Starting with "Whats *HOT* in data science?" Kirsten COLPAERT
- hype cycle for data science & ML in 2022.
- innovation trigger -> peak of inflated expectations -> trough of disillusionment
#LIVES2022
#datascience #WeAreICU #TogetherICU #IntensiveCare #CriticalCare @ESICM
- hype cycle for data science & ML in 2022.
- innovation trigger -> peak of inflated expectations -> trough of disillusionment
#LIVES2022
#datascience #WeAreICU #TogetherICU #IntensiveCare #CriticalCare @ESICM
Still some time for newer tools. Aiming for path towards productivity. 
Potential usage of "Synthetic data" to "preserve privacy" and "augment unbalance/or biased data". This is potentially interesting as it allows more transportability of models from one health care system to others.
e.g., 7000,000 patients training set with only 6.8% female. thus will not be applicable to female patients. Chen et al 2021 for synthetic data.
nature.com/articles/s4155…
nature.com/articles/s4155…
2. Deep-Learning : learning by example using "neural networks". nature.com/articles/s4174… interesting approach with structured and unstrcutured data turned into tokens without much feature engineering (I am sure people will debate on that ;) ) I find interesting thing though
Deep-learning models widely used in antibiotic discovery in research to help predict antibiotics.
- Reinforcement learning is still in upper slope of hype cycle. 2500 patients ICU Fluid guidance using reinforcement learning.
- Reinforcement learning is still in upper slope of hype cycle. 2500 patients ICU Fluid guidance using reinforcement learning.
Difficulty with DRL strategies in real life compared to say computer game is that there is no unobservable counterfactual data.
To unpack this, means cant go back in time and keep all yr variables and re-assign patient to other treatment for example. (easier in computer game)
To unpack this, means cant go back in time and keep all yr variables and re-assign patient to other treatment for example. (easier in computer game)
Natural Language Proccessing : medical records have a lot of texts unfortunately. NLP is helping unlock this further.
- Explainable AI. Contrasting interpretability with Predictive power. model evaluations (F1 score) will not be sufficient. need to understand mechanistically
- Explainable AI. Contrasting interpretability with Predictive power. model evaluations (F1 score) will not be sufficient. need to understand mechanistically
Saliency mapping as an example to overcome "one-pixel attack". (one pixela attack means adding or changing 1 pixel results in wrong classification by AI model).
How about "Trust worthy AI"? AI algorithm in sepsis learning to say "i dont know"
How about "Trust worthy AI"? AI algorithm in sepsis learning to say "i dont know"
nature.com/articles/s4174…
- Federated Learning has a lot of potential but early days.
Summary : exciting new growths but need to go to bedside to make it more useful.
@ESICM #LIVES2022
#ai #criticalcare #datascience
- Federated Learning has a lot of potential but early days.
Summary : exciting new growths but need to go to bedside to make it more useful.
@ESICM #LIVES2022
#ai #criticalcare #datascience
@threadreaderapp unroll
• • •
Missing some Tweet in this thread? You can try to
force a refresh
