HOSPITAL READMISSION

---

The primary objective of this project is to predict hospital readmission using data for hospital admission. The secondary objective is to evaluate whether an initial diagnosis of diabetes can serve as a reliable predictor for hospital readmission.

The data contains patient information - age, initial duration at the hospital, tests, diagnosis (primary, secondary and tertiary) and readmission status. Given the target variable is categorical (yes/no) for readmission, two classification models were employed: Logistic Regression and Neural Networks.

For the Logistic Regression model, due to a high number of predictors, feature selection

techniques were applied to enhance model performance, like Backward Elimination, Forward Selection, and Stepwise Selection. While the Exhaustive Search method was excluded due to its high computational cost, a Grid Search was implemented to optimize the neural network model by identifying the most effective number of hidden layer nodes for improved performance.

For the secondary objective, to assess if diabetes can be used to predict hospital readmission, a graphical presentation was used to compare the results to other diagnoses to aid the data analyst to understand the effect to readmission; while the coefficients were used to finalize the conclusion. The application of the model shall help in better hospital operations - readmissions are costly and often indicative of gaps in care delivery. High readmission rates are associated with increased

healthcare expenses and poor patient outcomes. Predicting which patients are at high risk of readmission can enable healthcare providers to deliver preventative interventions and reduce overall readmission rates.

Hospital readmission rates are widely recognized as a critical indicator of healthcare quality, efficiency, and patient outcomes. Elevated readmission rates often point to underlying systemic issues, such as insufficient discharge planning, inadequate patient education, or a lack of proper follow-up care. In the context of a healthcare industry moving steadily toward value-based care models, reducing avoidable readmissions has become a key priority. Addressing this issue not only improves the quality of care delivered to patients but also helps healthcare providers avoid penalties, reduce operational costs, and enhance long-term patient satisfaction and health outcomes.

Creating a comprehensive data analysis can significantly aid in identifying the factors contributing to readmissions, allowing healthcare providers to implement targeted interventions. This proactive approach not only benefits patients but also aligns with industry standards aimed at fostering accountability and excellence in healthcare delivery.

By leveraging the Kaggle hospital readmissions dataset, we were able to simulate real-world

hospital scenarios and develop predictive models that offer meaningful insights into readmission patterns.

A. Develop the Understanding

• The primary objective of this project is to predict hospital readmission using data for hospital admission.
• The second objective is to determine if patients with diabetes can be used to predict hospital readmission.
• Once the model is created, it can be used to identify admitted patients' likelihood of readmission, and they may use it to improve hospital services to reduce it.
• One of the outputs of the project is to identify diabetes as a criterion of readmission, if this is determined, hospitals can adjust patient care targeting diabetes patients improved health care services.

B. Obtain Data for Analysis

To conduct our analysis on hospital readmissions, we first identified and sourced a publicly available dataset that provides relevant and comprehensive healthcare data. After exploring various repositories, we selected a dataset from Kaggle titled "hospital Readmissions".

This dataset includes detailed patient-level information from over 100,000 hospital encounters for diabetic patients. It contains variables related to demographics, admission and discharge details, medical diagnoses, length of stay, medications, and whether the patient was readmitted. Specifically, it allows us to investigate factors influencing hospital readmission, making it suitable for predictive modeling and exploratory analysis in our case.

C. Explore, Clean and Preprocess Data / Reduce the Data Dimension

1. Number of Records: 25,000

2. Number of Columns: 17

• Hospital Readmissions Dimension: (25000, 17)

3. Columns 

 Index(['age', 'time_in_hospital', 'n_lab_procedures', 'n_procedures','n_medications', 'n_outpatient', 'n_inpatient', 'n_emergency','medical_specialty', 'diag_1', 'diag_2', 'diag_3', 'glucose_test','A1Ctest', 'change', 'diabetes_med', 'readmitted'],dtype='object') 

4.  Column Data Types 

age object

time_in_hospital int64

n_lab_procedures int64 

n_procedures int64 

n_medications int64

n_outpatie
nt int64 

n_inpatient int64 

n_emergency int64 

medical_specialty object 

diag_1 object 

diag_2 object 

diag_3 object 

glucose_test object 

A1Ctest object 

change object 

diabetes_med object 

readmitted object 

dtype: object 

 5. Converted Column Data Types 

 The fields that are non-numeric are converted to dummy variables to be used in the analysis. The result is 46 columns. 

