Source-linked AI summary
VBridge: Connecting the Dots Between Features and Data to Explain Healthcare Models
Furui Cheng, Dongyu Liu, Fan Du, Yanna Lin, Alexandra Zytek, Haomin Li, Huamin Qu, Kalyan Veeramachaneni
TL;DR
Clinicians face barriers to using ML explanations in clinical decision-making, including unfamiliar features, missing patient context, and a need for cohort-level evidence. VBridge connects feature explanations with patient records in clinician-informed workflows; case studies and interviews suggest that these associations can help clinicians interpret and use predictions.
Problem
Clinicians may find ML features unfamiliar, lack contextual information, and need cohort-level evidence to assess feature-contribution explanations in clinical settings.
Source: §1 · p. 2Method
VBridge is a clinician-informed visual analytics system that connects feature explanations with patient records through forward and backward analysis workflows.
Source: §1 · p. 2Results
Takeaways & Limitations
Visually associating model explanations with patients’ situational records can help clinicians better interpret and use predictions.
Source: §8 · p. 9VBridge focuses on interpretable features with clear meanings and links to relevant health records; it may be less effective for complex features built from representation learning.
Source: §7.2 · p. 9
Abstract
from arXiv · showhide
Machine learning (ML) is increasingly applied to Electronic Health Records (EHRs) to solve clinical prediction tasks. Although many ML models perform promisingly, issues with model transparency and interpretability limit their adoption in clinical practice. Directly using existing explainable ML techniques in clinical settings can be challenging. Through literature surveys and collaborations with six clinicians with an average of 17 years of clinical experience, we identified three key challenges, including clinicians' unfamiliarity with ML features, lack of contextual information, and the need for cohort-level evidence. Following an iterative design process, we further designed and developed VBridge, a visual analytics tool that seamlessly incorporates ML explanations into clinicians' decision-making workflow. The system includes a novel hierarchical display of contribution-based feature explanations and enriched interactions that connect the dots between ML features, explanations, and data. We demonstrated the effectiveness of VBridge through two case studies and expert interviews with four clinicians, showing that visually associating model explanations with patients' situational records can help clinicians better interpret and use model predictions when making clinician decisions. We further derived a list of design implications for developing future explainable ML tools to support clinical decision-making.
1 INTRODUCTION
EHR prediction models and explainability techniques have potential in clinical care, but clinicians face barriers to using feature-contribution explanations in practice. VBridge was designed with clinicians to connect model explanations to patient records and support decision-making workflows.
1 INTRODUCTION
- Clinicians may struggle to interpret engineered ML features, connect feature contributions to original records, and judge predictions against cohort-level evidence.For example, aggregate values such as Trend can be unfamiliar; the authors propose comparing patient feature values with reference values from similar patients.Source: §1 · p. 2
- Six clinicians with an average of 17 years’ experience informed VBridge’s design requirements and forward and backward analysis workflows.The requirements and workflows guided the system’s design and development.Source: §1 · p. 2
- VBridge combines SHAP feature contributions, hierarchical feature display, reference-value and what-if analysis, and links to influential records in coordinated views.Deep Feature Synthesis provides traceable paths between features and raw records.Source: §1 · p. 2
- Two case studies and expert interviews with four clinicians demonstrated the system’s usefulness and efficiency.The authors report that visually associating explanations with patients’ situational records can help clinicians interpret and use model predictions.Source: §1 · p. 2
2 RELATED WORK
Related work covers approaches to explaining clinical prediction models and visualizing EHR data. VBridge builds on post-hoc feature contributions and patient-record visualization, while integrating cohort reference values with explanation views.
2.1 Explainable Machine Learning in Clinical Predictions
- Intrinsic models offer direct explanations, but their performance can be bounded on complex tasks; added ensemble complexity may impair interpretability.Attention-based explanations can also leave unclear how highlighted historical data relate to predictions.Source: §2.1 · pp. 2–3
- Post-hoc methods explain existing black-box models, and feature-contribution approaches such as SHAP support local prediction explanations.VBridge uses SHAP to compute each feature’s contribution and presents those contributions in a scalable, user-friendly visualization.Source: §2.1 · p. 3
2.2 Electronic Health Records Visualization
- EHR visualization research spans individual records and multiple-patient data; VBridge combines patient-record views with cohort-derived reference values for feature analysis.Its reference-value visualizations are integrated with feature explanations and raw-record views.Source: §2.2 · p. 3
3 INFORMING THE DESIGN
A clinician-informed pilot study shaped VBridge’s requirements and workflows. The resulting design organizes feature exploration, record exploration, and the connections between them.
