Predicting In-hospital Mortality of Patients with Febrile Neutropenia Using Machine Learning Models

Several commonly used machine leanring models, as well as deep neural network, were built upon Healthcare Cost and Utilization Project (HCUP)’s National Inpatient Sample (NIS) data to predict the mortality of adult cancer febrile neutropenia admissions. Machine learning models achieved areas under the receiver operating characteristic (AUROC) above 90%.

Variable importance analysis was also conducted with ridge logistic regression and gradient boosting tree. The variable importance analysis results consistent with previous studies.

• You are welcome to re-use any code from this repo for your project, please cite this paper: https://www.sciencedirect.com/science/article/pii/S1386505619314078?via%3Dihub.
• Please refer to our wiki for more information regarding how to use our code to do different things related to HCUP data

---

Requirements:

• Python 2.7
• Python 3.6
• Pandas 0.21.1
• Numpy 1.14.5
• Pyhcup 0.1.6.4
• Scikit-learn 0.19.2
• Pandas-summary 1.11.0
• Scipy 1.14.5

---

Workflow

Datasets:

• Healthcare Cost and Utilization Project (HCUP)’s National Inpatient Sample (NIS) from 2007 to 2015 third quarter.

---

Experiment Procedure (Use python 3 if not specified):

Descriptive statistics, trend analysis and data preparation

• Run src/sas2csv.py with Python 2 to convert ASCII files to csv. Use the following code to convert the ASCII files of core, hospital, severity and diagnose/procedure group of each year to csv.

python src/sas2csv.py -d {path of ASC file} -l {path of corresponding SAS load file} -o {path of the directory of output csv file} -y {year of data}

• Use nbs/files_combination_each_year.ipynb to merge core, hospital, severity and diagnose/procedure group files of each year to one file.
• Run through nbs/train_test_generator.ipynb to get the number of cancer, febrile neutropenia (FN) and FN death of each year, and generate a csv file including all adult cancer FN admissions of each year.
• Run nbs/merge_all_year_data.ipynb to merge all years' adult cancer FN admissions to one file, which is for machine learning models training and evaluation.
• Run nbs/combined_inputs_generator.ipynb to do one hot encoding for diagnose/procedure variables.
• Run nbs/missing_data_imputation.ipynb to do data imputation (using mean values for continuous variables, most common values for categorical variables)
• Use nbs/desctiptive_statistics.ipynb to get statistics of selected data. P-values for proportions were calculated with z score calculator online; p-values for continuous variables were calculated with python.

Sample Selection Bias Test

• Use nbs/file_preparison_for_selection_bias_test.ipynb to generate files containing samples with age value and without age value separately.
• Use nbs/train_test_generator.ipynb and nbs/merge_all_year_data.ipynb to get adult FN samples with missing died information and without.
• Use nbs/Selection_bias_test.ipynb to test whether removing samples with missing age/died value will introduce bias or not.
• Consort diagram:

Diagnose/procedure ontology evaluation

• Use random forest as the learning algorithm. Run nbs/dx_pr_selection.ipynb to test predictive power of different diagnose/procedure coding systems.

Model training, evaluation and predictor analysis

• nbs/Model_Development.ipynb

More

• Please post your question in issue tracker