Feature Engineering

[data-colorid=mr5bz5wdao]{color:#bf2600} html[data-color-mode=dark] [data-colorid=mr5bz5wdao]{color:#ff6640}[data-colorid=qilzv6b9qc]{color:#ff5630} html[data-color-mode=dark] [data-colorid=qilzv6b9qc]{color:#cf2600}[data-colorid=hisiwhv6cc]{color:#bf2600} html[data-color-mode=dark] [data-colorid=hisiwhv6cc]{color:#ff6640}[data-colorid=bh1towvo59]{color:#bf2600} html[data-color-mode=dark] [data-colorid=bh1towvo59]{color:#ff6640}[data-colorid=zb0xpg0375]{color:#ff5630} html[data-color-mode=dark] [data-colorid=zb0xpg0375]{color:#cf2600}[data-colorid=z31xsyttsu]{color:#ff5630} html[data-color-mode=dark] [data-colorid=z31xsyttsu]{color:#cf2600}[data-colorid=xnzgmg3vbw]{color:#ff5630} html[data-color-mode=dark] [data-colorid=xnzgmg3vbw]{color:#cf2600}[data-colorid=mcumx5ysfw]{color:#ff5630} html[data-color-mode=dark] [data-colorid=mcumx5ysfw]{color:#cf2600}[data-colorid=rb7jg0wu4u]{color:#ff5630} html[data-color-mode=dark] [data-colorid=rb7jg0wu4u]{color:#cf2600}[data-colorid=faaqfxh949]{color:#bf2600} html[data-color-mode=dark] [data-colorid=faaqfxh949]{color:#ff6640}[data-colorid=zi1jtfpncn]{color:#ff5630} html[data-color-mode=dark] [data-colorid=zi1jtfpncn]{color:#cf2600}[data-colorid=abjqq4wub2]{color:#ff5630} html[data-color-mode=dark] [data-colorid=abjqq4wub2]{color:#cf2600}[data-colorid=eu6av933ww]{color:#bf2600} html[data-color-mode=dark] [data-colorid=eu6av933ww]{color:#ff6640}[data-colorid=s78qiewozi]{color:#bf2600} html[data-color-mode=dark] [data-colorid=s78qiewozi]{color:#ff6640}[data-colorid=uaqeoynofg]{color:#ff5630} html[data-color-mode=dark] [data-colorid=uaqeoynofg]{color:#cf2600}[data-colorid=z9d8osui3r]{color:#ff5630} html[data-color-mode=dark] [data-colorid=z9d8osui3r]{color:#cf2600}[data-colorid=nthibw39rq]{color:#ff5630} html[data-color-mode=dark] [data-colorid=nthibw39rq]{color:#cf2600}[data-colorid=th1tatf7r9]{color:#ff5630} html[data-color-mode=dark] [data-colorid=th1tatf7r9]{color:#cf2600}[data-colorid=x9kalryowd]{color:#ff5630} html[data-color-mode=dark] [data-colorid=x9kalryowd]{color:#cf2600}

EDA: https://docs.google.com/spreadsheets/d/1CxsfVNbub2SrAsxHV5_0btlO-qTjPww0/edit?usp=drive_link&rtpof=true&sd=true

//

• Difference between Train and Test
• Null Percentage
  • # Feature Correlation Matrix
• PCA

Difference between Train and Test

CV Name	Train Percentage of Null Values	Test Percentage of Null Values	Difference
CV_GFR_latest	76.5	3.4	73.1
CV_MAP_cuff_mean	38.8	18.3	20.5
CV_MAP_cuff_max	38.8	18.3	20.5
CV_MAP_cuff_min	38.8	18.3	20.5
CV_MAP_cuff_latest	27.9	12.3	15.6

• Drop these because it has a high difference between train and test dataset

Null Percentage

1. Drop temperature_skin and troponin based on high null %

CV	Train Percentage of Null Values	Test Percentage of Null Values	Difference
CV_temperature_skin_latest	99.8	99.9	-0.1
CV_troponin_HS_latest	97.8	95.4	2.4
CV_temperature_bladder_latest	85.2	86.3	-1.1
CV_PO2_art_istat_latest	85.1	88.3	-3.2
CV_TCO2_art_latest	85.1	88.3	-3.2
CV_PCO2_art_latest	85.1	88.2	-3.1
CV_pH_art_istat_latest	85.1	88.2	-3.1
CV_HCO3_latest	85.1	88.1	-3
CV_FiO2_latest	85.1	88.1	-3
CV_base_excess_istat_latest	85.1	88.1	-3
CV_MAP_invasive_latest	84.7	86.8	-2.1

# Feature Correlation Matrix

1. Manually look at highly correlated features (defined by abs(correlation) >= 0.7)

   • Used Spearman correlation (instead of Pearson) because relationship may not be linear.

2. To Do

   • Remove time series feature engineering for GCS components (motor_response, verbal_response, eye_opening)

     • The min, max, mean, latest tend to be highly correlated
     • However, I think it’s a good idea to keep min

   • Remove GCS as a feature

     • Since we have the individual components – having GCS does not add more information

   • Other highly correlated

Features	Correlation	Action
GFR Creatinine	-0.79	*Drop GFR (higher null rate)* Keep Creatinine
HCO3, TCO2, Base Excess	0.99 and 0.96	Drop HCO3 and TCO2 Keep Base Excess (highest scaled variance)
INR, prothrombin	0.99	Drop prothrombin (lower scaled variance) Keep INR
systolic_blood_pressure, MAP	0.75	Keep both

• Basically remove things 0.9 and higher

PCA

• Rank each feature by importance and explained variance
• Interestingly enough, many of the “mean” time series ones low variance and low importance

• We may consider dropping these