3.1 Pilot Study
- The pilot involved six clinicians and used interviews, a low-fidelity mockup, and a later prototype exploration to refine design requirements.Participants included two chief physicians and four residents; three explored the prototype in prediction tasks.Source: §3.1 · p. 3
3.2 Design Requirements
- Seven requirements span hierarchical features, cohort reference values, flexible and what-if exploration, patient record overviews and details, feature-record links, and influential temporal patterns.The requirements were grouped into feature, data, and bridge explorations.Source: §3.2 · pp. 3–4
3.3 Analysis Workflow
- Forward analysis starts from patient records and clinicians’ own predictions, then checks model predictions and explanations; disagreement sends them back to relevant records.Clinicians may accept or reject a prediction, or gain knowledge from the evidence and reference values.Source: §3.3 · p. 4
- Backward analysis starts with predictions and feature explanations, then traces selected features to records; if those records are insufficient, clinicians consult broader summaries for related data.This workflow is useful when clinicians begin without a clear diagnostic prediction or investigate unfamiliar features such as SD or Trend.Source: §3.3 · p. 4
4 PREDICTIVE MODELING
The paper describes the pediatric ICU dataset and a surgical-complication prediction example, then explains how patient records become model features and how influential records are identified.
4.1 Data
- The PIC Database contains over 14,000 admissions from 12,000 pediatric patients admitted to the critical care unit between 2010 and 2019.It includes demographics, surgery information, vital signs, laboratory results, symptoms, medications, diagnostic codes, and mortality.Source: §4.1 · p. 4
4.2 Running Example: Surgical Complication Prediction
- The running example predicts five postoperative complication types for cardiac-surgery patients using five binary classifiers.The modeling cohort comprised 1,826 patients, of whom 456 (25.0%) developed complications; Cardea evaluated eight classifiers using averaged AUC across 10 cross-validation folds.Source: §4.2 · p. 4
4.3 Feature Extraction
- Deep Feature Synthesis follows relationships across EHR tables to create patient features with traceable paths back to source records.The extraction targets surgery and includes patient profile, surgery information, low-resolution lab and chart events, and high-resolution vital signs.Source: §4.3 · p. 5, Figure 2 · p. 5
4.4 Feature Explanation
- For dynamic features, the method occludes record windows, recalculates feature values, and accumulates each record’s relative influence on the feature.Window-level linear curve fitting smooths replacements, while traceable paths allow real-time feature recalculation.Source: §4.4 · p. 5
- A non-parametric dynamic threshold identifies influential periods by maximizing the percent decrease in the influence array’s mean and standard deviation after above-threshold values are removed.The resulting continuous sequences represent the most exceptionally influential segments for a feature.Source: §4.4 · p. 5
5 VBRIDGE
VBridge links patient records, feature contributions, and model predictions through an interface supporting forward and backward analysis. Its views provide feature-level explanations, clinical-record context, and interactions for exploring their connections.
5.1 System Overview
- VBridge’s storage, analysis, explainer, and interface modules support cohort-based reference values, real-time what-if analysis, SHAP contributions, influential-period detection, and forward or backward workflows.Storage holds original records, patient feature vectors, and predictions.Source: §5.1 · p. 6
5.2 Feature View
- Feature rows show values and contribution magnitude and direction; hierarchical groups sum included feature contributions, letting clinicians inspect features at different levels.Reference ranges use the 95% confidence interval of the selected cohort’s low-risk group; what-if analysis changes one abnormal feature to fit that range and displays resulting prediction and contribution changes.Source: §5.2 · p. 6
5.3 Temporal View
- The Temporal View adds selected clinical time series to contextualize feature explanations, marking out-of-range records and influential, potentially overlapping time segments.The chosen highlighting design combines advantages of alternatives for showing influential regions.Source: §5.3 · pp. 6–7
5.4 Timeline View
- The Timeline View summarizes a patient’s health records and supports both forward analysis and adding records to the Temporal View for comparison.Its matrix summarizes events from different sources across equal time intervals; cell color encodes event count and inner-box width encodes the proportion with out-of-range values.Source: §5.4 · p. 7
5.5 Interaction
- Colored source bars and curves connect feature elements to medical-record elements, while pins mark records and highlight associated features for forward exploration.The links update when users scroll or add time-series records.Source: §5.5 · p. 7
6 EVALUATION
Two case studies with clinicians familiar with VBridge examined whether its backward- and forward-analysis workflows could support clinical decision-making. In both, connecting feature explanations with patient records helped clinicians interpret predictions and refine their judgments.
