Project 01 · Philips · Georgia Tech HCI Capstone · Spring 2026
Why clinical AI goes unused and whether fixing it is worth the investment.
Clinical imaging specialists were underusing Philips AI diagnostic tools — not due to capability gaps, but trust deficits. An HFE & HCI lens on physician distrust, mapped to ROI.
影像临床医生对飞利浦AI诊断工具的使用率偏低——根源不在能力缺口，而在信任缺失。从人因工程与人机交互视角系统分析医生不信任AI的行为成因，并构建信任干预措施与ROI的联合评估框架。
Human Factors Engineering
HCI
Strategy
Team of 3
02 / Problem Statement
Users struggle to trust AI with their health
用户在健康领域难以信任AI
AI-assisted health consultation apps are rapidly growing in popularity, bringing both significant opportunity and new challenges to product design. The core tension is not capability — it's calibration.
AI辅助健康咨询应用快速兴起，带来巨大机遇的同时也为产品设计引入了新挑战。核心矛盾不在于技术能力——而在于信任的校准度。
Background
AI Health Tools Are Booming
AI健康工具蓬勃发展
AI-assisted health consultation apps are rapidly growing in popularity, bringing both significant opportunity and new challenges to product design.
AI辅助健康咨询应用快速普及，为产品设计带来显著机遇的同时也引入新挑战。
Problem
Two-Sided Trust Crisis
双向信任危机
Over-trust may lead users to follow advice blindly, posing potential health risks.
Under-trust may cause users to dismiss or ignore the system entirely, preventing these tools from delivering their intended value.
过度信任可能导致用户盲从建议，带来健康风险；信任不足则可能让用户完全忽视系统，使工具无法发挥价值。
Research Question
Calibrated Trust
校准信任
How can user interface design strategies help users establish calibrated trust in AI-assisted health tools?
界面设计策略如何帮助用户在AI辅助健康工具中建立校准信任——在可靠时信任AI，在不确定或高风险时主动核查？
03 / Clinician Trust Research
7 interviews → 5 core trust factors
7次访谈 → 5个核心信任因素
Semester 1 combined an initial literature review on clinician trust in AI, decision-making, and risk evaluation with 7 semi-structured interviews with clinicians and healthcare UX designers. The interviews surfaced five recurring trust factors that became the foundation carried into Semester 2.
第一学期结合文献综述（临床AI信任、决策制定与风险评估）与7次半结构化访谈（受访者含临床医生及医疗UX设计师），梳理出五个反复出现的信任影响因素，构成延续至第二学期的研究基础。
01
Workflow Fit
流程融合
AI must integrate into existing clinical processes without disruption
AI须无缝嵌入现有临床工作流程，不造成干扰
02
User Control
用户控制权
Clinicians need the ability to override or reject AI output
临床医生需要能够覆盖或拒绝AI输出的能力
03
Clinical Reasoning
临床推理对齐
AI outputs must align with how clinicians think and diagnose
AI输出须与临床医生的思维方式和诊断逻辑保持一致
04
Reliability + Transparency
可靠性 + 透明度
Consistent, explainable, and predictable performance
稳定、可解释且可预测的性能表现
05
Ethical Data Use
数据伦理
Patient safety, privacy, and clear accountability
患者安全、隐私保护与明确的责任归属
Key takeaway: Trust depends on control, reasoning, and accountability. These factors became the research foundation bridging clinician research (Semester 1) to consumer product design (Semester 2) — confirmed by a literature cross-comparison showing the same core drivers hold across both populations.
核心结论：信任的根基在于控制权、推理逻辑与责任归属。文献交叉对比同样证实，这些驱动因素在临床医生与普通消费者两类群体中具有普遍性。
04 / MaxDiff Survey
Quantifying what drives trust
量化信任驱动因素
To translate shared trust factors into product priorities, we ran a MaxDiff (Best-Worst Scaling) survey on Qualtrics. Forced-choice methodology removes rating-scale bias — participants select the MOST and LEAST important feature each round, producing relative importance scores across all 12 items.
