Federated (Machine) Learning spins off to Med apps
Telcos, Med, Pharma, and on it goes. Looking back, the old saying “AI is what was impossible yesterday” fits on FL. I wrote about AI in health in my posts before this summer; now, those seem to converge in health and care, thanks to the CORD-19 AI Challenge and even smaller regional initiatives.
Compared to other domains, health, pharma, and care apps are very sensitive on privacy, data confidentiality, integrity, and even data tenancy; all of this makes FL come in handy.
As we know from applied AI in automotive, avionics, cybersec, natural language (NLP) and elsewhere, augmentation (collaborative apps) comes years ahead of automation (autonomous systems). Already before the CoViD crisis in healthcare, studies showed that diagnosis accuracy by either a doctor or an ML-based app was still significantly below the accuracy by both in collaboration.
So, autonomous or collaborative AI? ML works either way. Initially, a log of all cases where the app and the doctor arrived in opposite conclusions becomes even an extremely valuable addition to the training-data set. In FL, you can interleave periods of runtime and (re)training time, on the same local HW device or devices, whenever needed.
Spacing apart: when physical distancing matters
NASA initially intended surgical robotics to be used onboard the International Space Station ISS, and aircraft carriers. Just like GoreTex or Teflon, the technology made it all the way back down to Earth.
The first robotic surgical platform, commercially available in the US, was cleared by the FDA for laparoscopic surgery 20 years ago. Today, more than 1,700 of them are in hospital use worldwide. In complex laparoscopic arterial surgery, entering the body through several keyhole cuts, Prague’s university hospital took the lead this summer: 500 operations completed so far, by doctor Štádler’s team. Collaborative robotics turned out even better during Covid-restrictions. The tools of the robot work on/inside the patient, whereas the surgeons are several meters away.
FL gaining momentum
The Covid crisis speeds up even Med-ML activities other than the CORD-19. Large hospitals and professional organizations of doctors are open to AI-initiatives.
A few weeks ago, Stanford and professor D. L. Rubin implemented a publicly available FL framework, the Unified CT-COVID AI Diagnostic Initiative(UCADI), “allowing any hospital or institution around the world with the right infrastructure and data to join”.
An open-source framework of AI models, baselines, and evaluation metrics, fastMRI by Facebook AI and NYU School of Medicine aims at 10X faster MRI scans of joints (because MR is both expensive and time consuming). Last week, Norway´s medical daily Dagens Medisin let doctor Helga Brøgger, chair of the Norwegian society of radiologists, comment on the first results of fastMRI. If you don’t read any Nordic language, you might try ML by GoogleTranslate (IMHO, ready for collaborative use but not for autonomous. Yet… )
Don’t get me wrong: I’ve no connection whatsoever to social media firms, or search engines, or old media. Perhaps, some architectural foresight. In ML right now, things keep coming, well… automatically. 🙂
Summing up, FL combines the qualities of privacy/local ownership of data, secure international involvement in ML projects, accuracy built by ML from big data, resilience to data heterogeneity, interleaved runtime and training time, minimized network footprint (no training-data transfer), and not least, off-the-shelf public-domain frameworks.
See Milan’s courses:AI, machine learning