6.1 Case Study I - Backward Analysis
- P5 found surgery time and CPB time among the top-contributing features; normalizing surgery time reduced its contribution, but did not noticeably change the prediction.Source: §6.1 · pp. 7–8
- Tracing Lactate, Oxygen Saturation, and Pulse records led P5 to reject CPB as the start of lactate accumulation, while he generally judged the prediction trustworthy.Source: §6.1 · p. 8
- P5 understood the features behind the prediction, explored their records, and decided to monitor the patient more closely and consider proactive postoperative treatment.Source: §6.1 · p. 8
6.2 Case Study II - Forward Analysis
- After initially finding no solid evidence of risk, P1 used feature contributions and influential periods to examine the prediction and sought more context for the End-Tidal CO2 pattern.Source: §6.2 · p. 8
- P1 found high contributions from Systolic Blood Pressure, COHb, and pre-surgery RDW; COHb and RDW prompted further clinical hypotheses and interest in follow-up study.Source: §6.2 · pp. 8–9
- P1 judged the patient likely to develop lung-related complications, decided to pay closer attention, and considered studying COHb further.Source: §6.2 · p. 9
7 DISCUSSION
Case studies and interviews with four clinicians informed VBridge’s design implications and suggested it could support clinical decision-making. The discussion identifies limits involving feature interpretability, cognitive bias, EHR data quality, reference values, visual scalability, and generalizability.
7.1 Design Implications
- Clinicians saw VBridge as useful for diagnosis, workflow improvement, team communication, and helping junior doctors; reference-value explanations were easier to understand and accept than contribution scores alone.The reference-value example describes a feature below its reference range as contributing highly to a patient’s cardiac complication.Source: §7.1 · p. 9
- Clinicians praised the feature hierarchy for avoiding unnecessary detail and valued visual links between explanations and context, including the ability to mark explanations for later inspection.There is no standard for healthcare-feature hierarchies; one clinician suggested borrowing ideas from clinical communication forms.Source: §7.1 · p. 9
7.2 Limitations and Future Work
- VBridge focuses on interpretable features linked to relevant health records; connections for harder-to-understand features, such as representation-learning features, can be complex and make the system less effective.Tracing and storing such complex connections remains to be explored.Source: §7.2 · p. 9
- Potential confirmation bias from backward-oriented reasoning and the effects of analysis workflows on clinicians’ decisions were not fully evaluated; the authors plan to assess decision precision.The cited work suggests that first seeing predictions and then seeking supporting evidence may lead to confirmation bias.Source: §7.2 · p. 9
- Misrecorded EHR data can produce false-positive patterns, and clinicians’ prior knowledge is currently required to detect these defects.One apparent cardiac-arrest pattern was traced to a faulty sensor.Source: §7.2 · p. 9
- Reference-value precision remains under development, with planned cohort recommendations, time-varying values for temporal records, and experiments on stability as cohorts change.Time-varying values are intended to better fit surgical scenarios composed of multiple stages.Source: §7.2 · p. 9
- The temporal view scales poorly for signals with many records, while more test-item rows make exploration less efficient by requiring more scrolling.Planned extensions include multiscale signal segmentation and search and filtering.Source: §7.2 · p. 9
- VBridge can support other prediction problems and models using the PIC dataset, but other EHR datasets require adaptations to entity and feature descriptions.The authors plan to define system inputs using the FHIR standard to improve generalizability.Source: §7.2 · p. 9
8 CONCLUSION
The paper identifies challenges in applying ML to clinical decisions and presents VBridge, designed around requirements from a pilot study. Case studies and interviews with four clinicians suggest that linking explanations to patient records can help clinicians interpret and use predictions in clinical decisions.
8 CONCLUSION
- Clinicians face unfamiliar ML features, insufficient context, and a need for cohort-level evidence when using ML in clinical settings.These are the three challenges identified by the authors.Source: §8 · p. 9
- VBridge is a visual analytics system designed to support clinical ML decisions through forward and backward analysis workflows.Its design followed requirements identified in a pilot study.Source: §8 · p. 9
- Two case studies and interviews with four clinicians suggest that visually associating explanations with patients’ situational records can help clinicians interpret predictions and use them in clinical decisions.Source: §8 · p. 9