为将共同信任因素转化为产品优先级，我们在Qualtrics上进行了MaxDiff（最优最劣量表）调研。强迫选择方法消除评分量表偏差——参与者每轮选出最重要与最不重要的功能，生成所有12项功能的相对重要性得分。
Top Ranked
最高排名功能
- Error correction control错误纠正控制
- Clear AI explanationAI解释清晰度
- Privacy label transparency隐私标签透明度
- Institutional trust signals机构信任信号
- Data control rights数据控制权
Bottom Ranked
最低排名功能
- Confidence score置信分数
- Non-directive AI guidance非指令性AI引导
- AI frequency controlAI推送频率控制
User Differences
用户群体差异
- Lower AI familiarity → privacy + trust signalsAI熟悉度低 → 优先隐私+信任信号
- Higher AI familiarity → explanation + source transparencyAI熟悉度高 → 优先解释+来源透明度
Users prioritize the same factors clinicians did —
control, transparency, and accountability — reaffirming these as universal trust drivers across both clinical and consumer populations.
用户与临床医生优先考量同样的因素——控制权、透明度与责任归属——印证了这些因素作为跨群体普遍信任驱动力的核心地位。
05 / Design Requirements
From insights to design criteria
从洞察到设计标准
Each design requirement maps directly from a user insight surfaced through clinician interviews, comparative analysis, and survey results. Four requirements, four interventions.
每项设计需求均直接来源于临床访谈、竞品分析和调研结果所揭示的用户洞察。四项需求，四个干预方向。
More detailed answers, but not at the cost of cognitive overload
需要详细答案，但不能造成认知负担
→
DR-01
Provide summary and collapsed reasoning
提供摘要与折叠推理
Anxiety when they have to always use AI to ask health questions
总是需要用AI询问健康问题会产生焦虑
→
DR-02
Explicit AI ON / OFF state indicator
明确的AI开/关状态指示
Concern about persistent data storage, use, and privacy
担心持续数据存储、使用方式和隐私问题
→
DR-03
Incognito private mode + self data management
无痕私密模式 + 自助数据管理
AI is very likely to make mistakes
AI很可能会出错
→
DR-04
Review-and-correct button + trusted resource
审核与修正按钮 + 可信资源链接
06 / Prototyping Iteration
Three rounds, from 64 prompts to vibe code
三轮迭代，从64条Figma AI提示词到代码
Prototyping moved through three distinct modes — each building on the previous, tightening the alignment between research insights and implemented features. Figma AI for ideation; Claude Code for implementation.
原型设计经历三种不同模式——每一轮在前一轮基础上深化，持续收紧研究洞察与功能实现之间的对齐度。Figma AI用于原型制作，Claude Code用于工程实现。
01
Prompt Prototyping
提示词原型
Defined the product via structured prompts to Figma AI, establishing scope before building.
通过结构化提示词驱动Figma AI定义产品形态，在动手构建前先确立边界。
- User: general consumers, not clinicians目标用户：普通消费者，非临床医生
- Three screens: Home · Q&A · Report Upload三个核心页面：首页 · 问答 · 报告上传
- Informational only — no medical diagnosis仅提供信息参考，不做医学诊断
- Apple Health-inspired aesthetic参考Apple Health的视觉风格
02
Research-Aligned Iterative
研究驱动的迭代原型
Applied the five trust factors directly to the prototype across 67+ versions.
将五个信任因素直接映射到原型，历经67+个版本迭代。
- Transparency: data usage visible before AI answers透明度：AI回答前始终展示数据来源
- Privacy: incognito mode + repeated assurances隐私：无痕模式 + 多次明示隐私保护
- Control: data permissions + accept/reject output控制权：数据授权 + 接受/拒绝AI输出
- AI OFF state with inline warning bannerAI关闭状态 + 内联警告横幅
03
Vibe Coding
氛围编程
Moved from Figma prototype to working app using Claude Code with Figma MCP.
通过Claude Code接入Figma MCP，将Figma原型转化为可运行的应用程序。
- Generated initial app from Figma via MCP通过MCP连接从Figma生成初始应用
- Integrated Anthropic API for live health Q&A接入Anthropic API实现真实健康问答
- Deployed to public URL for usability testing部署至公开链接以便可用性测试
- Internal testing before participant sessions正式测试前完成内部验收
07 / Usability Testing
6 participants, 5 tasks, think-aloud
6名参与者，5项任务，出声思维
With the working app deployed, we ran remote moderated usability sessions to evaluate whether trust-informed features actually work in practice. Each session was ~45 minutes with a think-aloud protocol throughout.
