PaperlayerView exact revision on arXiv

Source-linked AI summary

Visualizing and Interpreting Temporal Care Team Dynamics for Cancer Survival Risk Analysis

Yuhua Huang, Hsiao-Ying Lu, Kwan-Liu Ma

arXiv:2610.30748v1cs.LG

TL;DR

Prior studies link care-team structure to outcomes, but the temporal evolution of collaboration and its relationship to changing patient risk remain poorly understood. This work combines temporal collaboration networks, risk estimation, feature explanations, and interactive visual analysis; case studies and expert feedback illustrate its use for comparing care-team patterns and developing hypotheses, not definitive clinical conclusions.

  • Problem

    Prior work links care-team structure to patient outcomes, but often uses static networks, leaving temporal collaboration dynamics and their relationship to evolving risk less understood.

    Source: §1 · p. 1

  • Method

    The framework uses temporal graph-based risk modeling on evolving EHR-mediated collaboration networks and coordinates surrogate feature attribution with interactive visual analysis.

    Source: §1 · pp. 1–2

  • Results

    Case studies and expert feedback demonstrate how the system supports exploration of collaboration patterns and risk trajectories and generation of hypotheses.

    Source: §1 · p. 1, §7 · p. 9

  • Takeaways & Limitations

    The framework offers a coordinated workflow for human–AI analysis of complex, evolving cancer-care collaboration, intended for insight generation rather than definitive clinical conclusions.

    Source: §7 · p. 9

    EHR access-log networks omit communication outside the record and do not encode clinical intent, decision-making context, or communication quality.

    Source: §6 · p. 9

Abstract

from arXiv · show

Cancer care is a longitudinal, time-sensitive process requiring coordination among diverse healthcare professionals (HCPs), whose collaboration evolves throughout treatment. Prior studies link care team structure to patient outcomes, but the temporal dynamics of collaboration and their relationship to evolving patient risk remain poorly understood. We introduce a visual analytics framework for modeling and interpreting time-varying care team interactions in oncology. We construct dynamic collaboration networks from electronic health record (EHR)-mediated communication data and use a temporal graph neural network (TGNN) to estimate patient outcome risk over time. We then investigate how changes in care team composition and specialist participation relate to shifts in estimated risk across treatment stages. Our multi-view interactive system combines TGNN-based risk modeling with coordinated visualizations of care team networks, feature-level explanations from a surrogate logistic regression model, and exploratory what-if analysis. Without making causal claims, the system helps analysts examine how risk estimates co-evolve with collaboration patterns and identify care team characteristics associated with these changes. We demonstrate its utility through expert-guided analysis scenarios involving 505 patients with breast, colorectal, and lung cancers, evaluating its ability to reveal clinically meaningful collaboration patterns. This work improves the interpretability of temporal graph-based models in longitudinal, team-based healthcare, with potential applicability to other evolving collaborative settings.

1 INTRODUCTION

Care-team collaboration changes throughout cancer treatment, but prior work often represents it as a static network, and EHR usage metadata has received comparatively little attention as a collaboration signal. This work aligns evolving collaboration networks, temporal risk estimates, and feature explanations in a visual analytics system for exploration and hypothesis development.

1 INTRODUCTION

  • Prior care-team research often models collaboration as a static network, while EHR metadata such as access logs, shared documentation, and co-signatures remains comparatively underused.These timestamped digital traces can support construction of time-evolving collaboration networks.Source: §1 · p. 1
  • Predictive modeling, collaboration visualization, and feature-level interpretation are often treated separately or use different temporal bases.The paper uses a shared temporal basis to align risk estimation, explanation, and exploration.Source: §1 · p. 1
  • The system supports analysts in examining how collaboration patterns and model-estimated risk co-evolve, with the aim of generating insights and hypotheses rather than definitive clinical conclusions.Its utility is demonstrated through case studies and expert feedback.Source: §1 · p. 1
  • The system aligns temporal graph-based risk modeling, surrogate feature attribution, and interactive exploration on the same evolving collaboration networks.It supports cohort-level, comparative, and longitudinal analysis, including exploratory what-if analysis.Source: §1 · pp. 1–2

2 RELATED WORK

Prior work established EHR-derived networks and dynamic graph methods as ways to represent collaboration and model outcomes, while leaving gaps in patient-level temporal interpretation. Existing explanations and visual analytics systems do not jointly connect changing collaboration, estimated risk, and feature attributions on a shared temporal basis.

