easy to locate in the three panels. Adding a legend or enlarging the relevant regions would improve readability.","section":"Fig. 1"},{"comment":"The explanation of Class II findings would benefit from an explicit statement that the causality is only hypothesized: the current wording ('this finding should require further investigation') is appropriate, but the earlier 'could be related to' phrasing could be mistaken for an established result.","section":"Sec. 5.2, Class II"}],"recommendation":"major_revision","confidential_remarks":"The paper is a short VIS paper with a plausible but unvalidated central claim. The lack of any empirical or comparative evaluation is the main obstacle; however, the core encoding and interaction design are internally consistent, and the case study is real. I see no evidence of intent to mislead; rather, the claims simply outpace the evidence. A revision that adds at least one of (a) a baseline comparison with existing tools, (b) a perceptual study of the event-box encoding, or (c) statistical validation of the case-study findings would be sufficient to bring the paper to publishable strength."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe one thing to know: the core idea is a genuinely new event-box encoding that puts duration on the horizontal axis and time of occurrence on the vertical axis inside quartile bands, integrated into a single overview with multiple-event alignment. That encoding is worth a look. The paper is not, however, able to support its own headline claim that the visualization yields findings 'otherwise could not be obtained' with traditional tools. That claim needs either a baseline comparison, a user study, or at least a statistical check on the reported patterns; none is present.\n\nWhat the paper does well: the related work is carefully positioned. The authors correctly note that most prior systems encode average duration or relegate time of occurrence to secondary views. The proposed encoding is clearly described with levels of detail, and the case study with real hospital data is a good demonstration of use. The taxonomy of findings (sequential patterns, event impact, duration trends, temporal distribution) is sensible and could generalize beyond healthcare.\n\nThe soft spots are real and concentrated in validation. No comparison against EventFlow, Outflow, LifeLines2, or simple boxplots and time-of-day histograms; no user study on readability of the event box; no quantitative verification of findings like 'late arrival duration decreases as the day goes by' (Finding 3). The frequency threshold for selecting unique sequences is unspecified, and there is no code or data artifact. One more weakness: the discussion of limitations only mentions scalability, not interpretability, which is telling. These are significant gaps, but they are gaps in evaluation, not a fatal flaw in the concept. The encoding itself stands as a design contribution.\n\nWho is this for? Visualization researchers working on temporal event sequences. It would be a reasonable short-paper contribution, and a serious referee could help sharpen the claims. If the authors added a small controlled study or even a systematic analysis of the case-study findings, the contribution would be much stronger.\n\nMy recommendation: engage with it, send it to peer review, but expect the 'otherwise could not be obtained' claim to be revised or supported.","headline":"A well-designed event-box encoding for duration and time of occurrence in event sequences, but the paper's central claim that it reveals what 'otherwise could not be obtained' is not supported by any validation.","tokens_in":7834,"tokens_out":1654,"would_cite":true,"duration_ms":16205,"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":"The paper claims that a single overview, built from an event-box encoding of duration and time of occurrence, lets analysts identify trends and outliers in event sequences that standard visualizations cannot show.","keywords":["temporal event sequences","time attributes","visual analytics","event box encoding","duration outliers","patient flow","sequence alignment","boxplot visualization"],"falsifier":"A controlled reading test would settle the claim: take the same outpatient event dataset, ask experienced analysts to find duration outliers and time-of-occurrence trends using the proposed overview versus a conventional sequence overview, and check whether the proposed method yields more correct findings or faster answers. A statistical check would also do: compute the raw distributions behind the case-study findings, for example late-arrival durations by hour of day, and test whether the claimed decreasing trend and morning concentration are significant in the source data.","tokens_in":6956,"feed_emoji":"⏱️","tokens_out":6107,"duration_ms":52473,"temperature":0.7,"pith_summary":"Most visualizations of event sequences show the order of events but hide two time attributes: when each event happens and how long it lasts. This paper claims that both attributes can be encoded in a single overview, called the Sequential and Time Patterns overview, using an event-box encoding that places each occurrence by duration horizontally and time of occurrence vertically, with quartiles and outliers drawn in boxplot style. Applied to one year of outpatient clinic records, the overview produced four classes of findings, from main pathways to duration trends tied to times of day and week. If the claim holds, analysts working with event logs in healthcare, clickstreams, or daily-living data could spot duration outliers and time-of-day patterns that standard sequence views leave invisible.","feed_headline":"Event boxes expose duration trends hidden in sequence views","feed_subtitle":"Outpatient clinic data gains a single overview that surfaces duration outliers and time-of-day trends.","key_machinery":"The load-bearing mechanism is the event box, a boxplot-inspired glyph that aggregates one event type's occurrences: height is proportional to record count, width to the duration range, and interior colored sub-boxes mark the quartiles of the duration distribution. Each data point sits inside the box at coordinates (duration, time of occurrence), so the figure simultaneously shows the distribution of how long events last and when they happen. A standard boxplot outlier rule using 1.5 times the interquartile range marks unusual durations as points outside the colored bands. What this element does is carry the paper's main promise: it replaces the traditional average-duration bar with a full distribution, so individual outliers and time-of-day trends become visible in situ rather than