Introduction of Fei-Fei Li:
Fei-Fei Li, a renowned AI researcher and professor at Stanford University, is the keynote speaker. She has made significant contributions to AI, particularly in computer vision, through projects like ImageNet, Visual Genome, and ActivityNet. Li is also a strong advocate for diversity in AI through her co-founded program, AI for All.

Fei-Fei Li’s Connection to Michigan:
Li spent a sabbatical at Michigan in 2013 and fondly remembers her time there. She expresses gratitude for being invited back to Michigan, albeit virtually, to share her work.

Collaboration in AI and Healthcare:
Li emphasizes the collaborative nature of her research, involving computer scientists, AI researchers, clinicians, and clinical researchers from Stanford Medicine and partner hospitals. Professor Arne Milstein is highlighted as a key collaborator, having introduced Li to healthcare research. PhD student Albert Hawk is acknowledged for his significant contributions to the work presented.

Main Talk: AI and Healthcare
Li expresses shared passion for healthcare and highlights the remarkable advancements in modern medicine. She emphasizes the critical role of AI in healthcare, particularly in understanding diseases, developing treatments, and improving patient care.

Introduction:
Fei-Fei Li highlights the importance of healthcare advances and the need to address medical errors, which are still the third leading cause of death. She emphasizes the complexity of healthcare as a human behavioral system and the challenges in preventing unintended errors.

Autonomous Driving as Inspiration:
Li draws a parallel between autonomous driving and healthcare, both involving complex human behavior. She discusses the emergence of enabling technologies in autonomous driving and their potential to revolutionize healthcare.

Envisioning Ambient Intelligence in Healthcare:
Li proposes endowing healthcare spaces with ambient intelligence, similar to smart cars, using sensors, machine learning, and computer vision. This transformation aims to make healthcare spaces visible and visualizable, improving patient care outcomes.

Research Journey and Nature Review Article:
Li mentions a comprehensive Nature Review article summarizing the learnings from their research journey. She focuses on two critical healthcare spaces: hospital spaces and daily living spaces.

Hospital Space Opportunities:
Li explores potential applications of ambient intelligence in various hospital settings, including ICUs, inpatient units, and operating rooms. She presents an example project in the ICU, aiming to improve patient mobility through sensor-based monitoring.

Challenges in Patient Mobility Monitoring:
Li discusses the challenges of ensuring proper patient mobility in ICUs, emphasizing the limitations of direct observation and charting. She highlights the issues of charting errors, labor intensity, and the lack of nuance and accuracy.

Sensor-Based Monitoring in Patient Rooms:
Li describes a collaboration with two hospitals to install sensors in patient rooms for monitoring patient mobility. She explains how the sensors capture data on patient movement, enabling more accurate and objective monitoring.

AI-Based Mobility Monitoring in ICU:
Depth sensors capture data without compromising privacy. Machine learning algorithms label human behaviors, such as patient movement activities. Action detection technology provides detailed data insights over time. Continuous objective measure of patient mobility aids clinicians.

Hand Hygiene Compliance Monitoring:
Hand hygiene compliance is crucial in preventing hospital-acquired infections. Traditional methods rely on human observers or secret shoppers. AI-based systems can provide continuous, accurate monitoring. Real-time feedback and alerts improve compliance and patient safety.

Challenges of Traditional Hand Hygiene Monitoring:
Traditional methods of hand hygiene monitoring are sparse, time-consuming, expensive, and prone to human bias. RFID technology is coarse, does not accurately capture human behavior, and disrupts workflow.

Sensor-Based Hand Hygiene Monitoring System:
Implemented in Intermountain Hospital in Utah and Stanford’s Children’s Hospital. Utilizes depth sensors to collect data on clinician behavior. Employs convolutional neural network action detection algorithm to identify hand hygiene behavior temporally and spatially. Addresses technical challenges such as combining multi-sensor views and finding a common space to project people in the video.

