How AI and ML Will Save Lives in Healthcare, Episode 3

Ep 3 Stroke Detection at the Scanner

When a clot blocks one of the large arteries of the brain, the patient loses about 1.9 million neurons every minute. The treatment, thrombectomy, is one of the most effective procedures in medicine: in the HERMES pooled analysis of five randomized trials, it left 46% of patients functionally independent against 26.5% without it. The catch is time, and finding the patient in the first place.

This is the third episode of a series on where AI and machine learning actually save lives in healthcare, one clinical area at a time. This article gives you the ideas and the math. The full deep dive, with the code, the queueing model, the network simulation and the study design, is on Substack.

Every hour costs 5.2 points

The benefit of thrombectomy fades at a measurable rate. In the HERMES time analysis, each hour of delay to reperfusion was associated with a 5.2 point drop in the share of patients who ended up independent. That turns minutes into outcomes:

ΔI = N × Δt × β / 60

N is thrombectomies a year, Δt the minutes saved per patient and β the 0.052 drop per hour. A hub doing 150 thrombectomies a year that saves 11.2 minutes on each gets about 1.46 more independent patients a year. A lifetime model built on the same trials puts every 10 minutes at 39 days of disability-free life and $10,593.

What the best trial found

The only randomized trial of AI large vessel occlusion detection randomized the order in which four Houston stroke centers switched it on. Door-to-groin time fell by 11.2 minutes from a median of 100. The trial was too small to show a difference in outcomes, and it measured patients already at a thrombectomy center.

In England, a national evaluation found thrombectomy rates rose from 2.3% to 4.6% at the 26 hospitals that adopted AI imaging software, against 1.6% to 2.6% elsewhere. That study was observational, and several authors have ties to the vendor. But it points at the bigger opportunity: patients at hospitals that cannot do the procedure.

The minutes are in the queue

An AI detector does not read a CT angiogram better than a neuroradiologist. It changes the order of events. Much of the time it saves is time a study spends waiting on a busy reader’s worklist, and queueing theory says how much.

Wq = ρS / (1 − ρ)

With a reader busy 80% of the time and 10 minute reads, a study waits an average of 40.0 minutes. Flag suspected occlusions to the top of the list and they wait 9.4 minutes. Everything else waits 6.8 minutes longer, because every false flag jumps the line too. The full article includes the code for both, checked against a simulation.

Faster transfers treat more patients

In a 2026 US study of 22,410 patients transferred for thrombectomy, the median time at the first hospital was 121 minutes and only 26.3% left within the 90 minute target. Patients who left after 270 minutes had 0.35 times the odds of being treated at all.

I simulated a spoke hospital transferring to a hub. With parallel AI alerts, the median time to departure fell from 119 minutes to 72, and the share leaving within 90 minutes rose from 24.6% to 72.9%. Per 1,000 transfers, that meant 34.5 more functionally independent patients: 20.4 treated sooner, and 14.1 who reached thrombectomy only because they arrived in time. Most vendor studies never count the second group.

What’s in the Full Article

• The alert math: how many calls a day the stroke team gets, and how many are false
• The read queue model, with code, and why results vary so much between sites
• A simulation of a stroke network, with every assumption stated
• The economics: extra independent patients, disability-free years and cost per outcome at three software prices
• How to prove it: why 85 patients per arm can prove the minutes and 41,446 would be needed to prove the outcomes
• An executive brief with three questions for your stroke and imaging leaders

It is written for stroke program leaders, neuroradiologists, informatics teams and health system executives deciding whether, and where, to deploy AI stroke detection.

Read the full deep dive on Substack → Here

Paid subscribers get the full article and key code. Team Licensing members get the complete code, implementation plan and reference architecture for running this inside a health system.

About the Author

Brian James Curry is a Kansas City AI researcher, applied scientist and data scientist working in machine learning, causal inference and healthcare intelligence. He is the founder of Vector1 Labs, an independent applied AI research lab, where his Causal Care series applies causal machine learning to health system decisions: which programs work, for whom, and what they are worth. He also founded KC AI Lab, which began in 2016 as the Kansas City Machine Learning Group and has grown to more than 1,500 members.

He has more than twenty years of experience leading analytics, data science and machine learning at Hallmark, Garmin, Koch Industries, Tractor Supply Company and Vail Resorts. His open-source research tools on GitHub include PyCausalSim (causal discovery through simulation), Papilon (marketing mix modeling and complex systems) and tincast (forecasting for AI-critical commodities), along with MeaningFlow and Daedalus. A Kansas State University alum from a small Kansas farming town, he writes about agentic AI, causal inference, healthcare AI and the economics of AI on Medium and Substack. Contact: brian@vector1.ai.