Das KI-System kann eine Sepsis mit einer Genauigkeit von 99 % diagnostizieren, bevor sie lebensbedrohlich wird
https://news.northeastern.edu/2025/10/15/ai-agent-helps-er-doctors-predict-sepsis-shock/
Das KI-System kann eine Sepsis mit einer Genauigkeit von 99 % diagnostizieren, bevor sie lebensbedrohlich wird
https://news.northeastern.edu/2025/10/15/ai-agent-helps-er-doctors-predict-sepsis-shock/
5 Kommentare
I can predict sepsis with 100% accuracy by diagnosing every ED patient with sepsis. Lots of false positives, but I’d never miss a diagnosis of sepsis. Without any further information regarding Sensitivity and Specificity, we can’t really say anything about the clinical utility of the tool.
Is it not just the collection of specific diagnostic criteria at various points adding up to the diagnosis of sepsis that it’s doing?
I’m not exactly sure how that needs to be ai in particular and not just an algorithmic tool that can calculate the risk based on certain data collected and input into a chart. Wouldn’t that be a less energy intensive way of doing it rather than using ai specifically?
I’m also curious because in practice wouldn’t that mean the provider needs to input information somewhere before making the diagnosis themselves? What if the issue pertaining to failed diagnosis is the charting and missed information?
Also, there was no comparison of whether a doctor given all the same collected data fed to the ai would come to the same conclusion as the ai model…without even needing the ai model itself.
I fully expect that in the near future all our diagnostics data will be fed into AI screening systems. Including a high definition video so it can see our skin and behavior.
Got a chest x-ray for a spinal injury? AI will go ahead and do a thorough screening for cancer while you’re at it.
Eventually it’ll be predictive. Using medical history and demographic data it’ll suggest screenings you might not have otherwise received.
99% accuracy—but what’s the specificity? If I think every person in the world is named Jack, I will correctly identify Jacks with 100% accuracy.
Especially important with infection. Overusing antibiotics is a great way to hasten our demise.
This paper – in a predatory MDPI journal – is just…kinda mad?
They never actually define the sepsis outcome, but the models use SOFA scores…? That’s…circular?
The study is a synthetic dataset of 70 sepsis patients and 70 nonsepsis patients.
They do no training/testing split – the 99% claimed accuracy is based on the entire dataset. This is fatally poor. And, these accuracy claims (even if they were true) based on a synthetic dataset are not relevant to real world use.
There is no early detection timeframe, despite the claim. Half of the models are using labs from late in the clinical course?
The input symptom data, clinician rated symptom severity across a number of domains, are nonstandard, and results are not reported in any detail. It is difficult to believe these were defined prospectively (this seems to be a repurposed dataset), so how were they consistently defined…? The actual patient characteristics of the dataset are completely undefined – instead they pad the paper with a lot of guff and “representative reports” – so much stuff is missing from the TRIPOD AI reporting guidelines.
Finally there doesn’t seem to be any ethics statement.