Goals

To preprocess PROACT dataset

Data files You can find the data files at the following url.
https://www.dropbox.com/sh/3lmi5ii3sgyi7o3/AACeVLTGtSDXSKIaX6P7Hxj2a?dl=0

or in the following folders.
Dropbox/ALSmaster/PROACT/PROACT_rawdata
Dropbox/ALSmaster/PROACT/PROACT_preprocessed

Data source

PRO-ACT database https://nctu.partners.org/proact
The above data source was accessed in Sept. 2017

DB contains data of 10,723 patients (including all datasets for training, training2, leaderboard, validation)
4,456,146 records

Workflow

preprocess.R

Preprocess 1

1. Filtering
   • exclude duplicate records, records with inconsistent or errorneous values (feature_delta < 0, diag_delta > 0, onset_delta > 0, diag_delta < onset_delta, both Q5a and Q5b filled-in)
   • exclude patients with incomplete ALShx or ALSFRS-R all item scores

Preprocess 2

2. Creating meta-features (mean and slope values)

   • calculate mean value for time-varying features (ALSFRS_R_Total and item scores, fvc, bmi, vital sign, and lab)
   • estimate slope for ALSFRS_R_Total, subdimension (bulbar, motor, respiration), and fvc over the first 3 mo, using linear model (time interval between the first and last measurements should be > 1.5 mo & the number of measurements should >= 2)
   • calculate onset2dx (onset to diagnosis)

3. Transforming features

   • convert the sign of onset_delta and diag_delta from minus to plus
   • convert the unit of feature_delta, onset_delta and diag_delta (days to months)
   • combine the categories of onset_site (Limb, Limb and Bulbar, and Others) into Nonbulbar
   • create a new feature "Gastrostomy" (True/False)
   • convert Q5a/Q5b items into Q5

4. Preprocessed data files

   • Predicting features data PROACT_preprocessed_rev.csv
   • Target data: Loss of functional autonomy in swallowing LOA_swallowing.csv

Imputation

merge_predictors_target.R
merge data (predictors and target)

• target data
• predictors: static and meta-features over the first 3 mo
• evaluate missing data pattern and proportion: missing_prop_pattern.png
• create a merged file (for missing value imputation) merge_predictors_target.csv

impute4slope_rev.R

• impute missing values in merged_slope_rev.csv
• multiple imputation using mice package
• exclude features with missing data proportion > 30 % except for bmi_mean
• data distribution before imputation: distribution_data.png
• missing data pattern: missing_pattern.png
• check convergence following multiple imputation: convergence.png
• check imputed value density plot (continuous features): imputation_densityPlot.png
• check imputed value proportion plot (categorical feature, riluzole in this case): imputation_riluzole.png
• save imputed dataset object: imputed_rev.RDS

Feature selection

Prediction