Introduction:
Stanford hosted a discussion on the future of healthcare and artificial intelligence (AI) with over 18,000 attendees from 140 countries. Speakers included Fei-Fei Li, Sequoia Professor in Computer Science and Denning Co-Director of the Stanford Institute for Human-Centered Artificial Intelligence, and Andrew Ng, founder of deeplearning.ai and adjunct professor at Stanford. The event was co-hosted by Stanford High, deeplearning.ai, and supported by the Amy Center at Stanford.

Questions for the Speakers:
Kurt Langlotz began by asking each speaker why healthcare was an important area of focus for them and their research. Fei-Fei Li, a computer scientist, explained her motivations for focusing on healthcare despite having many other options.

Motivations for Entering Healthcare AI:
Fei-Fei Li: Personal experience caring for her chronically ill mother inspired her to pursue AI in healthcare. Believes in the importance of healthcare for individuals, families, and humanity. Sees AI as a tool to improve healthcare and patient outcomes. Andrew Ng: Inspired by his father, a doctor, and his stories about patients. Wanted to apply AI beyond consumer internet, recognizing healthcare’s social importance and AI readiness. Fortunate to collaborate with experienced partners like Kurt Langlotz and Matt Lundgren.

Challenges in Implementing AI in Clinical Use:
Proof of Concept to Production Gap: Research papers often show impressive AI performance, but translating that to clinical use is challenging. Data drift and variations in hospital environments can significantly degrade AI performance. Need for a Full Cycle Approach: Machine learning projects require more than just modeling. Data collection, deployment, monitoring, auditing, and safety are crucial for successful implementation. Common Challenges Across Industries: The gap between proof of concept and production is not unique to healthcare. However, healthcare has unique challenges due to its complexity and sensitivity.

Similarities and Differences between Healthcare and Other Domains:
Andrew Ng believes that the technical challenges in healthcare are more similar than distinct compared to other domains. The business and go-to-market aspects of healthcare are more complicated due to the need to navigate the payer and provider ecosystem.

Holdout Test Set vs. Practical System Deployment:
Machine learning practitioners often excel at performing well on holdout test sets, but deploying a practical system in healthcare requires more than just good performance on a test set. Additional factors, such as safety, privacy, and human factors, need to be considered for successful system deployment.

Ambient Intelligence in Healthcare:
Fei-Fei Li introduces the concept of ambient intelligence in healthcare, which focuses on the human behavior space that is not digitized but is crucial to patient care. Ambient intelligence involves combining smart sensors and back-end AI algorithms to collect data about human behavior in healthcare settings. This data can be used to improve patient safety, prevent medical errors, and provide better care.

Challenges in Ambient Intelligence for Healthcare:
Fei-Fei Li highlights the challenges in obtaining data for ambient intelligence in healthcare due to privacy concerns, data accuracy, and the sparse nature of charting by nurses. These challenges require innovative approaches to data collection and analysis to ensure the success of ambient intelligence in healthcare.

Ethical Considerations:
Healthcare AI systems should be designed with ethical principles in mind, such as privacy, fairness, integrity, and human dignity. Privacy measures include using more privacy-protected sensors, encryption, and secure data storage. Ethical considerations should be integrated into the technical system from the start, not as an afterthought. A team of bioethicists, ethical philosophers, and legal scholars can help guide the ethical design of AI systems in healthcare.

Collaborative Approaches:
Collaboration between researchers, clinicians, patients, and other stakeholders is crucial for developing AI systems that meet the needs of healthcare. Humility is an important trait for successful collaboration, as it allows researchers to listen to and learn from stakeholders with different perspectives. Multi-stakeholder collaboration can help align objectives and ensure that AI systems are developed with the needs of patients and healthcare providers in mind. Building relationships with experts in different fields can help researchers access knowledge and resources that they may not have on their own. Slack and other communication tools can facilitate collaboration and information sharing among team members.

Building Trusted Relationships:
Successful AI projects often involve interdisciplinary teams with expertise in different domains. Establishing trust and effective communication among team members is essential. Project selection should involve brainstorming and evaluating technical feasibility, value to patients and doctors, and project alignment with healthcare goals.

Empathy and Understanding:
Empathy is crucial for effective collaboration between computer scientists and healthcare professionals. Both sides should make an effort to understand each other’s perspectives, languages, and problem statements. Shadowing doctors and observing patient interactions can help computer scientists develop empathy for the healthcare context.

Palliative Care Recommendations:
AI systems can estimate the risk of death for patients based on electronic health records. Implementing such systems requires careful consideration of how to communicate risk to patients and how to integrate AI recommendations into healthcare workflows. Injecting data into healthcare systems can democratize decision-making and create space for more stakeholders to contribute.

Addressing Racial Biases:
Standards for bias and fairness in AI are still unclear. Ethical considerations should guide project selection and development. Collaborating with experts in social sciences and ethics can help mitigate biases and promote fairness in AI healthcare applications.

Economic Case:
The economic case for ethical AI is strong, but it should not be the sole factor in determining whether a project should be pursued.

Bias Standards:
The standards for bias in AI are still ambiguous, leading to companies being hesitant to launch products due to fear of negative publicity.

Proactive Brainstorming:
Teams should proactively brainstorm all potential types of bias that could occur in their AI systems. This should include racial, gender, and other protected characteristics.