requiring a separate view.","core_discovery":"The central discovery is an event-level visual encoding that aggregates all occurrences of an event type into a single 'event box'. The box's width is proportional to the full range of observed durations, its interior is split into quartile bands, and each occurrence is a point placed horizontally by its duration and vertically by its time of occurrence. The encoding can be collapsed to a point event or expanded through several levels of detail, and time-of-occurrence can be re-scaled from hours to days, weeks or months, with color adding another time scale. Embedded in the overview and combined with user-driven alignment by multiple events and similarity-based sorting of unique sequences, this encoding lets a user compare duration and timing trends both within and across sequences. The paper reports that, on a real-world hospital outpatient dataset, the approach led to four classes of findings that, in the authors' judgment, could not be obtained with traditional visualizations.","pith_inferences":["If the event-box encoding is perceptually sound, the same visual grammar could be applied to any interval-valued event attribute beyond time, such as cost or resource usage, turning the overview into a general multivariate sequence display.","The authors' most ambitious claim, that findings could not be obtained with traditional visualizations, is an empirical statement; a direct comparison with existing sequence tools would convert it from a claim into a measured result.","Because the method sorts unique sequences by edit distance and then aligns by multiple events, it effectively builds an approximate multiple sequence alignment; borrowing richer alignment cost models might yield better grouping of clinically similar pathways.","The claimed taxonomy classes could be turned into automated pattern-detection routines, such as statistical tests for duration shifts between sequence groups, making the visual exploration semi-automatic."],"forward_implications":["A user can inspect the distribution of event durations and times of occurrence without leaving the sequence overview, reducing the need for separate secondary views.","Duration outliers become visible as points outside the quartile bands, so analysts can identify unusual individual cases rather than relying on averages.","Aligning sequences by multiple user-selected events and sorting by similarity exposes how the presence or absence of an early event relates to the duration of later events.","The same overview can be filtered by specific days of the week and broken down by day, revealing temporal patterns such as shorter consultations on certain days or a morning concentration of visits.","The four-class taxonomy of findings (main sequential patterns, impact of an event on later durations, duration trends over time, and temporal distribution of time attributes) is offered as a generalizable scheme for analyzing event data in other domains."],"supporting_citations":[{"why":"Defines the traditional overview approach based on frequency and average duration that the paper argues leaves time attributes under-represented.","marker":"[16]"},{"why":"Introduces the boxplot display that the event-box encoding is inspired by.","marker":"[19]"},{"why":"Supplies the 1.5-interquartile-range outlier rule used to flag duration outliers.","marker":"[6]"},{"why":"Demonstrates multiple sequence alignment for event logs, which the proposed user-driven multiple alignment builds on.","marker":"[4]"},{"why":"Provides the agglomerative clustering procedure used to sort unique sequences by similarity.","marker":"[9]"},{"why":"Provides the edit-distance measure used to compare and sort unique sequences.","marker":"[12]"}],"fun_headline_variants":["Event boxes merge duration and timing into one overview","Duration and timing trends exposed by event box encoding","Sequence alignment reveals duration outliers in event data","New visual encoding compares durations across event sequences"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim that these findings could not be obtained with traditional visualizations depends on an untested assumption: that a user looking at the event-box layout will read the plotted points and quartile bands correctly and that the patterns highlighted in the case study are genuine properties of the data rather than artifacts of the chosen display.","fun_headline_variants_meta":{"raw":{"variants":["Event boxes merge duration and timing into one overview","Duration and timing trends exposed by event box encoding","Sequence alignment reveals duration outliers in event data","New visual encoding compares durations across event sequences"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000226,"raw_usage":{"total_tokens":1429,"prompt_tokens":867,"completion_tokens":562,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":483,"completion_tokens_details":{"reasoning_tokens":505}},"tokens_in":483,"tokens_out":562,"duration_ms":5937,"temperature":1.0,"reasoning_tokens":505,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:27:48.967666+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled reading test would settle the claim: take the same outpatient event dataset, ask experienced analysts to find duration outliers and time-of-occurrence trends using the proposed overview versus a conventional sequence overview, and check whether the proposed method yields more correct findings or faster answers. A statistical check would also do: compute the raw distributions behind the case-study findings, for example late-arrival durations by hour of day, and test whether the claimed decreasing trend and morning concentration are significant in the source data.","supporting_citations":[{"cited_title":"Monroe, R","cited_arxiv_id":null,"evidence_quote":"Defines the traditional overview approach based on frequency and average duration that the paper argues leaves time attributes under-represented."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the boxplot display that the event-box encoding is inspired by."},{"cited_title":"Frigge, D","cited_arxiv_id":null,"evidence_quote":"Supplies the 1.5-interquartile-range outlier rule used to flag duration outliers."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Demonstrates multiple sequence alignment for event logs, which the proposed user-driven multiple alignment builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the agglomerative clustering procedure used to sort unique sequences by similarity."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the edit-distance measure used to compare and sort unique sequences."}],"review_version":1}