What Gets Measured

An exception-led clinical benchmarking command center. Faculty performance at an academic medical center spans 4 incommensurable domains that cannot be honestly collapsed into one score, and when the exception gets averaged away it becomes a care-quality, equity, or workload risk that surfaces too late to act on.

Healthcare analytics
Clinical operations
Data visualization
Concept prototype

4 domains, one screen, no blended score

01 · Premise

The design problem is how to make 4 incommensurable things comparable without flattening them into 1 number.

Faculty performance at an academic medical center is the contested intersection of 4 incommensurable domains: clinical productivity, research output, financial contribution, and educational impact. A high-volume surgeon fills OR hours and trains fellows. A principal investigator on 3 grants earns the institution funding and intellectual standing that no collection ratio can capture. A primary care physician who sees twice the patient volume of her colleagues but publishes nothing is, by different measures, both the most and least valuable person on the department’s roster.

I was brought in to design an interface that could show all of these dimensions simultaneously, compare any physician against any reference point, and support decisions about resource allocation, promotion, and strategic planning without pretending the underlying tensions had been resolved. The constraint was preserving 101 physician-level records and 4 parallel measurement axes on one legible screen.

Healthcare benchmarking sits under HIPAA-era data-sensitivity constraints. The interface had to make physician-identifiable data legible to authorized reviewers while keeping every public demo fully de-identified. That discipline carries over into any domain where data visualization and clinical analytics touch protected information.

101 Physicians in the benchmark set

4 Years of monthly historical data

8 Performance axes mapped to 5 chart types

4 Incommensurable domains kept separate, never blended

Engagement: Product designer and sole UX lead. Brought in by a leading New York City cancer research institution to design a faculty performance benchmarking interface spanning clinical, financial, research, and educational dimensions simultaneously.
Owned: I mapped the 8 performance axes to 5 chart types, designed and validated the parallel-coordinates compound brush, and set the HIPAA de-identification rules for the public demo.
Provenance: Concept prototype. Every public demo runs on simulated data, not production metrics.

3 other showcases in this portfolio deal with the same underlying problem: when dense information is misread, the resulting decision becomes expensive to reverse. Uncertainty as Signal keeps probabilistic climate risk legible for capital allocation. North Bridge keeps workflow signal tied to company context and follow-through. HR as DJ turns staffing uncertainty into actionable allocation moves. This showcase brings that same design problem into clinical operations, where missed exception patterns create care-quality risk.

One chart, one comparison question

02 · Core insight

The design principle that unified the entire interface was this: every visualization answers exactly one comparison question.

Bar chart: How does this year compare to prior years, for the same metric?
Treemap: How does this service compare to all others, by area?
Parallel coords: How does this physician compare to every other across 8 dimensions?
Spider diagram: How does this individual compare to the department average?
Bubble chart: How does financial efficiency compare to revenue capture, per service?

5 visualizations. 5 comparison questions. No chart that answered more than 1, because a chart that tries to answer 2 usually answers neither cleanly. Benchmark-driven comparison is the chart type and the comparison question being the same decision.

The compound brush

03 · Mechanism

Before: a single continuous brush

The core interaction is the parallel coordinates brush filter, showing every physician as an individual polyline crossing 6 clinical axes simultaneously. The brush let users hold 1 dimension constant while immediately observing whether that dimension correlated with others or was entirely orthogonal. That is analysis at the precision of a continuous brush, not the coarseness of a dropdown.

Fig. 01 Continuous brush on parallel coordinates

Parallel coordinates in D3.js. Each line is a simulated entity; each vertical axis is a clinical dimension. Brush filtering holds one dimension constant and reveals how the others move with it. Simulated data, not production metrics.

14 entities · brush any axis to filter

After: compound brush composition

The current case study extends the brush into compound cohort definition: OR within an axis, AND across axes, with a readable audit trail after overlaps resolve. Multiple ranges can live on one axis, colliding ranges merge automatically, and brushes across axes combine into a single cohort definition.

Fig. 02 Compound brush composition

The upgraded brush model turns parallel coordinates into cohort definition: multiple ranges per axis, collision merging, cross-axis composition, and an audit trail of active ranges. Simulated data, illustrative.

Two independent Equity ranges qualify entities before the Workload gate is applied.

Demo scenes

Before: hold Workload high · hold Care Quality high · reset
After: split Equity ranges · collide and merge · compose across axes · reset

What changed for the reviewer

In pilot review of the rebuilt interface, reviewers used the compound brush to define a physician cohort in one move where the prior dashboard needed 3 separate filter steps, observed in walkthrough, not a timed study.

Behavior shift the case study is designed for: before, reviewers navigated view → chart → filter to find outliers; after, exceptions surface ranked on load and the brush defines a cohort in one gesture. This is the intended behavior of the design, not a measured result.