Metrics and Auditing:
Design metrics to measure and track bias across different slices of data. Regularly audit AI systems against these standards to ensure they are not exhibiting unacceptable bias.

Human Responsibility:
Recognize that AI bias is a result of human actions, not independent machine values. The responsibility for ethical AI lies with the people involved in its design, development, and deployment.

Bias in the AI Pipeline:
Bias can occur at any step of the AI pipeline, from data curation to decision-making. It is important to mitigate bias at all stages, not just in the final outcome.

Ethics in Design:
Incorporate ethical considerations into the design of AI systems from the very beginning. Involve stakeholders, ethicists, and legal experts in the design process.

Data Bias:
Data scientists should not view bias as solely a problem of the data. Bias can often be addressed before data collection by carefully defining the problem and choosing appropriate data.

Key Points:

Striking the Right Balance:
Transparency in AI algorithms can build trust, but it can be challenging when proprietary technology is involved. Companies need to disclose information necessary for healthcare providers and patients to make informed decisions.

Transparency as Information Accessibility:
Andrew Ng emphasizes providing information based on stakeholders’ needs to do their jobs effectively. Healthcare providers need transparency to assess AI systems’ safety and performance. Patients and doctors require information to make decisions in their best interests.

The Problem of Tangled Explanations:
Ng highlights the confusion surrounding explainability due to mixing different purposes and stakeholders. Explanations can serve various purposes, such as convincing doctors, auditing for bias, and guiding error analysis. Separating out the needs of different stakeholders helps design targeted information.

Example of Misunderstanding Explanations:
Ng shares an example from a palliative care system using a learning algorithm to predict patient mortality. Doctors initially questioned the system’s predictions, leading to the development of explanations. Surprisingly, doctors quickly accepted the explanations but rarely referred to them again. Ng realizes that doctors didn’t need explanations to understand high-risk patients; they needed them to trust the AI system initially.

Multi-Stakeholder Involvement:
Ng emphasizes the need for a multi-stakeholder approach to AI in healthcare. Explanations are crucial for aligning the complex activities of various stakeholders. Collective understanding and participation are essential for delivering excellent patient care.

Different Types of Interpretability:
Fei-Fei Li discusses various types of interpretability and transparency beyond algorithm-level engineering. Causal interpretability focuses on establishing cause-and-effect relationships for trust-building. Direct trustworthy interpretability involves providing information directly relevant to decision-making.

Healthcare Challenges:
Medical errors contribute to 250,000 annual deaths in the US. Improving care delivery through AI can save lives, enhance family well-being, and reduce healthcare costs. Mental health and diagnostic capabilities are areas with significant opportunities for AI interventions. Operational aspects of healthcare, such as scheduling and resource allocation, can benefit from AI optimization.

AI Opportunities:
AI can assist clinicians in delivering intended medical procedures and timely diagnoses. AI can monitor chronic disease patients and aging seniors to prevent deterioration. AI can aid in mental health interventions and provide diagnostic capabilities, especially in regions with limited access to healthcare professionals. AI can improve operational efficiency in healthcare systems, leading to cost reduction and increased accessibility.

Data Access Barriers:
Despite HIPAA regulations, accessing healthcare data for research and development remains challenging. There is a need for modernized data curation and repositories to facilitate broader access to aggregated and de-identified healthcare data. Policymakers should prioritize efforts to modernize healthcare data access while protecting patient privacy. Stanford University has a progressive view of patient data, considering secondary use a public good and actively working to make it more available for research and development.

Data Access and Privacy:
* Balancing patient privacy with the need for accessible data.
* High penalties for privacy breaches, but patient support for responsible data use.
* Striking a balance between data-sharing for research and individual privacy.

Shifting from Model-Centric to Data-Centric AI:
* Historically, AI research has focused on refining models using fixed datasets.
* A new emphasis on improving data quality to enhance model performance.
* Need for MLOps tooling and processes to manage messy datasets effectively.

Challenges to AI Adoption in Healthcare:
* Lack of data interoperability and standardization.
* Difficulties in integrating AI systems with existing healthcare infrastructure.
* Concerns about bias and algorithmic transparency.

Future of AI in Healthcare:
* Prediction of widespread AI adoption in healthcare within 10 years.
* AI-driven improvements in disease diagnosis, treatment planning, and patient outcomes.
* Potential for AI to address healthcare disparities and make care more accessible.

Challenges:
* Managing the full life cycle of machine learning projects.
* Tackling the right problems.
* Technical and non-technical change management problems.

Progress:
* Slow progress in healthcare AI implementation, with isolated examples of success.
* Expectation of rapid growth in AI adoption in healthcare in the coming decade.

Human Factor:
* The need for AI to demonstrate a fundamental human difference in care, recovery, or well-being.
* The importance of proven stories and products that benefit patients and healthcare workers.

Commercialization:
* The path to commercialization and widespread adoption of AI in healthcare is still being defined.
* The industry is inching closer to a watershed moment where AI’s impact will be widely recognized.

COVID-19’s Impact:
* COVID-19 has accelerated the recognition of the importance of digital healthcare.
* Telemedicine and remote care have become more prevalent due to the pandemic.
* Aging populations have been disproportionately affected by healthcare system issues during the pandemic.