If AI can predict your death, can healthcare save your life?
The scene opens on a ticking clock. The numbers are probably bright red.
The hour it displays is not the current time. It’s a countdown. A foreboding prediction of the exact day, hour, minute, and second when you’ll meet your maker.
This scenario straight out of a Black Mirror episode is what many people think of when we talk about AI being used to predict mortality. But that’s not quite what it looks like in real life (thankfully).
Instead, this tech is a tool that can help providers better focus their limited time and resources—and can save lives. If you ask us, it’s one of the most exciting applications of healthcare AI.
But how does it work right now in practice? Let’s take a look.
The new frontier of preventive medicine
If pressed, we bet you wouldn’t have a hard time finding an example of this kind of technology in the news.
A few examples of AI-enabled predictive medical analytics off the top of our heads:
That last study piqued our interest because of the latter risk prediction. Can specific body measurements really predict all-cause death?
According to the researchers, it’s possible.
They did find that their AI-obtained body composition measurements could inform all-cause mortality prediction. And as the power of these predictive models continues to increase with technology advancements and data quality, those findings become even more exciting.
The researchers of the lung cancer study also acknowledged that this area of predictive analytics is still understudied. So what’s new in this field now?
Over at NYU Langone, an LLM is predicting hospital readmission based on a patient’s EHR data. The model, called NYUTron, takes a wide variety of structured and unstructured data—from vitals to written medical notes—and forecasts readmission for a variety of issues.
NYUTron accurately identified 85 percent of patients who died in the hospital and correctly estimated 79 percent of patient admission length. These numbers both mark significant improvements over standard prediction models used by the health system before.
That kind of all-cause mortality and readmission prediction can save a lot of lives. And it can save providers and health systems a lot of time (and of course, money) as they focus interventions.
Plus, paired with other AI-enabled automation, patient intervention can get both more specific and smooth.
The risks of AI-driven mortality prediction
But predicting mortality using body measurements and imaging isn’t all fun and games. There are real stakes to making these kinds of predictions—and using them in patient care.
1. Correlation vs. causation
For one, when it comes to using body measurements in analytics, there’s a need to be really careful and differentiate between correlation and causation.
A great example of what we mean comes in a common criticism for weight management medicine.
This critique argues that the focus on higher weight as a risk factor for disease has turned weight loss into a panacea. Targeting the weight itself can then lead doctors and patients to neglect the underlying causes of disease (which may also be leading to the higher weight).
Another key caveat to keep an eye on?
2. Data diversity
We’ve said it once and we’ll say it again: Effective predictive analytics depend on data diversity.
Predictive accuracy (and usefulness) is closely tied to how representative the data is. Without data diversity, you get predictive algorithms working effectively only for some patients—and not others.
3. How we use these predictions
And as we’re ironing out the accuracy of these algorithms, we must safeguard the applications. The last thing we want is for risk predictions to be used as fact.
A powerful cautionary tale is in this investigation of how Medicare Advantage plans have used proprietary prediction scores to deny senior patients care.
Final thoughts from HAN
If you read our feature on the AI and Anti-Aging, you’ll know we’re excited about this topic.
AI has greater potential to predict health issues and thus save lives. There’s so much incredible innovation happening around this issue—and you can bet we’ll keep covering it.
But as we get excited about this technology, we must remember that predictions are just that. They’re not diagnoses in themselves.
AI-enabled predictive care shines when it’s used to help patients avoid poor outcomes they may be predisposed to. We must take care to not turn them into an immutable judgment call.