The VBridge workflow supports clinicians in interpreting and validating machine-learning predictions made on Electronic Health Records (EHRs). It connects the dots between ML features, contribution-based explanations, and patients' original records, so that clinicians — who reason through domain expertise rather than technical model internals — can decide whether to trust a prediction and act on it. The running example is the prediction of five types of postoperative complications after cardiac surgery.
The workflow begins with an automated data-preparation pipeline. Cardiac-surgery patients are defined from relational EHR tables (with Surgery as the target entity), and Deep Feature Synthesis (DFS) engineers a per-patient feature vector while retaining a traceable path between each feature and its source records. An AutoML framework (Cardea) trains five binary classifiers, which are then applied to produce the patient's complication-risk predictions. SHAP values are computed and organised into a feature hierarchy to explain each prediction.
To make these explanations evidence-based, the clinician selects a reference cohort of similar patients (split into low-risk and high-risk subgroups), from which reference value ranges (95% confidence intervals) are computed for every feature. An occlusion-based, non-parametric procedure identifies the most influential time periods within dynamic signals. Four coordinated views are then rendered: the Profile View, the matrix-based Timeline View (records overview), the hierarchical Feature View (contributions plus reference values and what-if support), and the Temporal View (time-series records with reference bands, anomalies, and influential-period highlights).
The clinician enters an iterative investigation loop (illustrated here through the backward analysis workflow). They inspect the feature and temporal views to identify candidate risk factors, then assess whether the contributions align with their clinical judgment and the evidence. If the evidence is insufficient, the clinician refines the focus — sorting and filtering features, expanding the hierarchy, marking records, adding time series, brushing timeline cells, and running reference-value-based what-if analysis — which re-renders the feature, temporal, and timeline views. The cycle repeats until enough evidence is gathered. (The complementary forward analysis workflow follows the same machinery but starts from the records and proceeds toward the explanations.)
Finally, in a decision step, the clinician integrates the explanations, reference-based evidence, and record context to accept or reject the prediction, identify risk factors, plan care, and occasionally derive new clinical knowledge.