They created a dashboard that shows patient basic information, prediction explanations translated for clinicians, and #SHAP plots of the most predictive variables

#AMIA2020

For this study, they tried to predict 90-day mortality using different data from labs, diagnoses, demographics, and flowsheets. The model will go live in Feb. 2021

#AMIA2020

There's an exponential growth of studies using #ML/#AI in #healthcare. However, there's a very slow adoption in #CDS. #XAI can help with the latter (for more about this, see this thread from a previous session x.com/amoncadatorres/status/13...

#AMIA2020

Last session of this bunch: @LorenzoARossi's "A Dashboard to Automatically Translate #SHAP Explanations to Clinicians for Mortality Prediction"

#AMIA2020

This was a test of the platform for acceptability. Its real effect still needs to be evaluated in real-life conditions (e.g., real tumor boards) and with a larger dataset

#AMIA2020

There was a positive evaluation from clinicians regarding attractiveness, perspicuity, and dependability. There was a large interest in having access to different guidelines

#AMIA2020

During 14 sessions across all parties, there were 125 (fully exploitable) clinical cases discussed corresponding to 110 patients. Cases were chosen to be representative of real-life cases

#AMIA2020

Evaluation was done in three steps. Close-to-real tumor boards were groups very similar to the real tumor boards, but without a time pressure

#AMIA2020

#DESIREE is a web-based software ecosystem for personalized, collaborative, and multidisciplinary management of primary breast cancer.

Evaluation was done in three clinical sites

#AMIA2020

#CDSSs have proven to have benefits in clinical decision-making (e.g., facilitate use of up-to-date clinical evidence, support patient-specific clinical pathway & guidelines [#oncoguide!])

#AMIA2020

Time for "Preliminary Evaluation of #DESIREE, a Decision Support Platform for the Management of Primary Breast Cancer Patients" by @SylviaPelayo

#AMIA2020

While a good overall performance (precision 96.4%, recall 98.3%, F-measure 77.3%, there were still some disagreements. This was attributed to too strict decomposition of active ingredients and to differences between hospital pharmacopoiea (that's a new word for me)

#AMIA2020

This is how #EzMedRec's categorization of the data look like.

#AMIA2020

Discrepancies were categorized in different groups

#AMIA2020

They developed #EzMedRec, a retroactive MR support system at hospital admission. They used expertise of 2 pharmacologists and the gold standard. Data were a set of manually-completed MR forms and the French public drug database

#AMIA2020

Now's the turn of Brigitte Seroussi: Easy Medication Reconciliation at Hospital Admission

Complex clinical pathways involve multiple settings and medical specialists. Medical reconciliation (MR) tries to prevent and correct errors at transitions of care

#AMIA2020

#AITE uses knowledge-graph based models to identify

- patient similarity
- the next best action (or question)
- recommendation inference

It's been heavily optimized for real-time applications. In a graph with 2M nodes and 20M edges, mean performance is of 0.2s/query

#AMIA2020

A key part of the system are the methods for ontology learning from real-world data. About 1M synonyms were merged into 17k medical concepts (linked by 62k relations) (!)

#AMIA2020

These are #AITE's building blocks

(Lots of moving pieces and combination of several technologies here!)

#AMIA2020

#AITE was validated, tested for end-user friendliness, and certified as a medical device (class 1) in the EU. It's been live since Q4 2019 and monitored ever since

#AMIA2020

Shortage of physicians and increasing costs have created the need for this type of telemedical services.

Care providers optimize their personnel. They can also offer a scalable easy-to-use solution, while ensuring that emergency cases are detected quickly reliably

#AMIA2020

Now with Chiara Marchiori with "#AI Decision Support for Medical Triage"

#AITE is an #AI application not for diagnosis, but for recommendation. It uses data from >900k (German) medical case records

#AMIA2020

Next session: S93 - #DecisionSupport!

(Due to some technical issues I missed the first talk 😓)

#AMIA2020

reach!)

#AMIA2020

A few interesting quotes from the juries

- It isn't essential to provide an explanation for an automated decision when it is a matter of life and death
- There must be an explanation in order to prove there is no bias in the criminal justice system

#AMIA2020

In non-healthcare scenarios, there wasn't a clear consensus on systems. However, explainability was even less important

#AMIA2020

In the healthcare scenarios, there was a large consensus on System C (beyond human performance). Explainability was "not very important" (!)

#AMIA2020

There were three hypothetical systems presented.

(Again, thought exercise, they weren't real systems)

#AMIA2020

Out of >250 people, 18 were recruited per jury. They were posed 4 real-life scenarios:

- Stroke diagnosis
- Shortlisting job applicants
- Identifying kidney transplant recipients
- Identifying suspects suitable for rehab programs

#AMIA2020

They used "citizen juries", which is exactly what it sounds like. Over 5 days, a group of people hear evidence, deliberate together, and answer general questions.

(Never heard of those before!)

#AMIA2020