2.1 Healthcare Team Collaboration Analysis

  • EHR-derived care-team networks support analysis of team structure, but prior work has largely emphasized descriptive or aggregate analysis rather than interactive interpretation of patient-level temporal dynamics in predictive modeling.Earlier work also used digital traces to construct patient-centered networks and visualized EHR-mediated collaboration.Source: §2.1 · p. 2

2.2.1 Modeling dynamic graphs

  • Healthcare graph-learning work includes fixed-window collaboration models and cumulative weekly snapshots, but understanding how changing collaboration relates to evolving risk and model behavior remains open.The temporal formulations represent care teams as sequences of cumulative weekly snapshots.Source: §2.2.1 · p. 2

2.2.2 Interpreting dynamic graphs

  • Surrogate explanations express graph predictions through interpretable features, but do not by themselves support longitudinal comparison or temporal drill-down alongside changing collaboration structure.Feature contributions at a time point can be related to underlying data dynamics in a broader temporal and structural context.Source: §2.2.2 · p. 2

2.2.3 Visual analytics systems for graphs

  • Existing visual analytics systems support inspection of graph predictions, clinical trajectories, or EHR collaboration, but do not jointly align temporal collaboration networks, estimated risk, and feature explanations.This leaves practitioners with limited support for reasoning about how evolving team structures relate to predicted-risk changes and associated attributions.Source: §2.2.3 · p. 2

3 METHODOLOGY

The methodology integrates temporal collaboration networks, patient risk estimation, surrogate-based feature attribution, and coordinated visual analysis. Figure 1 summarizes the system components, while the analyst-facing workflow is described separately.

3 METHODOLOGY

  • The methodology integrates temporal collaboration networks, risk estimation, surrogate-based feature attribution, and coordinated visual analysis.Source: §3 · p. 2
  • Figure 1 summarizes the system components.Source: §3 · p. 2
  • The analyst-facing workflow is described in Section 3.3, separately from the component overview.Source: §3 · p. 2

3.1 Data Processing

The study uses a curated cohort of Stage II and III cancer patients with clinical variables and timestamped EHR note-access logs. Patient-specific collaboration is represented as cumulative bipartite note–HCP graphs over weekly snapshots, which capture EHR-mediated interactions rather than all care-team communication.

3.1.1 Data Overview.

  • The curated cohort contains 505 patients with Stage II or Stage III cancer from an academic medical center.The cohort includes breast, nonsmall cell lung, and colorectal cancer patients; colorectal cases combine colon and rectum/rectosigmoid groups.Source: §3.1.1 · p. 2
  • The dataset combines clinical and demographic variables with timestamped EHR access logs covering a fixed 15-month window from three months before diagnosis to one year after diagnosis.Only note-access events are included, and patients without sufficient records for the full window are excluded.Source: §3.1.1 · p. 3

3.1.2 Care Team Network Construction.

  • Each patient’s collaboration is represented as a bipartite graph whose note and HCP nodes are connected by recorded writing or reviewing events.The formulation represents interactions mediated through notes, not direct HCP-to-HCP edges.Source: §3.1.2 · p. 3
  • Weekly cumulative snapshots add recorded nodes and interactions over time, preserving when providers first enter the collaboration record as the team structure grows.The snapshots use the earliest recorded interaction in the analysis window as their temporal reference point.Source: §3.1.2 · p. 3
  • The resulting networks are proxies for EHR-mediated collaboration because they omit communication outside the record and reflect local documentation and access practices.Source: §3.1.2 · p. 3

3.2 Model Pipeline

The pipeline models weekly changes in EHR-mediated collaboration to estimate survival risk, then uses an interpretable surrogate to examine which collaboration attributes are associated with those estimates. Its feature perturbations are exploratory projections, not causal clinical effects.

3.2.1 Temporal Prediction with EvolveGCN.

  • EvolveGCN processes weekly cumulative note-HCP snapshots, evolving graph-convolution weights over time to preserve snapshot order and accommodate changing node and edge sets.Unlike static or temporally aggregated approaches, it retains collaboration history accumulated through each week.Source: §3.2.1 · p. 3
  • EvolveGCN-O is selected because provider and note attributes are fixed categorical descriptors, while node presence and graph structure change across weeks.In this variant, an LSTM updates the convolution weights without using node embeddings in the recurrent update.Source: §3.2.1 · p. 3
  • For each cancer cohort, the model predicts deceased versus surviving outcomes at each patient-week using only snapshots available through that week.Recorded patient outcomes supervise the patient-week examples.Source: §3.2.1 · p. 4

