Sequen-C: Sequence Cluster Explorer

• Sequen-C is a visual analytics system that integrates hierarchical clustering, the Align-Score-Simplify representation, and interactive transformations for exploring temporal event sequences.
• It employs a multilevel overview with adjustable granularity, using both vertical clustering and horizontal simplification to balance detail and clarity.
• Sequen-C supports coordinated views and automated statistical analyses, making it effective for uncovering common patterns and anomalies in complex healthcare data.

Searching arXiv for Sequen-C and related papers. Searching arXiv for "Sequen-C". Sequen-C, short for Sequence Cluster Explorer, is a visual analytics system for exploring temporal event sequences through a multilevel overview whose granularity can be transformed both across sequence clusters and within each cluster’s summary (Magallanes et al., 2021). It was introduced to address the limitation of event-sequence overview tools that expose only a single clustering or a single summarization level, and it was later extended with EventBox, a visual encoding for analyzing groups of events and their temporal and multivariate attributes jointly (Montana et al., 19 Jul 2025). Across these developments, Sequen-C combines hierarchical clustering, the Align-Score-Simplify representation, multilevel coordinated views, user-driven transformations, and automatically generated statistical analyses to support the discovery of common temporal patterns, rare or anomalous sequences or events, relationships among event attributes, and statistically supported explanations for those relationships.

1. Problem domain and analytical scope

Sequen-C is designed for temporal event sequences: ordered event data over time in which each sequence may represent a patient pathway, an incident trajectory, or another temporally structured record. The original system emphasizes multilevel overview of such sequences, with two distinct notions of level-of-detail. The vertical level-of-detail is controlled by the number of clusters $k$, so that small $k$ yields a coarse overview and larger $k$ yields a finer one. The horizontal level-of-detail is controlled by an information threshold $I_\tau$, which determines how much detail is shown within each cluster representation (Magallanes et al., 2021).

The system was motivated by the observation that a single “best” overview can be analytically risky. A clustering that is too coarse may merge distinct pathways, whereas one that is too fine may overwhelm the analyst. Likewise, a summary that preserves all variation may be noisy, while one that collapses too much may erase meaningful permutations. Sequen-C was therefore developed to support summary of common and deviating pathways, exploration of alternative clusterings, meaningful comparison of sequences within a cluster, and detail-on-demand from clusters down to raw records. The 2025 extension broadens this scope from event order alone to explicit analysis of temporal and multivariate event attributes, including sequence-level attributes and event-level attributes (Montana et al., 19 Jul 2025).

The event attributes handled in the EventBox extension are explicitly typed as temporal attributes $\tau$, categorical attributes $\kappa$, and numerical attributes $\eta$. Examples given in the paper include duration and start time for temporal attributes; day of the week, gender, urgency, symptom, and treatment for categorical attributes; and age for numerical attributes. This suggests that Sequen-C is intended not merely as a sequence summarization environment, but as a system for joint structural and attribute-based analysis of event-sequence data.

2. Multilevel overview and the Align-Score-Simplify representation

Sequen-C constructs an aggregate tree over the set of unique temporal event sequences by agglomerative clustering. Each unique sequence begins as its own leaf node, the two most similar clusters are repeatedly merged, and the process continues until one root cluster contains the entire dataset. Each node is defined as

$\text{Node}(l,r,\lambda,\alpha)$

where $l$ and $r$ are the left and right child nodes, $k$0 is the alignment of the sequences at that node, and $k$1 is the simplified representation used to visualize the cluster (Magallanes et al., 2021).

The initial pairwise distance between sequences is based on a q-gram cosine distance $k$2, with $k$3. This makes the comparison depend on shared event types rather than strict order. The paper explicitly notes that if stricter sequence order were needed, a different metric such as Levenshtein edit distance could replace it. Cluster-to-cluster distance is computed with the average agglomeration method $k$4, defined as the average of all pairwise distances between sequences across the two clusters. The hierarchical tree thereby represents all possible clusterings from $k$5 to $k$6, where $k$7 is the number of unique sequences.

The cluster representation is produced by Align-Score-Simplify. In the alignment stage, sequences in a cluster are aligned using Multiple Sequence Alignment (MSA), with each event type encoded as a character and gaps inserted so that equal events align column-wise as much as possible. The implementation uses a progressive alignment strategy with gap open penalty $k$8, match score $k$9, and mismatch score $k$0. This is intended to expose not only similarity of content but also common backbones and permutations.

