Please enable JavaScript.
Coggle requires JavaScript to display documents.
Communicate Model Insights (Information to be communicated (Business…
Communicate Model Insights
Information to be communicated
Business Problem
Model Quality Metrics (confusion metrix)
Model Struggles
Predictive features
Features interesting for MGMT
Recommended business actions
Generalizable relationshipe between features and the target, understand and generalize for the future
3-7 things they didn't know
Unlocking hold out
Final opportunity to evaluate the model
Models stay in same place, or stay at the top
Business problem first
Data manipulations, discharge decisions
Updating problem statement, discuss with SME
Pre-Processing and model quality metrics
Confusion matrix - plan the quality metrics
Data processing should be just one slide
no code
Performance will deteriorate if model is given new data
positive predictive value
precision and its ability to predict
different probability distribution thresholds
Filter patients
remove all patients not in the subgroup
Predict
Select hold out as optional feature
Compute predictions
consider potential proposals
Areas where the model struggles
Internal data
Are repeat visitors the same problem?
Patient records contain additional data to mine
External data
Would public data actually be helpful?
Most predictive features
Directional results make intuitive sense
Provide only the data that's interesting to audiences at the appropriate level
Not all features created equally
Are these features something C-suite executives can change?
Features that need to be changed, that require a re-run of models
Features that require further examination
Why wasn't someone given an admission ID?
Add additional columns to evaluate the effect
Immutable features
Number ofyears passed since someone entered an industry
Mutable features
features mgmt could change
employee turnover and hours worked
Kaiser Permanente could detect the opioid epidemic much faster
as details are changed, models predictive value could change
Recommended business actions
Implement the model to find 62% of patients most likely to be readmitted
Institute a pilot program
data extraction and purchase pilot program