The hidden tax of multimodal AI: lung cancer study questions cross-validation gains
bravo_abad · x · 2026-09-14
Multimodal AI sounds automatically better in medicine, but every added modality costs: fewer patients have complete data and measurements are harder to keep consistent across sites.
The I³LUNG study makes the trade-off visible:
- Of 2,396 lung cancer patients, only 339 had all four modalities.
- In cross-validation, adding pathology and imaging raised AUC from 0.68 to 0.88 on a 24-month survival task.
- But those gains failed to consistently survive independent test and external validation, even for a fusion model designed to handle missing modalities.
Takeaway: multimodal AI carries a hidden tax of data completeness and poor external generalization; cross-validation results can be misleading.
More from AGI Musings
- sull Warns Against Restricting Open-Weight Models: Responsibility Isn't Global Advantage — sull · 2026-09-14
- Ex-OpenAI safety researcher: slowing AI is about human control, not self-improvement — sull · 2026-09-14
- VC to founder: the small fish turn into big fish in 3 months these days — Danhightower · 2026-09-14
- DeepSeek reportedly chases true continuous learning while OpenAI bets on agent swarms — teortaxesTex · 2026-09-14
- Grant Hawkins: holding AI firms financially liable beats adding external regulators — granawkins · 2026-09-14
- Researcher argues catching up to frontier AI costs less than proposed compute budgets — eliebakouch · 2026-09-14