Clinical Communication Engineering Foundation Standard

Research baseline only. This document does not prescribe care, clinical thresholds, or regulatory compliance. Every application requires local safety analysis and representative-user validation.

1. Clinical communication principles

1. Design for the decision, not the database. Begin with the reader's task, risk, and time horizon.
2. Preserve the evidence chain. Make source, author/system, observation time, update time, and transformation visible.
3. Separate epistemic classes. Visually and semantically distinguish observed fact, patient report, clinician interpretation, generated inference, plan, and unresolved question.
4. Lead with state and change. Show current severity, active threats, meaningful changes, contradictions, missing context, and next actions before stable history.
5. Layer, do not amputate. Compression must retain a path to source detail and make omissions inspectable.
6. Make ownership closed-loop. Every action needs owner, state, due condition/time, contingency, acknowledgement, and completion evidence where applicable.
7. Treat salience as scarce. Prominence is proportional to harm if missed, actionability, time sensitivity, and confidence—not simply abnormality.
8. Use redundancy for safety. Never rely on color, position, icon, typography, or sound alone.
9. Optimize for representative variation. Expertise, role, stress, disability, language, numeracy, device, and environment are design inputs.
10. Measure comprehension and error. Preference and aesthetics are secondary outcomes.
11. Expose uncertainty honestly. Unknown, not assessed, pending, contradicted, unreliable, and normal are distinct states.
12. Keep AI contestable. Generated content must be inspectable, attributable, editable under governance, and monitored.

2. Physician cognitive model

2.1 Reader loop

2.2 Expertise adaptation

• Experts often recognize patterns and seek discriminating cues; let them scan compact structure and reveal evidence quickly.
• Novices need explicit relationships and rationale; do not remove intermediate cues in pursuit of expert speed.
• Under stress, working context and attention narrow; keep identity, urgency, actions, and contingencies stable.
• Do not infer expertise solely from job title. Measure performance by task and setting.

2.3 Bias countermeasures

• Show disconfirming and contradictory evidence adjacent to the active interpretation.
• Label source and time to resist stale-data anchoring.
• Separate “not present” from “not assessed.”
• Make copied-forward and generated content detectable.
• Avoid sorting solely by newest, abnormal, or algorithmic score when it hides clinical relevance.

3. Information architecture standard

3.1 Canonical layers

1. Safety header: identity, encounter, location/context, allergies/safety constraints, data recency.
2. Now: severity, active threats, what changed, pending high-consequence items.
3. Working model: concise synopsis, active problems, confidence, supporting and contradicting evidence.
4. Action layer: decisions, owners, timing, contingencies, acknowledgement state.
5. Evidence layer: trends, results, medications, imaging, notes, procedures, and source documents.
6. Longitudinal context: baselines, resolved problems, prior episodes, social/family context when relevant.
7. Audit layer: provenance, edits, generated transformations, access/change history where required.

This is a semantic layering model, not a mandatory page order. Emergency triage may foreground threats; a referral may foreground question and prior workup; a patient result may foreground meaning and next steps.

3.2 Artifact contracts

3.3 Navigation and disclosure

• Preserve stable section names and anchors across devices and print.
• Make critical content available without hover, animation, or color decoding.
• Progressive disclosure may hide detail, never the existence of a critical state.
• Search results show context, source type, date, and matched passage—not title alone.
• Filters communicate active scope and never silently exclude critical items.

4. Visual and typography standard

4.1 Hierarchy

Use four intentional levels: safety/identity, section/task, record/item, metadata. Emphasis is earned by decision relevance. Avoid full paragraphs in bold, decorative uppercase, or simultaneous use of multiple emphasis channels.

