Label balance
Inspect class coverage before interpreting a result.
04 / Evaluation
Measure what matters.
Evaluation is framed around useful separation, calibration, and failure analysis—not a single headline number.
MEASURED / MODEL 0187.73%Accuracy · Held-out test cohort
MEASURED / MODEL 0292.56%ICBHI score · Average sensitivity + specificity
These scores come from different tasks and split protocols. They are complementary evidence, not a direct ranking.
Research artifactInteractive figure
Research artifactInteractive figure
Interactive evaluation lab
DEMO DATARead the model under pressure.
Move the operating point, inspect class behavior, and compare how the model generalizes. Every chart below uses illustrative demo data.
Headline metric
Accuracy87.73%
Confusion matrix · threshold linked
Pred +Pred −True +8515True −1486
Normalized illustrative countsROC curvethreshold marker follows the slider
Training / validation losseight illustrative epochs
stableClass-wise performancemacro F1 by sound type
90.3%Decision threshold0.50
Patient-level separation
Keep recordings from the same patient out of both train and test evidence.
Calibration
Compare confidence with correctness before review.
Error review
Return to the recording and understand the model’s miss.
Prototype statusThreshold interactions are illustrative until linked to persisted evaluation artifacts.