应用部署完成后，进行远程调节式可用性测试，评估以信任为导向的功能是否真正奏效。每场会话约45分钟，全程采用出声思维协议。
Method
- n = 6 participants, age 24–476名参与者，年龄24–47岁
- Mixed AI familiarity levels不同AI熟悉度水平
- Think-aloud protocol throughout全程出声思维协议
- Remote moderated sessions (~45 min)远程调节式会话，约45分钟
Measures
- Task success rate任务完成率
- SUS (System Usability Scale)系统可用性量表
- Qualitative feedback on feature understanding功能理解度的定性反馈
01
Login & Initial Setup
登录与初始设置
Understand Data Consent & Privacy features
了解数据同意与隐私功能
02
Sleep Study Query
睡眠研究查询
Understand Transparency, Explainability, User Control features
了解透明度、可解释性与用户控制功能
03
AI ON / OFF Toggle
AI开/关切换
Understand User Control + evaluate Trust Calibration — can users understand the difference AI makes?
了解用户控制 + 评估信任校准——用户能否理解AI带来的差异？
04
Incognito Mode
无痕模式
Understand Privacy & User Control features
了解隐私与用户控制功能
05
Open-Ended Health Question
开放式健康问题
Test Transparency, Explanations, and User Control with real AI responses
在真实AI回复中测试透明度、解释性与用户控制的表现
08 / Findings & Next Steps
From findings to design opportunity
从发现到设计机会
Testing revealed a clear split: features grounded in data visibility and explainability performed well; features relying on abstract indicators or binary controls confused users and undermined the trust they were designed to build.
测试揭示了一个清晰的分界：基于数据可视性与可解释性的功能表现良好；依赖抽象指标或二元控制的功能让用户困惑，反而削弱了本应建立的信任。
What Worked Well
有效的设计
- "What data was used?" → supported Transparency"使用了哪些数据？"→ 支持了透明度
- Data visualizations → supported Explainability数据可视化 → 支持了可解释性
- Data-backed specific insights → supported Reliability有数据支撑的具体洞察 → 支持了可靠性
- Explanations & sources → supported Institutional trust解释与来源 → 支持了机构信任信号
What Did Not Work Well
失效的设计
- Confidence score → undermined Explainability置信分数 → 破坏了可解释性（被误读）
- AI ON/OFF → undermined User ControlAI开/关 → 破坏了用户控制（混乱，价值感知低）
- Generic answers → undermined Reliability通用答案 → 破坏了可靠性（与Google无差异）
- Binary consent → undermined Privacy intent二元同意 → 破坏了隐私设计意图
Top Issues to Fix · 待修复的核心问题
Redesign AI confidence indicator
重新设计AI置信指标
Rename Edit / Accept / Reject actions
重命名编辑/接受/拒绝操作
Rewrite AI OFF warning message
重写AI关闭警告文案
Match data visualization to question topic
数据可视化与问题主题匹配
Add conversational follow-up input
添加对话式追问输入
Granular data consent
精细化数据同意设计
Iteration Plan · 迭代计划
Immediate
立即执行
- Redesign confidence score重新设计置信分数
- Rename Edit → "Add Note"; add Follow-Up input将"编辑"改为"添加备注"；增加追问输入框
- Rewrite AI OFF warning to neutral language将AI关闭警告改为中性描述语言
- Make data visualizations match question topic使数据可视化与问题主题匹配
Near-Term
短期规划
- Add conversational follow-up (multi-turn chat)添加多轮对话式追问功能
- Context-aware personalization beyond raw data超越原始数据的情境感知个性化
- Proactive "Insights for You" on home screen首页添加主动推送的"专属洞察"板块
Long-Term
长期方向
- Evaluate over-trust vs under-trust in extended use评估长期使用中的过度信任与信任不足问题
- "Health Log" organized by topic for doctor visits按主题分类的"健康日志"，服务就诊场景
09 / Live Product
Trust conditions, shipped as an app
将信任条件落地为C端应用
We extracted the five clinician trust factors, validated them with 59 consumer survey responses, and vibe-coded a working health Q&A app that embeds each trust intervention directly into the UI. Built with Claude Code + Anthropic API.
将五个临床信任因素与59份消费者调研结果结合，通过Vibe Coding将每项信任干预措施直接内嵌到界面中，最终交付一个可运行的健康问答应用。使用Claude Code + Anthropic API构建。
Note
The clinical research was conducted in an enterprise context with Philips. To present the work openly, the core trust factors and research findings have been translated into a consumer-facing product — allowing the design decisions to be explored directly in the browser without exposing proprietary data.
原始研究在飞利浦企业合作背景下进行。为方便作品集展示，我们将核心信任因素与研究结论迁移至一个C端产品形态，访问者可直接在浏览器中体验设计决策，无需接触任何企业敏感数据。
App not loading? Open in new tab → · 加载失败请在新标签页打开