Report

Predicting Risk of Re-hospitalization for Congestive Heart Failure Patients (in collaboration with ) Jayshree Agarwal Senjuti Basu Roy, Ankur Teredesai, Si-Chi Chin, David Hazel, Kiyana, Mehrdad, (UWT) Paul Amoroso, Yoshi Williams, Dr. Lester Reed, Sheila, Eric Johnson (MHS) Motivation 19.6% patients readmitted within 30 days [Jencks et al. 2009] 31.1% patients readmitted within 60 days [Jencks et al. 2009] Many hospitalizations readmissions $$$COST - 2004 unplanned re-admits = $17.4 billion [Jencks et al. 2009] Congestive Heart Failure(CHF) LOW Readmission rate = HIGH quality of care by hospital No reimbursement for readmission within 30 days 2 MHS - UWT Web and Data Science collaboration objectives Predict the RISK of Readmission for CHF patients Reduce the Readmission rate and cost Improve patient satisfaction and quality of care Appropriate pre-discharge and post-discharge planning Proper resource utilization 3 Problem Develop models that can predict risk of readmission for CHF patients within 30 days after discharge 60 days after discharge The readmission may happen for other reasons in addition to CHF 5 Overall Approach How to solve the problem? – Apply predictive data mining techniques such as, classification What do these predictive mining techniques require? – Data in homogeneous format • Information Extraction, Integration, and data preparation • Prepare labeled dataset to train the model; used later on for testing. 6 Our Challenges Building domain knowledge – Which variables to consider? – How to merge and unify them in a homogeneous format (information extraction and integration) – How to understand the relative importance of the variables in the prediction task? How to prepare data? – Class label generation – Noisy real world data (missing values, inconsistencies, etc.) – Serious skew in the dataset 7 Solution 8 Building Predictive Classification Models Data Understanding Data Preprocessing Modeling Evaluation 9 Data Understanding Collect initial data Acquire Domain knowledge Describe and explore dataset Create data visualization 10 Building Predictive Classification Models Data Understanding Data Preprocessing Modeling Evaluation 11 Data Preprocessing Finding Eligible CHF admissions Define class label Attribute selection Data Integration Removal of incomplete data 12 Eligible CHF admissions and Generating Class Labels All CHF Admissions In hospital deaths removed Eligible CHF Admissions X=30 X=60 The class label is assigned as 0 NO Is there any readmission within x days of discharge? YES The class label is assigned as 1 13 Attribute selection Yale Model [Krumholz et al] -Socio-Demographic variable(2) -Comorbidities(35) Chi-square correlation test “Baseline” “All” “Correlated” Additional predictor variables identified by us (14) “New” 14 Data Extraction Labeled data Patient details Primary and Secondary diagnosis Table Joins Incomplete data removed Data Data used for training the Models Lab measurement Administrative data 15 Data Distribution 30 days time frame 60 days time frame 12000 12000 10000 10000 8000 8000 Readmit 6000 Readmit 6000 No Readmit No Readmit 4000 4000 2000 2000 0 0 Readmissions Readmissions 16 Building Predictive Classification Models Data Understanding Data Preprocessing Modeling Evaluation 17 Modeling Balancing imbalanced data by under-sampling and over sampling Selecting modeling technique for Binary Classification Building prediction models • Logistic regression • Naïve Bayes classifier • Support Vector Machine 18 Logistic Regression Model = = 0 + 1 1 + 2 2 + ⋯ + 1 1+ − 0 +1 1 +⋯+ = 1 1+ − P (Probability of Y) ln 1− Z ------> 19 Naïve Bayesian Classification Statistical Classifier performs probabilistic prediction based on Bayes Theorem Assumes that the attributes are conditionally independent Given a data tuple X and m classes 1 , 2 , … ) = () Predicts X belongs to only if is highest among all the for all the m classes 20 Support Vector Machine A method of classification for both linear and non linear data Searches for optimal separating hyperplane separating the two classes 21 Building Predictive Classification Models Data Understanding Data Preprocessing Modeling Evaluation 22 Performance Evaluation Metrics Precision – percentage of tuples labeled as positive are actually positive = TP/TP+FP Recall – measures the percentage of positive tuples that are labeled positive = TP/TP+FN Accuracy – percentage of tuples correctly classified = (TP+TN)/P+N ROC curves and area under the curve (AUC) – Shows the trade-off between true positive rate and false positive rate. 23 Evaluation • Predictive models are assessed using 10 fold cross validation • The performance is compared using different evaluation metrics mentioned previously 25 RESULTS Logistic Regression for 30 days Area Under the Curve (AUC) Recall 27 Logistic regression for 60 days Area Under the Curve (AUC) Recall 28 Naïve Bayes classifier for 30 days 0.64 0.63 0.62 0.61 Baseline New 0.6 All 0.59 Correlated 0.58 0.57 0.56 Attribute Set Area Under the Curve (AUC) 29 Support Vector Machine for 30 days 0.635 0.63 0.625 0.62 0.615 Baseline 0.61 New 0.605 All 0.6 Correlated 0.595 0.59 0.585 0.58 Attribute Set Area Under the Curve (AUC) 30 Conclusion and Discussion It is one of the difficult problem to solve Feature selection gives the best results. With data balancing recall of the model improves 35 Future Work Investigate other classifier techniques like ensemble methods, neural networks To explore additional features and study their relevance To employ other feature selection techniques To device a method to impute missing values Deploying the predictive models 36 Acknowledgement Multicare health System (MHS) and Dr. Lester Reed for giving us this opportunity Data architects and domain experts in MHS for their inputs Professors Dr. Ankur Teredesai and Dr. Senjuti Basu Roy for their guidance Other team members in UWT for their support 37 References S. F. Jencks, M. V. Williams, and E. A. Coleman, “Rehospitalizations among Patients in the Medicare Fee-for-Service Program,” New England Journal of Medicine, vol. 360, no. 14, pp. 1418–1428, 2009. J. Han and M. Kamber, Data mining: concepts and techniques. Morgan Kaufmann, 2006 H. M. Krumholz, S. L. T. Normand, P. S. Keenan, Z. Q. Lin, E. E. Drye, K. R. Bhat, Y. F. Wang, J. S. Ross, J. D. Schuur, and B. D. Stauffer, Hospital 30-day heart failure readmission measure methodology. Report prepared for the Centers for Medicare & Medicaid Services. 38 Questions 39