3.2.2 Surrogate-Based Feature Attribution.

  • A logistic-regression surrogate predicts the same outcome from cumulative collaboration attributes, using coefficient-weighted contributions to identify features associated with higher or lower risk.The system retains the 80 features with the largest absolute contributions for visualization.Source: §3.2.2 · p. 4
  • The surrogate is an interpretable proxy, not a direct decomposition of EvolveGCN predictions or evidence of a causal clinical effect.Explanation stability across retraining is reported in the supplement.Source: §3.2.2 · p. 4
  • What-if analysis uses learned surrogate coefficients to project the direction and magnitude of feature perturbations relative to the original EvolveGCN risk estimate.The projected trajectory is intended for exploratory reasoning about directional sensitivity in the learned model representation.Source: §3.2.2 · p. 4

3.3 Visual Analytics Design

The design coordinates patient-level collaboration, model-estimated risk, and surrogate explanations on a shared temporal basis. Its workflow moves from cohort selection to temporal inspection, care-team interpretation, and exploratory what-if analysis.

3.3.1 Design Rationale

  • The design addresses the difficulty of interpreting high-dimensional patient-level risk trajectories, weekly collaboration networks, feature attributions, and HCP activity together.Clinical and informatics collaborators helped define analytical tasks spanning case selection, temporal inspection, care-team interpretation, and patient comparison.Source: §3.3.1 · p. 4
  • Radial glyphs provide compact trajectory-shape comparison rather than exact value reading, while ego networks support inspection of provider neighborhoods and specialty-integrated co-access.The coordinated views connect when risk changes, which attributes are emphasized, and how providers are organized.Source: §3.3.1 · p. 4

3.3.2 Workflow Overview

  • The workflow narrows from cohort filtering to selecting risk-relevant weeks, interpreting care-team attributes and connections, and testing exploratory feature perturbations.The what-if result is framed as a hypothesis for further investigation rather than a clinical claim.Source: §3.3.2 · p. 5

3.3.3 View Design

  • The cohort scatterplot supports case selection, while the radial glyph distinguishes absolute weekly risk from week-to-week risk change and synchronizes linked views to a selected week.The scatterplot can also compare specialty breadth with peak risk escalation.Source: §3.3.3 · p. 5
  • Feature bars show cumulative collaboration contributions and direction, specialty charts show team composition, and ego networks reveal providers connected through shared note access.The views support inspection at a selected week or across the broader observation period.Source: Figure 2 · p. 5, §3.3.3 · p. 6
  • The what-if panel overlays a surrogate-projected trajectory on the original estimate; reducing surgery-related activity by 20% from week 0 shifts projected average risk upward.This comparison explores model sensitivity rather than providing a clinical prediction or recommendation.Source: Figure 3 · p. 6, §3.3.3 · p. 6

4 CASE STUDIES

The case studies use coordinated views to compare temporal risk and care-team patterns, first for colorectal patients with similar network size and then for clinically matched patients. These exploratory comparisons surface differences in risk trajectories, team composition, and co-access structure without establishing causal effects.

4.1 Exploratory Comparison: Divergent Risk at Similar Care-Network Size

  • In the exploratory pair, average model-estimated risk was 86.5% versus 46.0% despite both patients having 230 unique HCPs and average team volatility of 2.0/week.The higher-risk trajectory remained largely red, while the lower-risk trajectory shifted toward green over treatment.Source: §4.1 · p. 6
  • At week 9, the lower-risk patient had 36 active providers across 11 HCP types versus 22 providers across 2 types, with a broader set of prominent risk-lowering features.The higher-risk patient’s feature profile was narrower and included a positive contribution from Family Practice.Source: §4.1 · p. 7
  • A 20% reduction in surgery-related activity raised deceased-patient surrogate-projected risk from 86.5% to 96.9% (approximately 10.4%) and surviving-patient risk from 46.0% to 50.6% (+4.60%).This was a what-if perturbation of the surrogate model, not an observed intervention.Source: §4.1 · p. 7

4.2 Expert-Guided Clinically Constrained Comparison

  • The second case selected same-stage Stage 2B patients with comparable provider counts and note volumes but average model-estimated risks of 96.6% and 6.1%.The deceased patient had 272 HCPs and approximately 22K notes; the surviving patient had 269 HCPs and approximately 21K notes.Source: §4.2 · p. 7
  • At a selected week, Patient B had 8 active HCPs, 9 co-access edges, and 7 HCP types, compared with Patient A’s 4 active HCPs and 1 edge.Linked views also showed multiple specialty-related features associated with lower model-estimated risk for Patient B.Source: §4.2 · p. 7
  • Across the cumulative networks, Patient B had 61 co-access edges versus Patient A’s 40 despite spanning 28 specialties rather than 32 and a shorter observed timeline.The timelines were 51 weeks for Patient B and 58 weeks for Patient A.Source: §4.2 · p. 7
  • The second case reproduces a structural contrast in an independently collected cohort, offering preliminary evidence it is not an artifact of a single site; increased activity may also reflect complications or clinical complexity.The authors call for follow-up analyses accounting for comorbidities, treatment pathway, hospitalization history, and major-event timing; denser connectivity is not evidence of improved survival.Source: §4.2 · pp. 7–8

