To Do.txt

X 1. Look at list of 200+ variable names to get general understanding of predictors.
X 2. Get domains of each predictor (boolean, whole number, real, polytomous) and min-max values - Row 16 has almost every column as non-null.
X 3. Put data into SQL Server for ease of checking data
X 4. Get Rsquared for each predictor
X 5. - Is actualLOS the training variable or adjustedLOS
6. Admit vs Visit vs Encounter?
7. Issue with history mean stats - no idea how many data points created the mean or their temporal spacing
8. Consider identifying and removing outliers - all are in nullable columns so only do when running nullable columns model and for features I know nothing about remove anything 3 standard deviations from the mean.
9. How to deal with substantial missing data
10. primaryDiagnosisChronicCondition is a real number? Why?
11. Duplicate EncounterId - no primary key in data
12. What about normalization of variables (depends on model - logistic regression, SVM, neural net are most likely classification models I'd use that need it)

ProcedureCount is only known post discharge (probably true)

Coefficient of Determination is low for all the non-null predictors other than procedures count (which is .2) Running model #13 has significant risk of poor accuracy.
Possibly any tree will overfit with such low coefficient of determination numbers.

Update LOS.yml with R needs and remove notebook.

No apparent variables about procedures

13. Run model with only non-null columns (47 predictors) and only records whose encounterid is unique
14. Run model with null columnns (except replace impossible values with nulls, possibly replace 3 standard deviation outliers with nulls
15. Run model without tossing anything (other than replacements mentioned in #14)
16. Run model in non C5 models
17. Run unsupervised clustering models on false predictions (whatever false is determined when dealing with integral days of 0 to 20)

Beyond Scope
Figure out impossible values on all the lab test results - and if like Fahrenheit vs Celsius reading mistake in temperatures, correct values where possible.