How AI and ML Will Save Lives in Healthcare, Episode 1

Ep 1 Sepsis Early Warning

About 1.7 million American adults develop sepsis every year, and at least 350,000 of them die in the hospital or are discharged to hospice, according to the CDC. Worldwide, The Lancet estimated 48.9 million cases and 11 million deaths in 2017, roughly one in five deaths on Earth. Sepsis kills through time. By the moment it looks obvious at the bedside, organs are already failing.

That makes sepsis the best place to ask whether machine learning can actually save lives in a hospital. This is the first episode of a series on exactly that question, one clinical area at a time. This article gives you the ideas and the math. The full deep dive, with the training pipeline, evaluation code and study design, is on Substack.

Two models, opposite results

At Johns Hopkins, a system called TREWS monitored 590,736 patient encounters across five hospitals. When a clinician confirmed a TREWS alert within three hours, the patient’s adjusted mortality was 3.34 percentage points lower than for patients whose alerts were dismissed or never evaluated, a relative drop of about 18%.

At the University of Michigan, researchers tested the Epic Sepsis Model, one of the most widely deployed sepsis tools in the United States. It reached an area under the curve of 0.63, well below the 0.76 to 0.83 reported elsewhere. It missed 67% of sepsis cases while alerting on 18% of all hospitalized patients.

Same disease. Same goal. One model is associated with saving lives and the other mostly generated noise. The difference is not the algorithm. It is everything around it.

Lives saved is a product, not a score

A sepsis model saves a life only when five things happen in a row: the patient has sepsis, the model flags them, the flag comes early enough, a clinician acts, and acting early changes the outcome. That gives a simple proposed model:

L = N × p × Se × c × Δm

N is annual admissions, p is sepsis prevalence, Se is the share of cases flagged in time, c is the share of true alerts clinicians act on promptly, and Δm is the mortality reduction when they do.

Put in real numbers. A hospital with 30,000 admissions a year, 6% sepsis prevalence and a model with 0.80 sensitivity, where clinicians act on true alerts at the TREWS rate of 0.61 and acting cuts mortality by 3.34 points, saves about 29 lives a year. Drop the sensitivity to the 0.33 Michigan measured and the figure falls to about 12.

Because the terms multiply, a weak link anywhere drags the whole product toward zero. A brilliant model that clinicians ignore saves fewer lives than a decent one they trust. Most sepsis programs measure AUROC and never measure c at all.

The full article walks through each term with code you can run against your own hospital’s numbers.

Every alert is a Bayes problem

Why do clinicians ignore sepsis alerts? Arithmetic. The chance that an alerted patient actually has sepsis is the positive predictive value:

PPV = Se·p / (Se·p + (1 − Sp)(1 − p))

At 6% prevalence, a model with 0.80 sensitivity and 0.90 specificity has a PPV of 0.338. In that 30,000-admission hospital it fires on 11.7 patients a day, and two of every three are false alarms. Raise specificity to 0.95 and the PPV climbs to 0.505, with 7.8 alerts a day. Small gains in specificity cut alert volume sharply, because non-septic patients outnumber septic ones nearly 16 to 1.

The Michigan figures fit the same formula. A PPV of 12% means a clinician reviews about 8.3 alerts to find one case of sepsis. That is how alert fatigue is built, one false alarm at a time.

On synthetic patients, the full article compares a classic SIRS rule with a gradient boosted model. At nearly the same sensitivity (74% against 76%), the model alerts on 33% of patients instead of 57%. That is 41% fewer alerted patients, with the PPV rising from 9.3% to 15.4%.

Lead time is the metric that saves people

AUROC tells you whether a model ranks patients well, not whether it warns anyone in time. A model that flags sepsis an hour before the team would have noticed anyway saves almost nobody.

The hourly cost of delay shows what lead time is worth. Using the New York finding that each hour of antibiotic delay multiplies the odds of death by 1.04, six hours of delay moves a 20% mortality risk to 24.0%. That is an observational association, not a clean causal estimate, but it is close in size to what TREWS reported. Johns Hopkins said that in the most severe cases TREWS flagged sepsis nearly six hours earlier than traditional methods.

There is a trap here too. A noisy rule that fires on more than half of all patients will often look “early” by coincidence. The full article shows how to measure lead time honestly, against when your clinicians would have acted anyway.

Proving a model saved anyone

The best real-world sepsis results come from observational designs. TREWS compared confirmed alerts with ignored ones, which invites confounding. At UC San Diego, the COMPOSER model was tied to a 1.9-point drop in sepsis mortality (17% relative) across 6,217 septic patients, but in a before-and-after design during years when emergency medicine changed fast.

Health systems can do better because they control the rollout. The full article covers stepped-wedge rollouts, silent-mode validation and regression discontinuity at the alert threshold. It also sizes the trial: detecting the TREWS effect takes about 1,957 septic patients per arm, and roughly 7,787 in total once you randomize by hospital unit.

What’s in the Full Article

• The lives-saved model with code to run on your own hospital’s numbers
• A sepsis training pipeline on synthetic patients, with the leakage traps that inflate published results
• Patient-level evaluation code: sensitivity, PPV, alert burden and lead time
• A threshold comparison of SIRS against a gradient boosted model
• Study designs and a sample size calculator for proving lives were saved

It is written for clinical data scientists, informatics teams and health system leaders who own a sepsis alert, or are about to buy one.

Read the full deep dive on Substack → click this link

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

About the Author

Brian James Curry is a Kansas City AI researcher and data scientist, the founder of Vector1 Research, an independent applied AI lab, and of KC AI Lab, which began in 2016 as the Kansas City Machine Learning Group and grew to more than 1,500 members. He has more than twenty years of experience leading analytics, data science and machine learning at enterprises including Hallmark, Garmin, Koch Industries, Tractor Supply Company and Vail Resorts. His open-source work, including Papilon, PyCausalSim, MeaningFlow, Daedalus and tincast, is on GitHub. He grew up in a small Kansas farming town and attended Kansas State University. He writes about agentic AI, causal inference and the economics of AI on Medium and Substack. Contact: brian@vector1.ai.