Machine learning methods to predict 30-day hospital readmission outcome among US adults with pneumonia: analysis of the national readmission database

Figures

• 

  Fig. 1. Top nine rules for the rule-based model. Notes These plots depict the rules in the final ensemble as a binary tree structure. We chose to present the nine most important rules in the fitted ensemble. This could be obtained using the inbuilt plot method in the PRE algorithm. (1) The risk factors involved in each rule and their cut-off values are automatically determined by the rule-based model. The involved variables and cut-point employed by each rule are highlighted in bolden black: these individuals endorsing these criteria will follow the path of the prediction rule while those individuals not meeting the criteria instead are assigned as zero in the square box. (2) For example, 1st important rule indicates that those satisfying (1) disposition location ≤ 6, (2) metastatic cancer = 0, (3) leukemia = 0, (4) solid tumor = 0, are less likely to have 30-day readmission outcome. Abbreviations and descriptions of data element name: (1) DISPUNIFORMrr: Disposition of patient; (2) CANCER_METS: Elixhauser comorbdity measure of metastatic cancer; (3) CANCER_LEUK: Elixhauser comorbdity measure of leukemia; (4) CANCER_SOLID: Elixhauser comorbdity measure of solid tumor without metastasis,…

• 

  Fig. 2. The individual variable importance for the rule-based model. Notes The individual variable contributing to the outcome is defined as the sum of the importance of the linear term and the importance of every rule in which the variable appears divided by the total number of conditions in the rule. The 12 most important input variables for predicting the readmission outcome are (1) I10_NDX: Number of ICD-10-CM diagnoses coded on the record; (1) APRDRG_Severity: Severity of Illness Subclass based on 3M All Patient Refined DRG; (2) DISPUNIFORMrr: Disposition of patient; (3) LOS: Length of Stay; (4) mortal_score: Elixhauser Comorbidity Index Score; (5) APRDRG_Severity: All Patient Refined DRG: Severity of Illness Subclass; (6) AGE: the age at admission; (7) Resident: patient as a resident of the State in which he or she received hospital care; (8) HTN_CX: Elixhauser comorbidity measure of Hypertension, complicated; (9) pay1r: Expected primary payer; (10) I10_NPR: number of ICD-10-PCS procedures on this discharge; (11) CANCER_METS: Elixhauser comorbidity measure of Metastatic cancer; and (12) ZIPINC_QRTL: Median household income for patient's ZIP Code (based on current year)

• 

  Fig. 3. The top 10 most important predictive factors for the random forest. Notes A higher "Mean Decrease in Gini" in the x-axis indicates a higher purity (less noise, less bias) contributed by the variable, and higher variable importance). 1Variable importance is reported as two measures: it is based on the mean decrease of accuracy in prediction on the out of bag sampled when a given variable is excluded; it is also computed using the mean decrease in Gini index. The variable importance for each variable is expressed relative to the largest. Abbreviations and descriptions of data element name: (1) APRDRG_Severity: Severity of Illness Subclass based on 3M All Patient Refined DRG; (2) mortal_score: Elixhauser Comorbidity Index Score for in-hospital mortality; (3) I10-NDX: Number of ICD-10-CM diagnoses coded on the record; (4) APRDRG_Risk_Mortality: Risk of Mortality Subclass based on 3M All Patient Refined DRG; (5) DISPUNIFORMrr: Disposition of patient; (6) AGE: the age at admission; (7) I10_NPR: Number of procedures coded; (8) LOS: Length of Stay (9) CANCER_METS: Elixhauser comorbidity measure of metastatic cancer; (10) HTN_CX: hypertension, complicated; (11) PL_NCHS: Patient location…

• 

  Fig. 4. The most important predictor variables in the XGBoost model1. Notes 1The bar graph represents the 10 most important predictor variables in the gradient boosting model sorted by importance. Features are shown ranked in decreasing importance order. The numbers represent the relative importance of each variable. The feature importance is calculated by the feature’s importance contribution relative to the most important feature. The variable importance measure uses the mean decrease in the Gini index to determine the contribution of each predictor variable. Abbreviations and descriptions of data element name: (1) APRDRG_Severity: Severity of Illness Subclass based on 3M All Patient Refined DRG; (2) I10-NDX: Number of ICD-10-CM diagnoses coded on the record; (3) mortal_score: Elixhauser Comorbidity Index Score; (4) AGE: the age at admission; (5) DISPUNIFORMrr: Disposition of patient; (6) I10_NPR: Number of procedures coded; (7) APRDRG_Risk_Mortality: Risk of Mortality Subclass based on 3M All Patient Refined DRG; (8) LOS: Length of Stay; (9) HTN_CX: Elixhauser comorbidity measure of complicated hypertension; (10) pay1r: Expected primary payer