In the scoring stage, each alignment column is assigned an information score

$k$1

where $k$2 is the set of unique event types in the alignment and $k$3 is the entropy of column $k$4. If a column is mostly one event type, $k$5 is close to 1; if it contains many different events or many gaps, $k$6 is closer to 0. The probabilities are weighted by the frequencies of unique sequences in the cluster, so common sequences contribute more than rare ones. In the simplification stage, consecutive low-information columns are collapsed whenever

$k$7

producing the simplified matrix $k$8. This mechanism defines the horizontal level-of-detail.

The default number of clusters is suggested by the average silhouette width. For a given $k$9, the average silhouette width $I_\tau$0 is the mean of

$I_\tau$1

over all unique sequences $I_\tau$2, where $I_\tau$3 is the average distance from $I_\tau$4 to other elements in the same cluster and $I_\tau$5 is the average distance from $I_\tau$6 to the nearest neighboring cluster. The largest global maximum of $I_\tau$7 gives the most optimal cluster count, and local maxima are also exposed so that analysts can explore multiple alternative optimal clusterings. This formalizes the vertical level-of-detail while preserving user-guided exploration.

3. Coordinated views and detail-on-demand exploration

The original Sequen-C system implements the method through three coordinated views plus an attribute analysis view (Magallanes et al., 2021). The multilevel overview, or cluster view, is the main summary view. Each cluster is rendered according to the simplified alignment matrix $I_\tau$8; colored boxes represent event types, box height is proportional to the number of records, gaps appear as spacing, and cluster height reflects frequency in the dataset. Small clusters can be scaled up and outlined with a dotted line so that rare or deviating pathways remain visible. A cluster slider changes $I_\tau$9, and an information score slider changes $\tau$0. A combobox lists the default clustering and alternative optimal clusterings suggested by silhouette peaks.

The unique sequence view displays the individual unique sequences contained in a selected cluster, together with their frequencies. These sequences are shown in full, without simplification. This view supports inspection of the exact sequence content underlying a summary, sorting by frequency or similarity, and alignment by a selected event. It functions as the bridge between cluster-level abstraction and record-level inspection.

The individual sequence view provides the deepest level of detail. It shows selected individual records as a Gantt-like timeline, with events positioned according to their timestamps. A table beside the timeline displays raw attributes at either the sequence level or event level. This makes Sequen-C a detail-on-demand system in the strong sense: the analyst can move from a cluster summary, to a unique sequence, to raw temporal records and metadata.

The attribute analysis view allows comparison of attributes for selected clusters, unique sequences, or individual records. For each attribute, it shows stacked bar charts in three modes: Selected data, Sequence, and Cluster. This capability is especially important in healthcare case studies, where pathway structure alone is insufficient and analysts also need to relate pathways to age, gender, area classification, diagnosis, discharge, length of stay, or similar context.

Selections propagate across views, and filtering supports queries based on data attributes, frequency, date range, event occurrence, and temporal features such as day of week or month. The result is an exploration model organized around overview, alternative abstraction levels, and progressively finer inspection.

4. EventBox, transformations, and statistical reporting

The 2025 extension integrates EventBox into Sequen-C as the core mechanism for analyzing temporal and multivariate attributes jointly (Montana et al., 19 Jul 2025). An EventBox is defined as

$\tau$1

where $\tau$2 is the number of time-stamped event occurrences, $\tau$3 is the $\tau$4-th event occurrence, and $\tau$5 are the five attributes used in the EventBox. Each event has a multivariate attribute vector

$\tau$6

The five selected attributes are assigned the roles primary horizontal $\tau$7, primary vertical $\tau$8, secondary horizontal $\tau$9, secondary vertical $\kappa$0, and breakdown $\kappa$1. By default, for temporal analysis, $\kappa$2 duration and $\kappa$3 start time, but other types can be selected.

EventBox is inspired by box plots, scatter plots, and histograms. Its container area is a rectangular enclosure for all events of the same type, with color hue encoding event type, height encoding the number of events $\kappa$4, and width encoding the maximum value of the primary horizontal attribute $\kappa$5. Quartile lines summarize the primary horizontal attribute by minimum, $\kappa$6, median $\kappa$7, $\kappa$8, and maximum. Outliers are defined using Tukey’s rule,

$\kappa$9

with $\eta$0. Each event occurrence is shown as a point whose horizontal position is $\eta$1 and whose vertical position is $\eta$2, with additional encoding through color hue for the breakdown attribute $\eta$3 and transparency to create density-like heatmap effects. Histograms display distributions of the primary and secondary attributes, and when secondary attributes are selected the histogram bars can become stacked bar charts. The breakdown attribute subdivides the data into one EventBox per unique value of $\eta$4.

