Stanford CS547 HCI Seminar | Spring 2026 | HCI and Human-Centered AI for Digital Health

Key Concepts

• Human-Centered AI (HCAI) in Digital Health: Integrating AI models into clinical workflows while prioritizing user burden, trust, and behavioral science.
• Personalized Machine Learning: Training individual AI models per user rather than a single generalized model, allowing for better adaptation to unique physiological baselines.
• Personalized Self-Supervised Learning: Using unlabeled wearable data to create a "foundation model" for an individual, which is then fine-tuned for specific health outcomes (e.g., stress, craving).
• Just-in-Time Adaptive Interventions (JITAI): Delivering digital health interventions at the precise moment they are most effective and least burdensome.
• Algorithmic Fairness & Bias: The challenge of defining and mitigating bias (e.g., disparate impact, equalized odds) in AI models.
• Clinical Actionability: Prioritizing "event spikes" (e.g., blood pressure spikes) over precise numerical values to ensure interventions are meaningful to clinicians.

---

1. AI Methodology and Frameworks

The speaker emphasizes a shift from traditional diagnostic AI to predictive, intervenable health events.

• Personalization Strategy: Unlike traditional models that use one global dataset, the lab trains models per person. This addresses the high variability in human physiology.
• Self-Supervised Learning: By treating wearable signals like language models treat text (predicting the "next word" or "missing signal segment"), the lab creates personalized foundation models. This bypasses the "labeling bottleneck," as raw sensor data is abundant.
• Active Learning: To reduce user burden, the model identifies "uncertain" predictions (e.g., 56% probability of stress) and only requests user input/calibration at those critical moments.

2. HCI Challenges in Digital Health

The speaker identifies three core HCI challenges that arise when AI meets real-world clinical practice:

• Challenge 1: Reducing User Burden: Balancing the need for data with the reality of patient life (e.g., a parent picking up a child from school cannot engage with an app). The lab uses causal diagrams to map how intervention timing and content affect user receptivity and adherence.
• Challenge 2: Evaluation Metrics: Standard metrics (Precision, Recall, F1) can be misleading.
  • Precision is sensitive to population prevalence, making it unreliable when moving from screening to specialty clinics.
  • Sensitivity/Specificity are preferred by clinicians because they are independent of prevalence.
• Challenge 3: Workflow Fit: AI models often fail because they do not fit into existing clinical tools (e.g., Epic EHR). The speaker notes that clinicians often ignore alerts due to "alert fatigue" caused by low-precision models.

3. Real-World Applications and Case Studies

• Substance Use & Vaping: Collaborating with behavior scientists to deliver interventions for young adults vaping nicotine or cannabis at the right time to maximize efficacy.
• Blood Pressure & Stress: Predicting "spikes" rather than exact numbers, as spikes are the clinically actionable events.
• Parkinson’s Assessment: A web-based tool using mouse/keyboard tasks.
  • Key Finding: The model performed better on MacBooks than Windows devices, highlighting how hardware heterogeneity and social determinants of health impact AI performance.
  • Clinical Lesson: The team initially failed to account for medication timing (Levodopa) and the asymmetric onset of Parkinson’s symptoms, proving that AI development must involve clinicians from the design phase.

4. Key Arguments and Perspectives

• The "Alert Fatigue" Problem: The speaker highlights the Epic sepsis model as a cautionary tale. Despite high sensitivity/specificity, the low precision led to constant false alarms, causing clinicians to lose trust and ignore the system entirely.
• Specificity vs. Sensitivity: In many clinical contexts (like the Apple Watch hypertension feature), high specificity is preferred to avoid overwhelming users with false positives, even if it means missing some true cases.
• Iterative Design: The speaker argues that initial studies should be viewed as "low-fi prototypes" that provide preliminary data for future, more robust clinical trials.

5. Notable Quotes

• "You can have a very efficacious intervention, but if it's delivered at a time where no one wants to use it or can use it, then you have no efficacy."
• "Just AI innovation by itself is not going to solve these fundamental issues with human behavior."
• "Clinicians are really burnt out... a big design challenge is how do we increase the information presented to clinicians while not making them stressed out."

6. Synthesis and Conclusion

The primary takeaway is that AI performance is not just a technical metric; it is a socio-technical outcome. The speaker advocates for a "human-in-the-loop" approach where AI models are personalized, evaluated using clinically relevant metrics (like specificity for alert management), and designed with deep stakeholder engagement. The future of digital health lies not in more complex algorithms, but in better integration with clinical workflows and a nuanced understanding of the human context in which these models operate.