Slashdot Mirror
Researchers Develop AI To Predict Hospital Readmission Rates From Clinical Notes
Researchers at New York University and Princeton have developed a framework that evaluates clinical notes and autonomously assigns a risk score indicating whether patients will be readmitted within 30 days. They claim that the code and model parameters, which are publicly available on Github, handily outperform baselines. VentureBeat reports: As the researchers point out in a preprint paper on Arxiv.org, clinical notes use abbreviations and jargon, and they're often lengthy, which poses an AI system design challenge. To overcome it, they used a natural language processing method -- Google's bidirectional encoder representations from transformers, or BERT -- that captures interactions between distant words in sentences by incorporating global, long-range information. Each clinical note is represented as a collection of tokens, or subword units extracted from text in a preprocessing step. From multiple sequences of these, ClinicalBERT identifies which tokens are associated with which sequence. It also learns the position of tokens from variables corresponding to the sequences, and inserts a special token used in classification tasks in front of every sequence.
To train ClinicalBERT, the team sourced a corpus of clinical notes and masked 15 percent of the input tokens, forcing the model to predict the concealed tokens and whether any two given two sentences were in consecutive order. Then, drawing on the Multiparameter Intelligent Monitoring in Intensive Care (MIMIC-III), an electronic health records data set comprising over two million notes from 58,976 hospital admissions of 38,597 patients, the researchers fine-tuned the system for clinical forecasting tasks. Tested on a sample set consisting of 30 pairs of medical terms designed to assess medical term similarity, the authors report, ClinicalBERT achieved a high correlation score, indicating that its tokens captured similarity between medical concepts terms. Heart-related concepts like myocardial infarction, atrial fibrillation, and myocardium were close together, they say, and renal failure and kidney failure were also close.
To train ClinicalBERT, the team sourced a corpus of clinical notes and masked 15 percent of the input tokens, forcing the model to predict the concealed tokens and whether any two given two sentences were in consecutive order. Then, drawing on the Multiparameter Intelligent Monitoring in Intensive Care (MIMIC-III), an electronic health records data set comprising over two million notes from 58,976 hospital admissions of 38,597 patients, the researchers fine-tuned the system for clinical forecasting tasks. Tested on a sample set consisting of 30 pairs of medical terms designed to assess medical term similarity, the authors report, ClinicalBERT achieved a high correlation score, indicating that its tokens captured similarity between medical concepts terms. Heart-related concepts like myocardial infarction, atrial fibrillation, and myocardium were close together, they say, and renal failure and kidney failure were also close.
Cue in greedy insurance companies digesting that data and kicking off their 'ai' designed to ramp up premiums. Excellent.
doctor handwriting ocr with an 95+ rate?
virtual doctor says "You get leprosy good bye"
https://www.youtube.com/watch?...
Re:virtual doctor says "You got leprosy good bye"
Merely a stupidly over-simplified (compared to real neurons) "neural" net function, based on a few matrix multiplications of fine-tuned weights that turn input into output.
A crutch function, for when you don't know the actual formula/algorithm to implement.
AI means it's basically an individual thinking artificial lifeform.
Keep your hospital bed warm for you?
an AI that given a bunch of symptoms, blood analysis, weight loss / gain, ... can tell you a narrow list of diseases you might have ; that would surely beat the best human specialists.
Slashdot, fix the reply notifications... You won't get away with it...
https://www.youtube.com/watch?...
They just think they did. More 'AI' bs.
It really is. Instead of dedicating their effort to improve humanity and put lawyers out of business, software engineers are throwing major efforts towards putting doctors out of business. Which just goes to show that they can't think further than the last curly brace in their programs. Once all the doctors are out of business, just exactly who you they think become the next target for all the starving lawyers out there left without doctors to sue?
Seven puppies were harmed during the making of this post.
Whoa! These related concepts are actually close together! Can u believe it ??
What if we could LITERALLY build a robot that points out the obvious?
It will be like: THESE THINGS ARE SIMILAR *beep beep*
Notes: Patient is deceased.
--sf
Why, oh why, do you guys keep publicizing research before it's peer reviewed?
"Interesting [x]"
If you read the paper, you’ll find this 15% improvement over previous work, for the 48-72 hour range, has a recall, at 80% precision, of 0.171 ± 0.107. In other words, the recall is somewhere between 7% and 27%, if you want to ensure that you don’t get more than 1 in 5 false positives. If you’re a clinician and you’re trying to use this for a specific patient, it’s useless; a 27% recall rate means that three quarters of the time, it’ll tell you a patient won’t be readmitted and it’ll turn out that she is, and remember, the precision is 80%, so 20% of the time if it says she’s going to be readmitted, she won’t. The work is really interesting; BERT is a really important advance. But clinical language understanding is unbelievably hard, and the area has a long way to go to be reliably useful. VentureBeat is clickbait.
Hey Bert
Yeah Ernie?
Ain't 'merkin medicine great?
Yeah Ernie.
So a little bit of context... this is actually a fairly active area of research and has been for decades. I know people who have been using AI techniques to try to predict readmissions probabilities since the 90s, and some of them have gone on to develop these into commercial products that hospitals use today... so the story here really is not the use case but instead the incremental improvement of this particular corner of AI. I am actually just getting off a project that was trying to apply another area within AI to the same basic problem, though since we were using agent based modeling we probably will not get much attention.
And this stuff is honestly pretty useful. It helps earmark which patients might need a bit of extra monitoring or guidance, or which ones might benefit from having a nurse check in on them at home. It means figuring out where to proactively spend limited human resources in order to decrease the chances someone is going to end up right back in your ER. Good stuff.