To support interactive creation and exploration of EventBoxes, Sequen-C adds four user-driven transformations. Alignment allows sequences to be aligned by selected event types; hard events are prioritized for alignment, and soft events are then aligned between hard-aligned events. The paper notes that alignment inserts visual gaps that are not real time gaps, but layout artifacts to make sequences comparable. Sorting orders sequences by an event of interest by using the chosen event index to define a suffix or subsequence and then comparing sequences using those subsequences. Substitution allows one or more event types to be replaced with a new event type, which is useful for simplifying complex sequences, merging semantically related events, and reducing visual clutter. Aggregation merges consecutive occurrences of events into a single aggregated event type so that higher-level process phases can be analyzed instead of many fine-grained events.

EventBox is paired with an automatically generated statistical report. The report includes averages and standard deviations for continuous attributes, with mean comparison tests at different granularities; contingency tables for associations among selected categorical attributes; and ANOVA tables relating one selected continuous attribute to one or more categorical variables. The system relies on the central limit theorem for these group comparisons because group sizes are large, so further distribution checks are not needed. The stated purpose of these analyses is to validate whether visual differences are meaningful, identify significant attribute interactions, and avoid over-interpreting visual patterns that might be noise.

5. Evaluation and healthcare case studies

The original Sequen-C paper reports two healthcare case studies, CUREd and MIMIC-III, to demonstrate how multilevel overview and detail-on-demand can aid summary of common and deviating pathways and support attribute inspection (Magallanes et al., 2021). In the CUREd dataset, the subset used contained 25,243 calls, 21,805 unique patients, and 57 attributes. With the default clustering, the overview showed 200 clusters, but the first four covered about 85% of the data. When reduced to four clusters, the major pathways were: ambulance service and attendance to the emergency department; ambulance service and conveyance to a hospital; ambulance arrives, but no conveyance; and call closed without ambulance service. Attribute inspection linked pathway type to area classification, age, and symptom. A further clustering with 11 groups among sequences containing an emergency department event exposed distinctions involving triage and follow-up by a health professional, and uncovered a cluster in which patients had multiple calls before the first ambulance response.

The MIMIC-III case study used a subset focused on patients with a first or second diagnosis of atrial fibrillation, containing 1,425 admissions and 448 event types, most of which were prescriptions. The overview exposed common care-unit patterns involving CSRU, CCU, MICU, and SICU. The cluster structure separated some pathways by medication and others by care unit with different internal treatment variants. Among the examples highlighted were CSRU clusters with distinct treatment differences such as Metoprolol versus Warfarin, and higher mortality in MICU and SICU groups when atrial fibrillation was a second diagnosis rather than a primary diagnosis.

The EventBox extension adds a formal evaluation with 21 participants: 3 domain experts and 18 novice data analysts (Montana et al., 19 Jul 2025). The novice study was in person and included training, task performance, and visualization value evaluation; screens and mouse clicks were recorded. Participants were aged 22–40, with 9 female and 9 male, had an engineering or data-analysis background, and had no prior exposure to Sequen-C or EventBox. They completed a 15-question multiple-choice questionnaire designed to cover all design requirements, with accuracy defined as correct answers ratio and completion time as response time per question. The study took less than 60 minutes. Questions Q1–Q7 did not require EventBox, Q8–Q9 were possible without EventBox but with much more effort, and Q10–Q15 required EventBox.

The novice results were an average accuracy of $\eta$5 and an average response time of 95 seconds per question. The hardest question was Q3, answered correctly by only 8 of 18 participants; the fastest was Q7 at 27 seconds; the slowest was Q9 at 221 seconds, though Q9 was still answered correctly by all participants. The visualization value assessment used the 21-question ICE-T questionnaire, covering Time minimization, Insight generation, Essence conveyance, and Confidence generation. The overall visualization value was $\eta$6, the best-rated component was Insight $\eta$7, the lowest component was Time $\eta$8, the highest individual heuristic in Essence was “The visualization helps understand how variables relate...” $\eta$9, and the lowest Confidence heuristic was $\text{Node}(l,r,\lambda,\alpha)$0, with concern about insufficient explicit communication of missing, duplicate, or invalid data.

