Life Sci AI: The Podcast

The episode contains a handful of genuinely useful specifics - competitor AUC scores from an FDA filing, the insight about training on pathological ground truth vs. clinical methodology, and the problem of university-only datasets not representing real-world hospitals. However, the opening third is largely biographical throat-clearing with limited operational value, and general AI optimism pads the second half.

“one of our competitors in their FDA filing, their area under the curve is 0.65, which isn't much better than tossing a coin”

“about 8 out of 10 hospitals in a country like the UK are not universities. Hospitals, their patients look different, they set their scanners up differently”

Two genuinely non-standard ideas surface: training on pathological truth rather than established clinical workflows like PI-RADS, and the argument that data diversity across non-academic NHS sites matters more than volume from single university centres. The broader AI commentary - embrace it, automate boring tasks, speech recognition just works now - is entirely recycled.

“we train our AI not on a clinical judgment, but on the pathology. And fundamentally, is this cancer or not?”

“I'm not a believer in the generalist AI that will rule the world”

Anthony Ricks is a legitimate practitioner - Cambridge-trained, a prior BT-funded PhD that generated patents and a spin-out, 12 years of consulting followed by founding a medical AI company that has achieved CE certification and conducted NHS clinical studies. He is not a career podcast guest, and his answers reflect real operational experience rather than thought-leadership posturing.

“Within four years we'd created an AI that was better at this task than people, got an international standard, got patents in IT and spun a company out”

“we've now secured CE certification for the product which has been a really major milestone under the mdr”

Several concrete data points land well: a named competitor's FDA AUC of 0.65, Lucida's 0.9 vs. the clinical average of 0.8 - 0.85, 12,000 UK male prostate cancer deaths annually exceeding female breast cancer mortality, and over a third of Scottish men first diagnosed at stage four. The NHS clinical study is referenced but never characterised in enough detail to evaluate independently.

“one of our competitors in their FDA filing, their area under the curve is 0.65, which isn't much better than tossing a coin. When you start to get to kind of 0.9 upwards, uh, here we're operating in territory that's typically better than doctors”

“literally 12,000 men each year in the UK die from m it more than women die from breast cancer in the UK”

The host lands one genuinely useful clarifying question (asking for a layman's explanation of AUC) and a decent structural question on challenges. However, most questions are long and self-answering, the host frequently interjects with affirmations and personal anecdotes rather than follow-ups, and no claim - including the competitor AUC figure - is ever probed or challenged.

“What would just so people understand area End of the curve, what would be the goal number?”

“could you have a self driving truck with a remote radiologist doing a screening program on someone in a car park in Germany”