It’s not easy to get depression-detecting AI through the FDA
Why it matters: Kintsugi's shutdown highlights the significant regulatory hurdles for novel AI in healthcare, impacting future mental health screening innovations.
- Kintsugi, a California-based startup, developed AI to detect signs of depression and anxiety from speech patterns over seven years, aiming to offer a more objective signal than traditional self-report tools.
- The company is shutting down and releasing most of its technology as open-source due to the inability to secure FDA clearance in time, with some elements potentially finding new applications beyond healthcare, such as detecting deepfake audio.
- Kintsugi's software analyzed subtle shifts in speech patterns like pauses, sentence structure, or speed, which are known indicators of mental health issues, and reported results in peer-reviewed research broadly in line with established self-report screening tools.
- The FDA's "De Novo" pathway, intended for novel, low-risk medical devices, proved challenging for Kintsugi, as the framework is designed for fixed devices rather than continuously optimized AI systems, requiring extensive time to educate regulators about AI.
- Mental health assessments largely rely on patient questionnaires and clinical interviews, such as the PHQ-9, which Kintsugi's voice-based model sought to complement or potentially replace, arguing for expanded, scalable, and more objective screening.
Kintsugi, a California startup, is shutting down and open-sourcing most of its AI technology designed to detect depression and anxiety from speech after failing to secure timely FDA clearance. The company's innovative approach, which analyzed how people speak rather than what they say, aimed to complement or replace traditional self-report screening tools but faced significant hurdles navigating the FDA's regulatory framework, which is ill-suited for rapidly evolving AI. While Kintsugi's core healthcare mission ends, elements of its technology may find new life in other sectors, including deepfake detection.
