Discrimination
How effectively the model separates patients with differing levels of observed risk.
Research & validation
Evidence should answer clinical and operational questions.
Our validation approach looks beyond a single performance measure to examine how decision support behaves across data, populations, time, and workflow.
We describe the current evidence status precisely: retrospective validation is complete, while external and prospective evaluation remain active areas of development.
Validation philosophy
Technical performance is necessary, but not sufficient.
A useful evaluation must consider whether a model performs consistently, whether its outputs can be interpreted in context, and whether the resulting experience fits the realities of clinical review. We treat those questions as connected rather than separate.
Our goal is to develop evidence that helps clinical and research partners understand where SEPTR may be useful, where uncertainty remains, and what should be tested next.
Current evidence status
A staged pathway from retrospective work to real-world evaluation.
Each stage is intended to answer a different set of questions. Progression depends on transparent methods, appropriate data, and study designs suited to the intended use context.
Complete
Retrospective validation
Initial retrospective evaluation has been completed as a foundation for further study. Detailed methods, cohorts, and findings should be reviewed in an appropriate research or partner setting.
In development
External evaluation
We are pursuing evaluation with additional data and care settings to examine generalizability and identify context-specific limitations.
Planned pathway
Prospective evaluation
Future prospective work is intended to examine model behavior, implementation conditions, and workflow utility in a real-world clinical environment.
Evaluation dimensions
A multidimensional view of performance.
No single metric can describe how a clinical decision-support system will behave. Our evaluation framework includes the following dimensions.
Calibration
Whether estimated risk remains aligned with observed outcomes across clinically relevant groups.
False-alert burden
How signal volume and specificity may affect attention, trust, and workflow.
Lead time
Whether information becomes available early enough to support meaningful clinical review.
Missing data
How availability, documentation patterns, and incomplete inputs influence model behavior.
Stability
Whether performance remains consistent as populations, practice, and data patterns change over time.
Workflow utility
How the product supports interpretation, prioritization, and next-step decisions in practice.
Limitations & interpretation
Evidence should be read in context.
Retrospective performance does not establish prospective clinical utility, and results from one dataset or setting may not generalize to another. Model behavior can be affected by population differences, clinical practice, documentation patterns, missing data, and changes over time.
SEPTR remains under development and evaluation. Its outputs are intended to support—not replace—independent clinical judgment, institutional protocols, or professional medical decision-making.
Research partnerships
The right collaborators sharpen the questions and the evidence.
We welcome conversations with health systems, researchers, and clinical leaders interested in external validation, study design, workflow evaluation, and responsible implementation planning.