Skip to content
AUA 2027 Research Amendment: 5×5 Repeated Nested CV · Bitwise Reproducible

Explainable clinical decision support

Every prediction, traced back to the pressure that caused it.

CURES-AI-UDS reads a urodynamic study, classifies lower urinary tract dysfunction across nine overlapping diagnostic labels, and decomposes each prediction into the exact measurements that drove it — then drafts the report and stops, for a specialist to sign.

Research prototype · benchmarked on 527 studies from a released public pressure-flow dataset.

Pressure–flow study study 03 · held out

What happens next

Fifteen released variables, nine overlapping labels, and one requirement: show which measurement moved which prediction.

The problem

A model that scores well is easy. One that can be trusted is not.

Urodynamic feature tables contain summary variables that sit very close to the diagnostic constructs they are used to predict. A classifier trained on them can look strong while largely restating its own inputs — and a clinician has no way to tell the difference from an accuracy figure alone.

The pipeline

Five stages. One study. Nothing hidden.

Follow a single held-out patient through the whole system — raw trace to signed report. Every value on screen is read from the committed benchmark artifacts.

    Scroll to advance, or select a stage.

    Ingestion

    study 03 · held-out partition

    FILLING VOIDING
    storage phase statistics voiding phase statistics

    The released MATLAB structure is flattened into a tabular frame. Storage and voiding statistics are computed separately, then audited for label proximity before any model sees them.

    One-vs-rest across all nine labels — the released annotations overlap, so the model is never forced to pick a single diagnosis.

    Tree SHAP decomposes the leading prediction into per-measurement contributions, with direction preserved.

    toward the label away from it

    Draft output — not a diagnosis. Awaiting clinician sign-off

    Simulator

    Change the patient. Watch the reasoning change.

    Five real held-out studies, with their committed attributions and report drafts. Withdraw the label-proximal features and the same patient's reasoning visibly shifts.

    Live inference · committed artifacts

    —

    Attribution

    —
    Leading label

    —

    toward the label away from it

    Report draft

    Requires specialist review. Awaiting sign-off

    The approach

    Three decisions that make it defensible.

    01 / Audit

    Leakage-audited by construction

    Every released variable is classified — primitive, engineered, derived index, or label-proximal summary — with a written rationale, before modelling. The audit decides what the conservative analysis may see. Not a hunch.

    15 variables classifiedpublished rationale

    02 / Attribute

    Attribution on every case

    Global rankings say what matters on average. This produces per-patient attributions, so a prediction can be traced to the specific values that moved it, in both directions.

    tree SHAPsigned contributions

    03 / Stop

    It stops at the draft

    The pipeline composes a report that states its own uncertainty and terminates in a required human sign-off. Decision support, scoped deliberately — never autonomous diagnosis.

    clinician-in-the-loopauditable template

    Evidence · AUA 2027 Amendment

    How much predictive information do urodynamic summaries add?

    Our primary amendment benchmarks full urodynamic feature summaries against a routine clinical pressure-flow comparator (Age, Sex, Pdet@Qmax, Qmax) across 5×5 repeated nested CV, and examines feature-family withholding.

    Evidence path

    released cohort → attributed prediction

      The gain is concentrated in specific feature families. While full summaries provide an incremental +0.144 Macro AP over clinical core inputs (paired bootstrap 95% CI: +0.110 to +0.183), per-label ablation shows that this benefit is not uniform across diagnoses: the largest gains occur where released summaries are near-definitional to the label (e.g., VinfDec and ComplianceDec), whereas key obstruction and detrusor weakness diagnoses are already largely captured by routine pressure-flow parameters.

      Scope & safety

      Stating the boundary precisely is part of the work.

      What it is

      • A modular, explainable classification prototype
      • A leakage-aware benchmark on a public dataset
      • A clinician-in-the-loop reporting workflow
      • Groundwork for raw-trace and PDF ingestion

      What it is not

      • An autonomous diagnostic system
      • A substitute for specialist interpretation
      • Prospectively or clinically validated
      • A medical device, or cleared as one

      Team

      Built by two engineers.

      Read the code, not just the claims.

      The pipeline, the leakage audit and every artifact behind this page are public. Reproduce the benchmark in one command — or challenge the analysis directly.