Three separate interactive sessions with healthcare domain experts were also reported, each lasting 90 minutes and including a walkthrough of new features, expert-posed questions answered using the system, and feedback on ICE-T categories, usability, and future work. Experts stated that the system quickly identifies patterns in complex processes, helps detect outliers and deviations, is valuable for identifying bottlenecks, supports rapid exploration without SQL, and provides multiple perspectives beyond a single sequence. They also noted that data quality issues are common in healthcare, that statistical analysis increases confidence, and that the system has a learning curve for advanced features.

The extension presents three real-world healthcare case studies. In ANCU, the dataset contained 73,279 events for 9,623 pregnant women over three months of clinic visits. The analysis selected 8 clusters, focused on cluster C1, aligned sequences using NYA, IS, WC, IC, and COM, treated WC and IC as soft events, sorted by WC, and built an EventBox for WC. Reported findings included a higher proportion of outliers on Fridays, longer waiting times on Mondays and Fridays mainly in the morning, shorter waits on Tuesdays to Thursdays across morning and afternoon sessions, and clinic-specific time slots revealed through ClinicCode. In the CUREd ambulance-service case, the dataset contained 25,243 calls, 21,805 patients, 34 attributes, and three months of emergency service data. The analysis selected 15 clusters, compared clusters C1 and C6, used CAL event alignment, and analyzed duration, start time, urgency, symptom, day of the week, and cluster. Findings included more children in C6 than C1, a long-duration outlier of over 5 hours in C1, more outliers in C1 than C6 at 10.58% versus 8.48%, more outliers in the afternoon in both clusters, and 75% of calls answered within 40 minutes for C1 and 30 minutes for C6. A heatmap EventBox showed most Red calls dispatched within 10 minutes, more variable durations for Amber calls, Chest Pain common for Red calls in C1, and Convulsions Fitting dominant in C6. The statistical report found a significant three-way interaction among cluster, day of the week, and urgency, with Psychiatric Suicide Attempt, GreenT, and GreenF among the strongest duration-increasing factors. In MIMIC-IV, the data were filtered for patients with Hypertension, Chronic Kidney Disease, and Diabetes, resulting in 1,776 patients. An original cluster with 102 event types and sequences up to 153 events was reduced by substitution and aggregation to 7 event types, with drugs merged into DRU and transfers into TRA, after which alignment and sorting clarified sequence structure, proportions of transfers, and how medication changes related to discharge.

6. Significance, limitations, and disambiguation

Compared with prior event-sequence visualizations, Sequen-C with EventBox does not just show event order or average attribute values. It explicitly visualizes temporal distributions, supports multiple multivariate attributes at once, allows interactive restructuring of sequences before analysis, and combines visualization with statistical evidence (Montana et al., 19 Jul 2025). This suggests a shift from overview-only sequence visualization toward a tighter integration of clustering, transformation, attribute analysis, and inferential support.

The original paper is explicit about several limitations (Magallanes et al., 2021). Scalability is constrained by event-type coloring when the number of event types becomes large. Long lists of event types are cumbersome, and a hierarchy of event types would be beneficial. Horizontal simplification can be visually confusing because users may have difficulty seeing which subsequences were collapsed when $\text{Node}(l,r,\lambda,\alpha)$1 changes. The choice of $\text{Node}(l,r,\lambda,\alpha)$2 is not yet optimal. Merged subsequences can still become cluttered when there are many event types. Alignment is computationally expensive and sensitive to gap and substitution costs. Meaning-based alignment might be preferable to name-based alignment, and alternative alignment strategies such as longest common subsequence could be explored. Generalization beyond healthcare remains to be validated.

The name also requires disambiguation across adjacent literature. A separate paper on SymSeqBench contrasts its own role as a unified, modular framework for generating, grounding, analyzing, and benchmarking symbolic sequences with a “Sequen-C-like system,” described there as a model or architecture for learning sequences; in that framing, SymSeqBench is the testbed rather than the model (Zajzon et al., 31 Dec 2025). This usage is distinct from Sequen-C as Sequence Cluster Explorer, which is a visual analytics system for temporal event sequences. The term should also not be conflated with Sequencing by Emergence (SEQE), a proposed single-molecule DNA/RNA sequencing technology in which the paper notes that “Sequen-C” is used in the context of the SEQE inference framework; that work concerns probabilistic sequence estimation from probe-binding localizations rather than visual analytics of event sequences (Boyd et al., 2021).

Within the event-sequence visualization literature, Sequen-C is therefore best understood as a multilevel visual analytics environment whose original contribution was the combination of hierarchical aggregation and Align-Score-Simplify for adjustable overview construction, and whose later contribution was the integration of EventBox, user-driven transformations, and automatically generated statistical analyses for joint temporal and multivariate attribute analysis.