4/ The math shows something pretty neat: if each individual clinician is better than 50% accurate at the labeling task, adding more clinicians to your voting ensemble will keep improving accuracy toward 100%. This is also the idea behind ensemble learning methods.

en.wikipedia.org/wiki/Condorc...

3/ To address this, a common approach is having multiple clinicians independently label the same data, then using majority voting to determine the "ground truth" label. But to what degree does this actually help? This is where Condorcet's Jury Theorem comes in.

2/ Instead, we'll often rely on clinicians to provide expert annotations of our data. These human-provided labels are used both for training and evaluating models. But even expert clinicians aren't perfect at all labeling tasks. Sometimes the relationship between data and labels is inherently noisy.

1/ When developing AI systems for medical diagnosis from images, waveforms, etc., one big challenges is getting reliable ground truth labels. Unlike many ML problems, we often can't get perfect quantitative measurements of what we're trying to predict/classify in medicine.

“Never underestimate your audience’s intelligence, always underestimate their vocabulary” is a rule I teach in my workshops.

And part of this is tone. Don’t act like your audience (family or not) is stupid for not knowing 15-syllable science words.

Right, like two Froude numbers or equivalent.

On one hand, it made some senior scientists seem more approachable at conferences, which is a good thing. On the other hand, social media in general seems to have a negativity bias, which is a bad thing. Hard to say overall.

Honestly, depending on the size of the library, it might be worth a try to upload it to a Claude Project and asking for help.

I've been thinking about this in light of the FDA's recently released guidance on "Assessing the Credibility of Computational Modeling and Simulation in Medical Device Submissions," which provides a really nice framework.

www.fda.gov/regulatory-i...

For example, what would it take science-wise for someone to be able to create a successful startup that uses OpenSim as the back end of a software-as-a-medical-device clinical decision support tool for orthopedic surgeons?

This is really cool work. Given the mixed results, do you have strong thoughts what the answer to your original question about apathy over time in people with PD?