Healthcare AI Faces a Human-Test Reality Check

Healthcare AI companies continue to make bold claims, but most companies have not achieved a medicine that worked in humans. Isomorphic has brought no treatments to market, Lila the same, and Anthropic recently acquired stealth startup Coefficient Bio for $400 million.

The article argues that there is only one true test of any healthcare AI: Did it work in humans? To test a new treatment, you need to take it through a Phase 3 clinical trial. That is typically 10 years and $2 billion. To test a diagnostic, you need to demonstrate clinical benefit, pass a rigorous third-party test, and build a full quality management system before the product is even permitted into the clinic. To uncover and prove new human biology could take decades of scientific experimentation.

Companies including Insilico Medicine and Recursion are advancing AI-discovered assets through clinical trials. At Owkin, OKN4395 has entered the Phase 1a clinical INVOKE trial. The company also said it trained its AI on real patient data for years and brought MSIntuit CRC through Europe’s CE mark into pathology practice.

The article said ongoing data from patient participants in the INVOKE trial are being used to improve the AI. Where predictions about patients’ responses missed the mark, the AI was retrained on real data to improve performance. It also said models trained on rich patient data, patient-derived organoids, and routine clinical use can bring model predictions closer to reality, but no drug-discovery, trial-design, diagnostic, or clinical AI can be successful without showing that the AI’s results work in humans.