A panel that existed. The usability that did not.

A two phase research arc

The study followed a generative to evaluative arc, built around three research questions. Why does a fully shipped panel go unused? What data does each clinical role actually need and at which moment of a live interaction? And how should that data be structured so retrieval survives the cognitive load of real triage?

Phase 01 mapped the operational problem space with an open mind rather than validating existing feature hypotheses. Phase 02 stress tested the structures that discovery produced. Generative AI tools were built into the analysis to process large volumes of qualitative field records cleanly, which freed me to focus on synthesis and interpretation.

The five panel environment

The core health software is a five panel workspace for supervising asynchronous clinical protocols. Panel 02, the Summary Panel, was built to give the primary clinical overview. It was fully live across teams, yet field observation showed its baseline data did not support real time triage.

Uncovering operational realities

The field study opened with an investigation into operational barriers. The workspace was available, but interaction stayed minimal. The goal was to diagnose the gap between the engineered layout and the way clinicians actually worked.

Methods deployed in the field

• Observational tracks

  Contextual inquiry

  Shadowed staff during high velocity consultation sequences. This exposed how they managed information hierarchies, where transitions broke and where operators left the platform to use their own record fragments.

• Targeted assessment

  Semi structured interviews

  Individual sessions built around critical scenario walkthroughs, tracing user motivations and local data adaptations without prompting feedback on layout limits.

• The shadow tracking discovery

  Artefact audit and shadowing

  Traced peripheral tooling to auxiliary logs and uncovered an unmapped 49 column spreadsheet maintained by hand across teams to fill the platform gaps.

Turning field notes into structural patterns

Before any automated processing, the first pass coded transcripts and logs by hand. That immersion kept the later structural models calibrated to real clinical constraints.

How generative AI was used

Validating the structure with the people who rejected the last one

With the schema drafted, testing moved to readability. The priority was confirming the structure held up under the cognitive load coordinators carry during emergency triage.

Evaluative methods

• 18 structural sessions

  Open and closed card sorting

  Participants organized the 49 metrics from the field study into intuitive clusters. This surfaced the clearest divergence of the study: doctors prioritize longitudinal diagnostic trends, while coaches need immediate habit adherence metrics.

• Think aloud tracks

  Low fidelity validation models

  Early wireframes ran under time constrained simulation. I watched navigation paths, logged search patterns and resolved friction before any high fidelity styling began.

The redesign, side by side with what it replaced

The findings converged on a three tier structure that separates variables by the operational moment they belong to inside a live clinical interaction: what you need in the first five seconds, what you reach for during the conversation and what you only open in a full review.

Before · one endless pane

The live panel: identity fields up top, then an undifferentiated scroll of topic snippets. Every retrieval was a hunt.

After · the three tier panel

Client dataBK

Tier 1 · Case summary

Client name, age, genderBarbie Kong, 25, F

Next actionAsk her not to eat pastries

Medical status, phase and dayS2, V1, D26

Progression score4

Diagnosed conditionsSerotonin deficiency, fat loss

Active medical concernsGERD, low energy

Current medications5 HTP 100mg, Antoxid HC

Dietary preferenceNon vegetarian

Country, timezoneIndia, GMT+5:30

Tier 2 · Further details

Type the data you are looking for

Client's initial data ⌃

Height 167 cm

Start weight 90 kg

Target weight 50 kg

Work schedule Mon to Sat, 11am to 7pm. Unpredictable

Food and diet ⌄

Active patient management ⌄

Surveillance data ⌄

Client status and workflow ⌄

Tier 3 · Deep archives

Reports, lab timelines, revision history ⌄

The redesigned panel: a five second case summary, searchable accordion detail and archives kept out of the way.

Closing the loop after rollout

Retrieval performance in testing was the leading indicator. The lagging indicator was behavioral: whether teams would abandon the shadow sheets they had built to survive the old panel.

Half the organization's care teams no longer maintain any external tracking sheet. The remaining teams are being migrated as the panel rolls out across the rest of the org.

Matching the layout to how clinicians think

The defining insight was structural, not visual. Teams often assume low adoption comes from visual friction and reach for a cleaner skin or more charts. Here the care team bypassed the platform because its data architecture did not mirror their professional cognitive sequence.

By finding the 49 column shadow spreadsheet inside their daily routine, the research did more than inform a dashboard redesign. It folded the team's informal operating system back into the core application.