ClarityCare
A risk score is a black box that says “trust me.” I designed the glass box around it.
Care coordinators have to catch the patients most likely to fall through the cracks before the next crisis — but the tooling hands them an undifferentiated wall of EHR data and no idea where to start. ClarityCare is a self-directed concept for turning an explainable AI risk model into a workflow a clinician can actually trust and stay in charge of.
A self-directed concept, scoped like a real brief.
No client handed me this. I wrote the brief myself and held it to the same bar as paid work: a real user with a real deadline, a genuinely predictive — and genuinely opaque — model, and a product that has to earn a busy clinician’s trust in seconds.
A concept project — the ML model, SDOH indices, and named patients are illustrative sample data, assumed to make the argument concrete. Nothing here is staffed, validated, or measured.
A wall of data, and no idea where to start.
Care coordinators own the patients most likely to fall through the cracks — missing appointments, cycling through the ER, quietly losing housing or food access. They’re meant to catch them before the next crisis. The tool hands them an undifferentiated EHR and a to-do list that never ends.
The important patient is invisible in the data.
No way to see who it is.
No time to reason it out by hand.
Enhance judgment. Never replace it.
The north star: the AI proposes, the clinician disposes. If a screen ever made the coordinator feel like a rubber stamp for the algorithm, it was wrong.
Predicts risk, surfaces the reasons, and suggests a next step — always with its own uncertainty on the table.
Accepts, overrides with a structured reason, or adds context the model never had. The last word is human.
The biggest health risks often aren’t medical.
Two kinds of input feed the SDOH layer — what a coordinator asks the patient, and public data about the neighborhood they live in. All four are validated, publicly documented standards; the design work was translating them into something usable in the field.
PRAPARE
A national, standardized questionnaire covering housing, food, transportation, employment, and social support.
nachc.org/resource/prapare ↗CDC / ATSDR SVI
Social Vulnerability Index — ranks how vulnerable a neighborhood is using census data: poverty, vehicle access, crowded housing.
atsdr.cdc.gov/…/svi ↗CMS AHC HRSN
Accountable Health Communities screen — a short 10-item form for utility, personal-safety, and interpersonal needs.
cms.gov/…/ahcm ↗ADI
Area Deprivation Index — ranks neighborhood disadvantage from income, education, employment, and housing quality.
neighborhoodatlas.medicine.wisc.edu ↗The real work wasn’t scores. It was trust.
The model reads four sources — and every feature answers a skeptical clinician’s question:
→ a genuinely predictive model, and a completely opaque one.
The Smart Triage Queue answering “where do I start?” — the ranked wall, severity and its cause on every row.
Built on IBM Carbon — on purpose.
The product shouldn’t look like the tools it replaces. I chose Carbon — IBM’s open-source, publicly available design system — as the foundation for every ClarityCare screen, for four defensible reasons.
Fit
Purpose-built for dense, data-heavy enterprise tools — exactly what clinical triage is.
Access
Audited contrast & keyboard support out of the box — non-negotiable in a clinical setting.
Tone
Sharp, quiet, grid-driven — reads trustworthy and clinical, not consumer-playful.
Continuity
Same IBM Plex type family as this deck — the product threads through the whole story.
One week, blank page to system.
The move that made a week possible: settle the judgment first, produce second. Decide the hard things up front, and production moves without second-guessing.
The Smart Triage Queue, three days apart · sketch → wireframe → hi-fi Carbon.
Lead with severity — and its cause.
Oriented in five seconds. Every row carries the risk level, the drivers behind it, and the next action — so the coordinator reads who and why at a glance. Everything else lives behind filters.
SHAP output, rendered as a sentence.
No coefficients. The model’s reasoning is a plain-language sentence, the score breakdown sits beside it, and the confidence cue is honest — a stale data source is stated outright, and it invites the override rather than hiding behind a number.
The override is the training signal.
Control and learning, made the same gesture. When the clinician overrides, the structured reason doesn’t just log a disagreement — it becomes labeled feedback the model learns from.
Coordinator overrides
“This is wrong because…” — a required, structured reason.
Captured as labeled feedback
Structured reason + free-text note, attached to the case.
Model improves
The humans correcting it are the humans improving it.
Context, not just a number.
A single risk number is a snapshot. The timeline shows the story behind it — clinical events, the AI’s re-scores, and staff actions on one thread, color-coded by who acted, ending with the override that fed back into the model.
Nothing critical waits for a login.
Every alert is actionable in place — a risk jump, an overdue outreach, an override flagged for review — each with the one action it needs, not a feed the coordinator has to re-triage.
Thirty seconds of attention, in someone’s living room.
Mobile wasn’t a shrunk-down desktop — it was re-architected for the home visit: a coordinator with thirty seconds of attention, needing to re-score risk on the spot after entering a new SDOH fact. Optimized for speed, clarity, and a single next action.
Stacked filters, high-priority flags, one line of reasoning per patient.
Flip a toggle → a banner fires → the model re-scores on the spot.
Same product. Two postures.
Documented as an explicit contract — the analyst at a desk, the coordinator on a doorstep — so neither surface drifts from the shared logic.
Each model capability created an interface obligation.
Visual severity indicators
Plain-language insight cards
Transparency cues that invite override
Risk-type filters & segments
Structured override capture
Live re-prioritization
Three tradeoffs I’d defend in a critique.
Lead with severity + its cause; filter the rest.
Oriented in five seconds beats complete. A spreadsheet is complete and useless.
Show confidence honestly; invite the override.
A tool that hides its uncertainty earns less trust, not more.
The sentence a colleague would say.
No clinician should decode coefficients mid-shift to act on a score.
A point of view on AI in care.
Never make someone obey a model they disagree with.
Never waste the disagreement — it’s the training signal.
Be honest about uncertainty; encourage overriding a shaky call.
The clinician is accountable — so the clinician stays in control.
With AI in the loop, the interface’s job is to make the model accountable to the human using it.
Settle the hard calls, produce second — sequencing I’ll keep.
It’s the whole relationship between the clinician and the AI.
A black box turned into a glass box the clinician stays in charge of.
The deliverable is the argument and the high-fidelity system that embodies it: a triage queue, an explainable insight card, an override flow that keeps the clinician in charge, real-time SDOH entry, a timeline, and a field surface — a black box turned into a glass box.
A concept project with no clinician validation or live model — so there are no metrics, and I don’t invent any. Named patients are illustrative sample data.