Benefits of Sensor-Based Monitoring:
Provides invaluable insights for optimizing the location of hand hygiene dispensers. Enables the study of hand hygiene compliance beyond entering rooms, including the five moments of interaction with patients as outlined by the WHO.

Research on Surgical Staff Activity:
Ambient sensors are being used to study human 3D pose in surgical environments for training and tracking purposes. Tracking objects in the surgical room is important to prevent errors, as many objects cannot be tracked by RFID or barcodes. Nurses spend a significant amount of time counting objects to ensure they are not lost in the patient’s body.

AI in Healthcare:
AI is revolutionizing healthcare by automating processes, tracking surgical instruments, and using smart sensors to understand patient behavior.

Surgical Instrument Tracking:
Laparoscopic videos are analyzed to track surgical instrument usage, providing valuable information for training and understanding surgical procedures.

Smart Sensors for Senior Care:
Smart sensors and ambient intelligence are being piloted in senior homes to monitor activities of daily living (ADL) and detect falls.

Benefits of Smart Sensors:
Smart sensors can provide real-time detection of falls, which is crucial for senior care. They can also collect data on sleep patterns and other ADLs, providing valuable information for clinicians and caretakers.

Challenges of Wearables:
Wearables are often inconvenient for seniors to wear consistently. They are limited in capturing nuanced human activities.

Complementary Information from Sensors:
Thermal sensors and step sensors provide complementary information for understanding ADLs. This data can be analyzed to detect falls and other important events.

Data Analytics for Clinically Relevant Data:
Smart sensor data can be analyzed to provide clinically relevant data for caretakers and clinicians. This data can help improve care and identify potential health issues.

Sleep Pattern Analysis:
Smart sensors can provide detailed information about sleep patterns, including different types of sleep and daily sleep data.

Senior Care:
Fei-Fei Li mentions an ongoing project focused on reducing 30-day readmission rates among seniors during the COVID-19 pandemic.

Physical Rehabilitation:
Physical rehabilitation is crucial for many individuals, especially seniors. Traditional methods involve visiting labs or clinics, which can be inconvenient and costly. Ambient intelligence, particularly camera systems, offers the potential for at-home physical rehabilitation. Researchers are exploring the use of depth sensors to understand human gestures, gait, and posture for rehabilitation purposes.

Mental Health:
Ambient intelligence has applications in mental health as well, but Fei-Fei Li does not delve into this topic due to time constraints.

Societal and Ethical Implications:
Stanford University emphasizes human-centered AI research, which includes considering ethical implications during the design phase. Collaborations with bioethicists, legal scholars, philosophers, and ethicists help ensure that technology aligns with ethical principles.

Privacy Concerns:
Technical Approaches: Face blurring, video downsampling, and human body masking can be used to protect privacy. Differential privacy and federated learning can ensure data privacy. Homomorphic encryption can provide stronger privacy guarantees. Physical privacy of devices must also be considered. Regulatory Considerations: Psychotherapist-patient privilege and the potential for ambient sensors to become witnesses pose legal challenges. Legal scholars are collaborating to forecast and address these issues.

Bias and Fairness:
Data sets can exhibit bias based on gender, skin color, and other demographics. Inclusion of different demographics is important to avoid bias. Algorithmic transparency and accountability are essential to address bias. Model cards can be used to ensure algorithmic transparency.

Ethical Issues in Scientific Research:
Ethical considerations must be addressed before mass application of technology. Meticulous care is taken in the scientific research protocol, including IRB approval, patient consent, data storage, and annotation.

Multi-Stakeholder Approach:
Ethical design should be incorporated into technology before development begins. Collaboration with policy leaders, ethicists, legal experts, clinicians, patients, and technologists is crucial.

Multimodality in Data Collection:
In addition to visual sensors, wearables, audio, depth, and thermal sensors can provide valuable multimodal data. The choice of sensors depends on the specific clinical context.

