'Doctor, I'm worried about this patient's observations.' We find them bolt upright, gripping the rails, wheezing with each gasp. RR & BP up, sats down. Classic APO - we got here just in time. Yet an AI for detecting deterioration might have missed this. Our paper explains why...

Our powerful @Daily_Express front page for tomorrow. The brutal suffering in Gaza must end. The shocking image shows Muhammad Zakariya Ayyoub al-Matouq, one, who weighs the same as three-month old baby due to the humanitarian crisis following the continued blocking of basic aid

Are Private Eye jokes now... criminal? PALESTINE Solidarity campaigner Jon Farley was arrested, bundled into a van and taken to the cells during a silent march in Leeds last weekend. What egregious act had prompted the police to act? He was, er, carrying a Private Eye joke piece

Age Ageing掲載論文 FI-Labsという検体検査だけでfrailtyをスクリーニングする方法の信頼性や妥当性をERでみたもの。 なかなかよさそうです。 https://t.co/eNKfHas9xI

@Krockdoc @jthteo @IbrahimZina @liamdunnell *at @KingsCollegeNHS and PRUH. @KCLDriveHealth

I honestly have the worst twitter game in the world "But in our last previous"

Thank you to all the patients whose data made this happen. @Krockdoc, @jthteo, @IbrahimZina for being incredible supervisors, @liamdunnell and the clinical teams at KCH who collected the scores, and anyone still reading for any engaging questions! Do fire away!

For the full story on this work, you can read the paper in @Age_and_Ageing:

But to truly know if any score, AI or rudimentary can make a difference for patients in practice, the ultimate test has to be a prospective RCT. That's my goal for FI-Lab 12/14

In fact, that's exactly what I'm working on now. A project comparing these simple, transparent FI-Labs to more sophisticated and complex AI-generated scores in a simulated clinical setting. 11/14

We think the future isn't about replacing clinical judgment but augmenting it. The next step is to get these scores in front of clinicians to see if they think they are helpful, and to see if they can help anchor assessments and improve care. 10/14

So which FI-Lab model performed best? Did adding more data and complexity help? Not really. Complex recipes came at the expense of the number of people we could score, for example, adding drug data excludes everyone without it. The difference in model performance was small. 9/14

Besides which, there's so much information can be captured with a human assessment which will NEVER make it into the EHR. It's the power of the end-of-the-bed-o-gram! (We discuss in more detail here how this will be a challenge for any AI: https://t.co/BrYViCHzW2) 8/14

The nurse's assessment was still better at predicting 90-day mortality (Hazard Ratio 1.55 vs. 1.29 for our best FI-Lab). Perhaps the "wobble" in nurse assessments isn't a bad thing; it's capturing the severity of the acute illness, which is powerfully predictive. 7/14

Here’s the core trade-off we found: Our automated FI-Lab was MUCH more stable visit-to-visit (Intraclass Correlation up to 0.76) compared to the nurse-assessed CFS (0.37). BUT… 6/14

We tested several FI-Lab recipes with different time windows, adding medication data, and changing how much data was required and how it was averaged. The big question: How do these automated scores stack up against a human assessment for stability and predicting outcomes? 5/14

So, we asked: can we build a more stable, data-driven measure of frailty? We used the averages of routine lab results from the past 12-36 months to create a "Chronic Frailty Index from Labs" (Chronic FI-Lab), which by design isn't affected by the immediate clinical picture. 4/14

Why was this happening? We looked at the factors associated with higher scores, and found high NEWS scores, and presenting with things like delirium tended to result in patients being rated as more frail. Though we still couldn't explain a lot of the variability 3/14

You can see that in this plot of 4 patients with the most attendances to ED during the study period. The scores wobble all over the place. That doesn't fit with our traditional understanding of frailty, as a concept which tends to change slowly. (NB fig captions as alt text) 2/14

And now for something completely different! https://t.co/TOiV0quYeH But in our last previous paper, we showed that CFS scores were useful in ED, but found a lot of variation. A patient could be rated as CFS 7 one week and CFS 2 the next. 1/14

I have written up my thoughts on "Physician Associate graduates have comparable knowledge to medical graduates" here on reddit (where I can do things like edit): https://t.co/653GvDeH2z *I strongly disagree with any comments criticising the authors personally*