4.2 Typography

• Use a highly legible system or tested clinical typeface with clear I/l/1, O/0, and punctuation differentiation.
• Body text baseline: 16 CSS px on interactive screens; allow user scaling to 200% without loss. Dense tables may use smaller text only after task testing and must retain a scalable alternative.
• Default line height: about 1.45–1.6 for prose; compact tabular rows may be tighter when row tracking is demonstrably reliable.
• Keep patient prose lines near 45–80 characters; avoid justified text.
• Use tabular numerals for aligned quantities; always display units and preserve meaningful precision.
• Use sentence case. Monospace is for identifiers/codes, not long clinical prose.
• These values are starting constraints from accessibility/legibility practice, not proven clinical optima.

4.3 Spacing and alignment

• Proximity indicates relationship; dividers do not compensate for ambiguous grouping.
• Align labels, values, units, reference context, and timestamps predictably.
• Whitespace separates decision groups; compactness is allowed within a group.
• Avoid card grids when comparison across rows or time is the primary task.

5. Color and urgency system

Color is supplemental. Every state also has a text label and at least one structural/iconographic cue.

• Do not map “outside reference range” directly to critical.
• Do not use green to mean globally safe; normal values can coexist with serious illness.
• Minimum WCAG contrast applies; critical data should target stronger contrast where practical.
• Validate in dark/light themes, grayscale, color-vision simulations, print, glare, and low-quality displays.

6. Clinical data presentation

Labs and vitals

Show value, unit, observation time, reference/target context and its source, trend, meaningful change, relevant interventions, and data-quality caveats. Distinguish physiological urgency from statistical abnormality. Never truncate units or silently mix unit systems.

Medications

Separate active, held, stopped, historical, and proposed. Show generic name, dose, route, frequency, indication when known, start/change time, reconciliation status, and unresolved discrepancy. Avoid color-only status and ambiguous abbreviations.

Imaging and procedures

Lead with impression and clinical implication only when authored/verified; retain full report, modality, body region, date, comparison, limitations, and pending status. Generated summaries must not masquerade as radiologist text.

Problems and diagnoses

Separate confirmed, suspected, ruled out, resolved, and historical. Show evidence, confidence owner, onset/update, and relationship to current decisions. Problem-list presence is not proof of current truth.

Timeline

Anchor events to an explicit time zone and distinguish occurrence, documentation, result, and communication times. Permit multi-scale views. Mark uncertainty in date and copied-forward entries.

7. Risk and uncertainty communication

• Prefer absolute risk with an explicit time horizon and consistent denominator.
• Present baseline and post-intervention risk together; do not lead with relative change alone.
• Use icon arrays or bar/number-line displays when tested for the audience; include accessible text equivalents.
• State confidence qualitatively only with a reason; do not fabricate numeric probabilities.
• Use controlled states: confirmed, probable, possible, unlikely, ruled out, unknown, not assessed, pending, conflicting, each governed locally.
• State what would change the assessment and what action follows uncertainty.

8. AI collaboration standard

Permitted candidate roles

Extraction, chronological organization, duplicate detection, discrepancy surfacing, draft summarization, plain-language translation, missing-context prompts, and retrieval—only within validated use cases.

Mandatory output envelope

Every generated artifact records model/system version, generation time, source scope, claim-level citations or links, transformations performed, unresolved contradictions, missing inputs, and reviewer state. Generated inference is distinct from source text.

Prohibited defaults

• Autonomous diagnosis, treatment, urgency, or disposition without authorized validated workflow.
• Silent replacement of clinician-authored text.
• Invented confidence, source, date, or negative finding.
• Hiding source disagreement to create fluent prose.
• Learning from patient data outside approved privacy/security governance.

Evaluation

Measure omission, unsupported claims, contradiction preservation, temporal errors, actionability, subgroup performance, automation bias, review time, correction rate, and downstream clinical outcomes. Monitor after deployment and provide rollback.

9. Print, mobile, and resilience

• Print preserves identity, generation time, page count, confidentiality marking, section continuity, and critical labels on every relevant page.
• Mobile preserves Now/Actions/Contingencies before deep detail and supports one-handed, zoomed, and interrupted use.
• Offline/export states communicate freshness and missing dynamic content.
• Never encode essential meaning only in interaction unavailable on paper or assistive technology.