Privacy and Accuracy Trade-Off:
Depth sensors offer greater privacy protection compared to RGB cameras. RGB cameras provide valuable clinical information, such as nuanced movement patterns and facial expressions of pain. A balanced approach is needed to ensure privacy while maximizing clinical benefits.

Ethical Considerations in Data Usage:
Balancing privacy concerns with the ethical obligation to help those in need is a complex challenge. Collaboration between researchers, clinicians, and policymakers is essential to address this trade-off.

Key Insights:
Fei-Fei Li emphasizes the importance of a multi-stakeholder approach to address ethical concerns in AI-assisted healthcare. This involves involving clinicians, patients, nurses, legal scholars, and technologists in the decision-making process. Li highlights the need for humility and respect when working with vulnerable populations, such as elderly care patients. She encourages AI students to shadow doctors and gain empathy for the challenges faced by these individuals. Li believes that balancing technology and human dignity requires careful consideration of necessary policy guardrails and leveraging technology’s intelligence to solve ethical issues.

Concerns about Autonomy and Independence:
Rada Mihalcea raises concerns about data collection potentially infringing on the autonomy and independence of elderly individuals receiving care. She questions whether such systems might discourage people from striving for autonomy. Li addresses these concerns by emphasizing the importance of informed consent and working with individuals who understand the purpose and implications of data collection. She also notes that ambient intelligence technology is primarily an observer and does not enact actions on behalf of individuals, thus preserving a sense of autonomy.

Augmenting Humanity with AI:
Li believes that AI has the potential to augment and enhance human capabilities rather than replace them. She emphasizes the value of using AI to improve safety and security, while also recognizing its potential to enhance healthcare and other domains.

Transfer Learning between Populations:
Mihalcea inquires about the potential for reusing AI algorithms across different populations, such as elderly care and infant healthcare. Li acknowledges the potential for transfer learning but cautions that careful consideration must be given to the specific needs and characteristics of each population to ensure effective and ethical applications of AI.

Public Trust:
Public trust is lacking in the use of AI in healthcare, due to concerns about data privacy and security. There is a need to address these concerns and build public trust before AI can be widely adopted in healthcare.

Ethical Considerations:
Ethical considerations are also a challenge, such as the need to ensure that AI systems are fair, unbiased, and used responsibly. There is a need for guidelines and regulations to ensure the ethical use of AI in healthcare.

Regulatory Framework:
The lack of a clear regulatory framework for AI in healthcare is also a barrier to adoption. There is a need for clear regulations and standards to govern the use of AI in healthcare and ensure the safety and efficacy of AI-based technologies.

Lack of Skilled Workforce:
The lack of a skilled workforce is also a challenge, as there is a shortage of healthcare professionals with the necessary skills and knowledge to use and interpret AI technologies effectively. There is a need for training and education programs to develop the necessary skills and knowledge among healthcare professionals.

Long-Tail Learning and Rare Events:
Healthcare often involves rare events, such as senior falls or certain diseases, which makes it difficult to collect enough data for traditional deep learning algorithms. Transfer learning, domain adaptation, and few-shot learning are suitable methods for handling long-tail learning and rare events in healthcare.

Technical Challenges:
Technical challenges also exist, such as the need for robust and interpretable AI algorithms that can handle the complexity and variability of healthcare data. There is a need for research and development to address these technical challenges and improve the performance and reliability of AI systems in healthcare.

Public Image of AI:
Fei-Fei Li is concerned about AI’s public image, particularly the low trust in the technology due to issues like bias and privacy. Li believes that gaining public trust is essential for the healthcare field.

Modernizing Policy and Regulations:
Li emphasizes the need to modernize policy and regulatory measures surrounding data and machine learning in healthcare. She acknowledges the importance of data protection but stresses the need to incentivize innovation. Li believes that modernizing regulations can help build public trust and encourage innovation within safe boundaries.

Conclusion:
Li identifies two key issues that need to be addressed: improving AI’s public image and modernizing policy and regulations. She highlights the importance of gaining public trust through responsible research and the alignment of technological aspirations with the goal of using AI for good.