ParcoursVis: Visual Analytics for Pathway Data

• ParcoursVis is a framework for visualizing and analyzing sequential event data, effectively encoding complex patient pathways, urban routes, and sports trajectories.
• It employs progressive visual analytics and LLM-mediated query synthesis to enable scalable, interactive exploration of large datasets with stable visual representations.
• Its applications span healthcare, urban design, architecture, and sports, offering actionable insights through intuitive visual summaries and expert-driven evaluation protocols.

ParcoursVis is a class of systems and models focused on the visualization and analysis of human or agent movement and event sequences—the "parcours" or path—across a range of domains including healthcare, architecture, urban environments, and sports. The term encompasses both large-scale interactive visual analytics platforms for Electronic Health Records (EHR) (notably the Progressive Visual Analytics system ParcoursVis for patient pathway analysis), as well as computational frameworks for understanding the embodied experience of route traversal in built environments and urban contexts. The following account integrates methodologies and findings from medical informatics, behavioral architecture, urban studies, and sports informatics as reported in the major ParcoursVis and related literature.

1. Foundational Principles: Pathway Aggregation and Event Sequence Visualization

ParcoursVis systems are fundamentally defined by their capacity to aggregate, simplify, and encode large collections of sequential event data—typically patient, user, or agent event trajectories—into interpretable visual summaries. In the EHR context, ParcoursVis constructs prefix trees over millions to tens of millions of patient event sequences, employing Icicle tree visualizations in which node height encodes frequency, width encodes mean or median duration, and color encodes event type. The architecture leverages event sequence simplification (ESS) to facilitate exploration of branching pathways, dominant flows, and rare trajectories—patterns of critical interest in healthcare and other domains where both the common and the anomalous are significant (Assor et al., 14 Aug 2025, Assor et al., 12 Sep 2025).

2. Progressive Visual Analytics and Scalability

ParcoursVis addresses the scalability barrier inherent to classic event-sequence visualization, where interactive exploration is limited by the need to process entire datasets before visualization is possible. Its progressive visual analytics (PVA) architecture partitions processing into chunks, incrementally updating aggregation trees and rendering visualization states as computation proceeds. This enables immediate, approximate feedback at scale—supporting datasets with tens of millions of entities and billions of events, with performance benchmarks reporting throughput on the order of 13 million patient sequences per second on multi-core systems (Assor et al., 14 Aug 2025).

A critical component is the maintenance of visual stability as node frequencies and orderings update with additional data. ParcoursVis implements hysteresis-based stable sorting: node order only changes when frequency differences exceed a specified inertia threshold, preserving the user's mental map across progressive updates.

3. Medical Use Cases: Patient Pathway Exploration at Population Scale

ParcoursVis has been developed and deployed in real-world healthcare analytics, notably through partnerships with 16 Parisian hospitals and the French Social Security. Use cases include exploration of emergency department flows and non-cancerous prostate adenoma treatment trajectories. The system operates over EHR datasets whose scale (millions of patients, billions of events) and coding heterogeneity far exceed previous systems such as EventFlow. This scale is essential for detecting rare, clinically important pathologies and patient behaviors—scenarios where data sampling would systematically exclude less common but highly actionable cases (Assor et al., 12 Sep 2025, Assor et al., 14 Aug 2025).

Core analytical capabilities include:

• Exploration of complex cohort event trees,
• Filtering and abstraction over multi-granular event types,
• Comparative analyses across institutions and timeframes,
• Identification of unexpected deviations from clinical guidelines,
• Detection of operational bottlenecks and data quality issues.

4. Interaction Paradigms: From Dynamic Queries to LLM-mediated Query Synthesis

A significant recent innovation in ParcoursVis is the transition from conventional dynamic query widgets (sliders, drop-downs) to LLM-mediated natural language-to-SQL translation (Assor et al., 11 Sep 2025). In domains characterized by high-dimensional, coded, and heterogeneous data (e.g., the French SNDS), LLM-based interfaces allow users to express rich analytical intentions in natural language, which are automatically converted into executable, editable queries. This paradigm reduces visual clutter and negates the cognitive burden of memorizing dataset schema or legal code values. Editable query transparency and natural-language paired explanations are required to address LLM-specific limitations such as nondeterminism and possible hallucination.

Nonetheless, ParcoursVis does not abandon dynamic queries; instead, widget-based dynamic query refinement is selectively reinstated for tractable query fragments, thereby recovering properties such as direct manipulation, immediate feedback, and reversible interaction where feasible.

5. Evaluation Methodologies: Insight-based Expert Assessment

To address shortcomings of traditional usability metrics in exploratory analytics, ParcoursVis research leverages insight-based evaluation protocols. These protocols focus on characterizing the quality and mechanism of discoveries made by domain experts during system use: assessing whether insights are hypothesis-driven or unexpected, their depth/breadth, actionability, domain value (1–5, from trivial to critical), and the mental strategies (top-down vs. bottom-up, sensemaking loops) that lead to discovery (Assor et al., 12 Sep 2025).

A two-stage protocol consisting of in-person exploration sessions with think-aloud reporting, followed by remote expert and participant review, enables triangulation and validation of found insights. Pilot studies confirm the importance of contextual data richness, institution-specific event semantics, and a staged separation of training from open-ended exploratory tasks.

6. ParcoursVis in Architecture and Route Experience: Visuo-Locomotive Complexity Models

In architectural and built-environment applications, ParcoursVis-style reasoning is formalized through visuo-locomotive complexity models, which quantify the cognitive experience of navigation as a function of visual, structural, and dynamic scene attributes encountered along a route (Kondyli et al., 2020). The path itself is represented as a polyline, along which quantitative (clutter, object density, luminance), structural (symmetry, openness, interconnection density), and dynamic (movement, flicker) attributes are evaluated. Complexity is aggregated and classified (1–5), with moderate levels targeted for optimal wayfinding. The model both diagnoses problematic path segments and supports parametric manipulation of environmental morphology during design—enabling evidence-based tuning of navigability and legibility in hospitals, airports, and large public spaces.

7. Broader Applicability and Limitations

ParcoursVis methodologies generalize to additional domains—sports (e.g., streaming trajectory overlays in skier broadcast footage (Dunnhofer et al., 2023)), urban route selection by visual appearance (Wu et al., 2024), and geospatial recreational analytics (Rauscher et al., 2023)—with a common emphasis on representing, simplifying, and interactively visualizing complex, sequential pathways.

Limitations include dependency on attribute coding quality, interpretability and stability of LLM-synthesized queries, the need for tailored preprocessing pipelines in heterogeneous data environments, and challenges in uniformly abstracting actionability or meaning across diverse application domains. In architectural applications, class-based complexity aggregation is context-dependent and often requires further empirical refinement.

---

References:

(Assor et al., 14 Aug 2025, Assor et al., 12 Sep 2025, Assor et al., 11 Sep 2025, Kondyli et al., 2020, Dunnhofer et al., 2023, Wu et al., 2024, Rauscher et al., 2023)