time_in_hospital int32 

n_lab_procedures int32 

n_procedures int32 

n_medications int32 

n_outpatient int32 

n_inpatient int32 

n_emergency int32 

age_[50-60) int32 

age_[60-70) int32 

age_[70-80) int32 

age_[80-90) int32 

age_[90-100) int32

medical_specialty_Emergency/Trauma int32 medical_specialty_Family/GeneralPractice int32 

medical_specialty_InternalMedicine int32 

medical_specialty_Missing int32 

medical_specialty_Other int32 

medical_specialty_Surgery int32 

diag_1_Diabetes int32 

diag_1_Digestive int32 

diag_1_Injury int32 

diag_1_Missing int32 

diag_1_Musculoskeletal int32 

diag_1_Other int32 

diag_1_Respiratory int32 

diag_2_Diabetes int32 

diag_2_Digestive int32 

diag_2_Injury int32 

diag_2_Missing int32 

diag_2_Musculoskeletal int32 

diag_2_Other int32 

diag_2_Respiratory int32 

diag_3_Diabetes int32 

diag_3_Digestive int32 

diag_3_Injury int32 

diag_3_Missing int32 

diag_3_Musculoskeletal int32 

diag_3_Other int32 

diag_3_Respiratory int32 

glucose_test_no int32 

glucose_test_normal int32 

A1Ctest_no int32 

A1Ctest_normal int32 

change_yes int32 

diabetes_med_yes int32 

readmitted_yes int32 

dtype: object 

 D. Determine the Data Mining Task 

• The identified output column is “readmitted” with two outputs “Yes” or “No.” From “readmitted” it was converted to “readmitted_Yes” with either “0” or “1” value. 

• The initial predictor variables are listed below: 

Index(['time_in_hospital', 'n_lab_procedures', 'n_procedures', 'n_medications',  'n_outpatient', 'n_inpatient', 'n_emergency', 'age_[50-60)',  'age_[60-70)', 'age_[70-80)', 'age_[80-90)', 'age_[90-100)',  'medical_specialty_Emergency/Trauma',  'medical_specialty_Family/GeneralPractice',  'medical_specialty_InternalMedicine', 'medical_specialty_Missing',  'medical_specialty_Other', 'medical_specialty_Surgery',  'diag_1_Diabetes', 'diag_1_Digestive', 'diag_1_Injury',  'diag_1_Missing', 'diag_1_Musculoskeletal', 'diag_1_Other',  'diag_1_Respiratory', 'diag_2_Diabetes', 'diag_2_Digestive',  'diag_2_Injury', 'diag_2_Missing', 'diag_2_Musculoskeletal',  'diag_2_Other', 'diag_2_Respiratory', 'diag_3_Diabetes',  'diag_3_Digestive', 'diag_3_Injury', 'diag_3_Missing',  'diag_3_Musculoskeletal', 'diag_3_Other', 'diag_3_Respiratory',  'glucose_test_no', 'glucose_test_normal', 'A1Ctest_no',  'A1Ctest_normal', 'change_yes', 'diabetes_med_yes'],  dtype='object') 

E. Partition the Data 

To ensure there is no overfitting, initial data was partitioned into 60% Training data, and 40% Validation data using the train_test_split function. Other models was adjusted accordingly - 80%/20% split, and 70%/30% split 

F. Techniques 

1. Since the identified output variable is categorical rather than numeric, a classification model was used. It was identified to use Logistic Regression as the initial model, and the Neural Net was selected to ensure complicated relationships can be covered.  
2. Logistic Regression - a method of reduction through Forward Elimination, Backward Elimination and Stepwise was used to improve the accuracy of the model. 
3.  Neural Nets - to refine the model, GridSearch was used to better improve the result. 

G. Charts, Algorithm and Measures 

LOGISTIC REGRESSION MODEL 

Bar Chart for Data Visualization of Diagnosis Readmission 

• All models' respective confusion matrix results indicate no overfitting (the training and validation result is close to each other)

• Logistic Regression comparison - Model #1 to Model #5 provided almost the same results of approximately 60%. The Backward Selection, Forward Selection and Stepwise reduction method provided no significant improvement.

• Neural Nets - initial model performed better even if the second model was enhanced using the Gridsearch.