4.1 Exploratory Comparison: Divergent Risk at Similar Care-Network Size

  • Fig. 6 contrasts a more interconnected lower-risk network, with surgical and medical oncology tightly integrated, against a sparser higher-risk network less integrated around surgery-related specialties.Node size and week count encode provider activity duration; edges connect providers who co-accessed notes, with thickness encoding repeated co-access.Source: Figure 6 · p. 8

5 EVALUATION

The evaluation combines expert review with clinical literature to assess whether the system supports plausible interpretation and hypothesis generation. Expert feedback shaped the comparison workflow and emphasized that the visualizations should explore co-evolving risk and care-team activity, not explain survival causally.

5 EVALUATION

  • The evaluation paired qualitative expert review with clinical-literature context to assess interpretive usefulness, not to establish definitive clinical conclusions.The review also identified constraints on how analyses should be framed.Source: §5 · p. 8

5.1 Expert Review Protocol

  • Two live expert sessions involved 13 participants and 11 participants, including clinicians, informatics and health-services researchers, and visualization and machine-learning team members.Some experts attended both sessions.Source: §5.1 · p. 8

5.2 Expert Feedback Themes

  • Clinician feedback led the follow-up comparison to match patients on cancer stage and overall EHR activity before temporal and structural drill-down.Suggested matching factors included age, gender, treatment pathway, time to first treatment, and first-therapy type.Source: §5.2 · p. 8
  • Experts framed the system as a tool for exploring how model-estimated risk co-evolves with EHR-mediated care-team activity and generating coordination hypotheses, rather than directly explaining survival.They valued linked views for relating risk, specialty composition, attributions, and co-access at the same time point.Source: §5.2 · p. 9

5.3 Connecting Expert Feedback to Colorectal Care-Team Literature

  • Access logs are digital traces of care activity, not direct evidence of communication quality, clinical intent, or causal effect.The authors describe follow-up matched-cohort analyses with richer clinical variables and downstream care-delivery outcomes as a way to test coordination hypotheses.Source: §5.3 · p. 9

6 DISCUSSION AND LIMITATIONS

The framework supports comparative and longitudinal analysis of care-team patterns alongside estimated risk, but its findings are exploratory rather than clinical conclusions or evidence of causal effects. Its scope is bounded by EHR-log proxies, illustrative case studies, and the need for stronger validation.

Discussion

  • The workflow links risk estimates, surrogate explanations, specialty composition, and HCP co-access to support longitudinal reasoning about care-team activity.Case studies and expert feedback show it can help analysts move from aggregate comparisons to temporally specific questions about care-team organization.Source: §6 · p. 9

Limitations

  • Observed co-access networks are proxies for EHR-mediated collaboration, not complete records of care-team interaction.EHR logs omit face-to-face discussion, calls, secure messages, informal coordination, and clinical intent; documentation and institutional workflows also affect observed activity.Source: §6 · p. 9
  • The case studies illustrate the analytical workflow rather than establish controlled clinical comparisons.Stronger validation would require larger matched-cohort analyses with richer clinical variables to account for patient complexity.Source: §6 · p. 9
  • Surrogate attributions and what-if projections are feature-level associations, not clinical interventions or evidence of causal effects.Future work should test whether similar signals remain stable under alternative temporal graph models, explanation methods, and feature definitions.Source: §6 · p. 9
  • Future validation should examine endpoints tied more directly to care coordination, including emergency-department use, readmission, and treatment delays.These endpoints could help determine whether surfaced patterns reflect clinically meaningful differences in care delivery.Source: §6 · p. 9

7 CONCLUSION

The conclusion presents a visual analytics framework that aligns temporal risk modeling with evolving collaboration networks. Case studies illustrate how coordinated views can help analysts compare care-team patterns that static or aggregated summaries may obscure, for insight generation rather than definitive clinical conclusions.

7 CONCLUSION

  • The framework aligns temporal graph-based risk modeling with evolving collaboration networks and coordinates risk trajectories, specialty composition, explanations, and HCP co-access views.Source: §7 · p. 9
  • Case studies and expert feedback illustrate that coordinated analysis helps surface and compare care-team organization patterns difficult to observe in static or aggregated summaries.Source: §7 · p. 9
  • The framework is intended for insight generation and hypothesis development, not definitive clinical conclusions.Source: §7 · p. 9
Loading 2610.30748v1…