Interview
a16z Podcast | Putting AI in Medicine, in Practice
- AI is expected to gain traction primarily where financial incentives like fee-for-value models reward accuracy and reduced hospitalizations, with implementation difficulty varying significantly by area due to commercialization and scaling challenges.
- Future AI applications will complement human limitations by analyzing high-volume, noisy data from wearables to predict conditions like sleep apnea that exceed human interpretive capacity, while also acting as proactive tools to prompt patient consultations rather than serving only as reactive measures.
- Automation is anticipated to progress from assistive physician roles toward potential full autonomy without a doctor in the loop, a transition characterized as a societal decision rather than a technical hurdle, with novel outcomes like fusing image and blood data deemed valuable even if error patterns differ from human benchmarks.
- Risks include model overfitting when fuzzy or noisy data elements are included, causing validation datasets to perform worse than training sets, and the potential for bad inputs to bias systems learning on the fly, necessitating strict validation with unseen datasets for each model version.
- Addressing statistical mismatches will require gathering data from diverse global populations, with future clinical trials expected to recruit 40,000 to 50,000 participants via platforms like ResearchKit and HealthKit, though initial medical apps risk losing 90% of participants within the first 90 days due to poor mobile design.
- Successful AI teams must be highly interdisciplinary, combining clinical, AI, and mobile design expertise, while healthcare entities must accept risk to achieve cost reductions and improved outcomes in an environment where regulatory challenges are being addressed by the Office of Digital Health and the FDA.
- AI systems are expected to version similarly to speech recognition software, involving batch training, careful testing, and date-tagged deployments to manage changes and avoid bias, potentially allowing for the identification of physicians with high error rates through comparison with generalizable algorithms.
- Clinical adjacencies for treatment management will include optimizing OR scheduling by analyzing physician speed, case risk, and drowsiness via data sources like Fitbit, which may also be used to monitor staff stress levels to ensure availability during peak demand.
- Full-stack healthcare provider models, such as Omada Health or Virta Health, are predicted to see the fastest adoption in the payer sector, potentially leading to the industry reconstituting vertically around AI-based diagnostics or therapeutics with providers oriented around specific data networks rather than geography.