The Benchmark Comparison Engine

04 · Scenario lenses

The parallel coordinates survived and deepened. The Benchmark Comparison Engine is the same analytical operation, hold one axis and observe all others, rebuilt in D3.js v7 with 6 clinical dimensions and overlaid reference lines for the review target and the scoped service mean.

4 scenario lenses scope that comparison. Each is a named review scenario that applies service-level and patient cohort adjustments before anything is ranked.

Quarterly Benchmark Review
Capacity Strain Watch
Access Parity Review
Quality Drift Watch

Fig. 03 Clinical Benchmark Command Center

An exception-led workspace where benchmarking, comparison, ranking, and cohort review are one loop. Simulated data, not production metrics.

Build

Current prototype: single-file HTML and D3.js v7, with no framework or build-tool dependency
5 coordinated analytical views: grouped bar, treemap, parallel coordinates, spider or radar, and bubble chart
101 physicians, 8 performance axes, 4 years of monthly historical data
Parallel coordinates on Canvas with live brush filtering across 6 clinical dimensions
4 named review scenarios with service-level and patient cohort adjustments
Exception surfacing: alert feed ranked by severity, drill-down panel, burnout risk toast
CSS custom properties and @layer specificity management, with no !important in prototype code

Burnout risk and patient equity are first-class signals

05 · Signals

They are treated as operational risk indicators, not secondary annotations. The domain model covers care quality, access reliability, equity gap closure, workload pressure, value realization, and team support. Each metric answers a reviewer question instead of becoming one blended score.

The interface is exception-led. Ranked alerts, selected exception context, and burnout-risk signals appear first because exploration should follow the signal. Putting equity and burnout on the same footing as productivity is what stops them from being the axes that get averaged away.

The metrics chosen for any performance dashboard are not neutral. They are an institutional statement about what gets valued. What gets measured defines what gets managed. A well-designed interface makes that argument legible rather than hiding it in a formula.

The decision-intelligence loop

06 · Loop

The interface is not 5 disconnected charts. It is 1 loop, and the order is the argument.

Benchmark Multi-dimensional benchmarking establishes the reference points for the cycle. review target · scoped service mean
Compare Parallel-coordinates comparison holds one axis constant and reads the rest. 6 clinical dimensions
Surface the exception Comparison surfaces exceptions where an entity diverges from its cohort. divergence from cohort
Rank it Exceptions become a ranked alert feed ordered by severity. severity order
Review the cohort Reviewers run cohort and access-parity review on the exception in context. cohort review · access parity
Commit a decision Reviewers commit a decision with an audit trail, and that decision tunes the next benchmark cycle. audit trail · feeds the next cycle

The same loop underlies exception monitoring, operational-risk review, and model-risk governance.

Design trade-offs

07 · Trade-offs

3 were made on this build, and each one bought something specific.

Continuous brush, not a dropdown

Traded a simpler control surface for analysis at the precision the comparison needs, holding one axis constant while reading the rest.

Reference lines, not one composite score

Traded a single sortable number for a comparison that survives review, because incommensurable axes should not collapse into one score.

Canvas for 101 polylines

Traded native per-line accessibility for every physician rendered at once, with the production path then owing the semantic summary and keyboard filtering noted below.

Accessibility and compliance posture

Canvas polylines are not screen-reader readable. A production version owes keyboard filters, a semantic summary table, focusable controls, color-independent severity, and reduced-motion support. This page is a prototype, designed to WCAG AA principles but not VPAT-certified. Naming the gap is more useful than overclaiming a compliance the prototype has not been audited against.

Clinical pattern, fintech equivalent

08 · Transfer

The transfer is concrete. The same interface operations map to risk, surveillance, fair-lending, and portfolio-review workflows.

Parallel coords: Multi-factor portfolio risk review.
Ranked exceptions: AML, fraud, and trade-surveillance alert triage.
Access parity: Fair-lending disparate-impact review, and model fairness.
Workload pressure: Operational-risk monitoring.
Bubble efficiency: Risk-return attribution.
Treemap contribution: Exposure sizing.

Ranked exceptions, cohort comparison, confidence cues, and audit-ready drill-downs are the portable parts of the workflow. They let a reviewer move from signal to context without flattening the decision into a single score. The same structure applies when the domain moves from physician benchmarking to portfolio risk, AML triage, fair-lending review, or surveillance operations: define the comparison set, surface the outlier, preserve the rationale, and make the final call reviewable.

The domain changes; the comparison, exception, and audit pattern stays the same. That is why a designer who can work in healthcare can also work in capital markets: the constraints differ, but the underlying decision geometry does not. 15 years of practice in this domain, from Protovis to D3, from reporting tools to command centers, from financial metrics to equity signals, produced that point of view.

Illustrative transfer map, no production metrics. Clinical signals are shown only as origin patterns; the transferable work is the decision interface.