Sepsis

Transcript

1. Automating Sepsis Diagnosis Stephen Thomas
 BMED 6517 - Spring 2019
2. 40 features 8 vital signs 26 lab results 6 dem ographic values 40,336 patients sepsis label 70% missing data 2932 sepsis positive patients 8 to 336 hourly measurements
3. Data Preparation 40 features 8 vital signs 26 lab results 6 dem ographic values 40,336 patients sepsis label 70% missing data 2932 sepsis positive patients 8 to 336 hourly measurements 12,237 patients 28,099 patients 2064 sepsis positive patients 30% Holdout For Testing Impute Temporal Data by Linear Interpolation/ Extrapolation Balance Training Data by Undersampling Augment Temporal Data with Rates of Change Normalize and Zero-Fill 34 tem poral variables velocity acceleration Δ(X) Δ²(X) Training Set: 199,686 observations × 108 features Test Set: 469,332 observations × 108 features 19,669 sepsis positive labels
4. 30% Holdout 40 features 8 vital signs 26 lab results 6 dem ographic values 40,336 patients sepsis label 70% missing data 2932 sepsis positive patients 8 to 336 hourly measurements 12,237 patients 28,099 patients 2064 sepsis positive patients
5. Impute Temporal Data 12,237 patients 28,099 patients 2064 sepsis positive patients Linear Interpolation & Extrapolation
6. Balance Training Data Undersample sepsis negative 26,035 → 2064
7. Capture Temporal History 34 tem poral variables velocity acceleration Δ(X) Δ²(X)
8. Normalize / Zero-Fill Training Set: 199,686 observations × 108 features Test Set: 469,332 observations × 108 features 19,669 sepsis positive labels
9. Is Data Separable?
10. Is Data Separable?
11. LSTM Network Results Sepsis Negative Sepsis Positive Predicted Class True Class 297 1172 571 10197 0 0.2 0.4 0.6 0.8 1 False prediction rate 0 0.2 0.4 0.6 0.8 1 True prediction rate Sepsis Negative Sepsis Positive
12. Results • All models very sensitive to overfitting • Problem is worthy of computational challenge Utility Notes Support Vector Machine 0.00 Gaussian kernel, other hyper-parameters had minimal effect Random Forest 0.02 Large leaf size and minimal splits to minimize overfitting Long Short-Term Memory Recurrent Neural Network 0.35 Limited epochs and relatively high learning rate to minimize overfitting