Foreign Systems: Track developments like DeepMind Health (UK NHS) for lessons on transparency and data access (e.g., blockchain for auditability).

Challenges in Creating Comprehensive Datasets

Data Quality & Access: Critical barriers to AI implementation in health.
Defining Datasets: Major challenge for AI aiming to find novel disease correlations and personalize treatments.

Concerns regarding Electronic Health Records (EHRs)

Caution: Use EHRs as training sets with extreme care.
Issues: Incomplete, lack interoperability, variable quality (not collected under research controls).
Risk: Incorrect or correlated data can lead to misleading AI outputs.
Study Example (CVD Prediction): ML algorithms offered only marginal improvements over standard tools with UK NHS EHR data.
Transparency: 'Black-box' nature of ML means weighted risk factors can change idiosyncratically; need for transparent reporting.

The Missing Environmental Data Stream

Impact: Environmental exposures and social behaviors account for $\approx 60\%$ of premature deaths.
Data Imbalance: Lack of diverse data capture for precision medicine, especially environmental toxicology and exposure.
Recommendations for Data Collection:
- Integrate toxin screening (dioxin, lead) into routine blood panels.
- Add diet and environmental toxin questions to health questionnaires.
- Start urban sensing and tracking programs aligned with large health studies (e.g., All of Us).
- Support wearable device development for sensing environmental toxins and pathogens.
- Develop protocols and IT capabilities to collect and integrate diverse data streams.

Advancing AI Development and Ensuring Success

Crowdsourcing and Data Generation

Value: Provides high-quality, labeled data and accelerates AI algorithm development (e.g., Kaggle competitions).
Examples: Improving lung cancer screening, classifying genetic mutations.
Recommendations:
- Support competitions to advance health data understanding.
- Share data in public forums to engage scientists.
Main Challenge: Need for large, well-labeled public/semi-public datasets.

Limitations, Transparency, and Safeguards

Guarding Against Misinformation ("Snake Oil"):
- Potential for misleading websites, apps, and companies (e.g., MTHFR treatments, non-validated diagnostic apps).
- Engagement of learned bodies recommended to endorse best practices for AI deployment.
Ensuring Reproducibility and Transparency:
- Deep learning models are complex and susceptible to adversarial examples; 'black-box' nature.
- Methods for transparent disclosure of large-scale computational models are essential but nascent.
Recommendations:
- Support development/adoption of transparent processes and policies for reproducibility (data, software, workflows, environment details).
- Involve researchers, technologists, health professionals, industry, regulators, and patients in discussions.

Summary of Findings and Recommendations

Category	Key Findings	Key Recommendations