AI and Machine Learning in Cancer Medicine

• The Melanoma Research Alliance (MRA), founded in 2007 by Deborah and Leon Black under the Milken Institute, has invested $110 million in melanoma research and leveraged an additional $140–$150 million over a decade.
• MRA has funded over 300 projects across the U.S. and 15 countries, with AI and machine learning now a key focus for improving early detection and staging.
• Melanoma incidence is rising globally, causing more than 10,000 deaths annually, with survival rates dropping significantly as tumor depth increases from 1 mm to 4 mm.

Dermatology and AI Imaging (Dr. Susan Sweater)

• Stanford researchers developed a deep learning algorithm trained on 139,000 curated global images to classify melanoma versus benign nevi and keratinocyte cancers.
• In testing, the algorithm performed as well as or better than 21 board-certified dermatologists in classifying static images.
• Current limitations include the need for prospective clinical validation outside of "in silico" environments and the necessity to break the "black box" of AI decision-making.
• Significant risks regarding algorithmic bias exist, particularly across different skin types, with current models potentially misclassifying lesions based on background markers or rulers.
• The FTC has fined developers for false claims that consumer apps could diagnose melanoma, with some algorithms underdiagnosing by 30%.
• Future goals include using AI for longitudinal change detection via serial photography and providing automated second opinions for primary care providers to reduce referral burdens in rural areas.
• Dr. Sweater anticipates AI could reduce the need for specialist referrals if high sensitivity and specificity are validated, though the technology is not yet ready for unassisted consumer diagnostic use.

Causal Modeling and Precision Medicine (Colin Hill, GNS Healthcare)

• GNS Healthcare utilizes causal modeling and simulation to "reverse-engineer" cancer patients in silico, specifically for predicting responses to stem cell transplants in multiple myeloma.
• A model developed in roughly three months identified a gene expression threshold predictive of 20-month progression-free survival, later validated by Dana-Farber.
• The approach moves beyond pattern recognition (deep learning) to answer counterfactual "what-if" questions regarding treatment effectiveness and personalized intervention.
• Dr. Hill predicts that within 3–5 years, AI will drive a "renaissance of precision healthcare," moving away from the term "personalized medicine" toward standard practice.
• He forecasts that the first AI-driven drug and truly personalized care pathways will emerge within 5–10 years, potentially transforming reimbursement models.
• Real-world evidence is currently being used to drive care recommendations for hundreds of thousands of patients, and Dr. Hill anticipates these packages will complement or begin to substitute for randomized controlled trials within 18–24 months.

Device Engineering and Imaging (Dr. Maurice Ferre, Insight Tech)

• Insight Tech applies AI and big data to model and steer acoustic ultrasound beams through the skull with sub-millimeter accuracy for focused ultrasound brain therapies.
• Every skull requires unique modeling due to physical deformations and scattering; AI enables the system to adapt to these individual anatomical variations.
• The technology integrates with MRI machines to perform tractography, modeling brain tracks to guide non-invasive lesioning treatments for movement disorders like Parkinson's and essential tremors.

Pharmaceutical Industry Perspective (Mike Nohaley, Amgen)

• Amgen distinguishes between "machine vision" (rapid classification) and "causal modeling" (predictive counterfactuals), viewing the latter as more transformative for long-term drug development.
• Dr. Nohaley emphasizes that current AI applications require specific, high-quality data conditions and cannot yet solve "totalizing" problems like a "magic button" for drug discovery.
• The company is focusing on integrating electronic medical records, omics data, and longitudinal wearable data to create rich datasets for predictive modeling.
• Amgen warns against the "push-button" hype, noting that while AI can aid researchers in literature review and data integration, human domain expertise remains essential.

Market Trends, Challenges, and Future Trajectory

• Data Gaps: Current healthcare data is often "dirty" (EMR) or siloed; significant investment is required in data curation and annotation to achieve clinical-grade inputs.
• Workforce Disparities: Rural areas face higher cancer mortality due to a lack of specialists (1 oncologist per 100,000 people vs. 5 per 100,000 in urban areas), a gap AI telemedicine aims to bridge.
• Tech Industry Entry: Panelists agree big tech companies (Google, Amazon, etc.) will likely partner with healthcare providers rather than disrupt the sector alone due to heavy FDA regulation and liability.
• Implementation Barriers: Success requires not just technology but "change management," including shifting fee-for-service payment models, ensuring interoperability, and addressing social determinants of health.
• Regulatory Landscape: The FDA is increasingly open to real-world evidence for approvals, though the pace of validation remains slow and rigorous.
• Ethical and Safety Concerns: AI lacks empathy and cannot handle medical "I don't know" scenarios; unassisted use in primary care could lead to over-diagnosis or workflow fatigue if not carefully managed.
• Future Vision: The field is moving from hypothesis-driven to data-driven research, with the ultimate goal of simulating millions of clinical trials in silico to reduce costs and improve patient matching.