We focus on machine learning applied to clinical decision support tasks that integrate with electronic health records (EHR) and rely on EHRs for input data. For more on the differences between decision support and automation check out @DrTJamiesonhttps://qualitysafety.bmj.com/content/28/10/778.info …
-
-
Prikaži ovu nit
-
We identify 4 steps along the path of translation, including (1) design and develop, (2) evaluate and validate, (3) diffuse and scale, (4) monitoring and maintenance. Within each step, we identify important milestones and activities that are common for teamspic.twitter.com/ROGSpQQPDw
Prikaži ovu nit -
For each of the 21 products, we provide details on the use case, the origin of the product development effort, and translational milestones along each step. While there are similarities, there are also important differences. Some examplespic.twitter.com/QOHbaz6Cxn
Prikaži ovu nit -
Some highlights: (1)Clinical integration remains the primary barrier for ML products. Many of the products have early evidence of statistical validity, some have evidence of clinical validity, but almost none have evidence of economic validity. It may be a while before we see ROI
Prikaži ovu nit -
(2)Approaches to external validation differ dramatically. Many products developed within health systems are put into practice with limited external validation. That being said, many of these undergo rigorous internal and temporal validation...
Prikaži ovu nit -
(2 cont'd)Two products have notable multi-national external validations (Kidney Failure Risk Equation and Colon Flag). Other products notably train new versions of models for new settings.
Prikaži ovu nit -
(3)The role of academic medical centers in driving ML innovation cannot be overstated. 16 of the 21 products were initially developed within a health system. Many are then licensed out or acquired by commercial entities. The products have collectively >$200MM
Prikaži ovu nit -
(4)There is a great deal of stealth science. Many of the products have 0 peer-reviewed published findings. This includes anything related to statistical, clinical, or economic utility. Unfortunately, it's common:https://onlinelibrary.wiley.com/doi/full/10.1111/eci.13072 …
Prikaži ovu nit -
(5)Still a long way to go for any of these products to successfully scale and diffuse across settings. Many challenges remain.
Prikaži ovu nit -
It was fun to place our own work in the context of many others paving the way in this field (incl
@suchisaria@CjBayesian@mdraugelis@JFutoma) and it's exciting to see what's to come. Hope this helps share insights with folks who've been asking for real use casesPrikaži ovu nit -
Prikaži ovu nit
-
Also features the great work out of
@DeepMind by@DrAConnell@weballergy@alan_karthiPrikaži ovu nit
Kraj razgovora
Novi razgovor -
Čini se da učitavanje traje već neko vrijeme.
Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.