American Journal of Engineering and Applied Sciences Original Research Paper Evaluating Patient Readmission Risk: A Predictive Analytics Approach 1,2,3 Avishek Choudhury and 4Dr. Christopher M. Greene 1 Applied Data Science, Syracuse University, New York, USA Process Improvement, UnityPoint Health, Iowa, USA 3 Systems Engineering, Stevens Institute of Technology, New Jersey, USA 4 Systems Science and Industrial Engineering, Binghamton University, NY, USA 2 Article history Received: 31-10-2018 Revised: 26-11-2018 Accepted: 06-12-2018 Corresponding Author: Avishek Choudhury Applied Data Science, Syracuse University, New York, USA Tell: +1 (515) 608-0777 Email: achoud02@syr.edu Achoudh7@stevens.edu Abstract: With the emergence of the Hospital Readmission Reduction Program of the Center for Medicare and Medicaid Services on October 1, 2012, forecasting unplanned patient readmission risk became crucial to the healthcare domain. There are tangible works in the literature emphasizing on developing readmission risk prediction models; However, the models are not accurate enough to be deployed in an actual clinical setting. Our study considers patient readmission risk as the objective for optimization and develops a useful risk prediction model to address unplanned readmissions. Furthermore, Genetic Algorithm and Greedy Ensemble is used to optimize the developed model constraints. Keywords: Prediction Model, Patient Readmission Risk, Healthcare Expenses, Healthcare Quality, Optimization Model Introduction It is a fact that the federal budget of the United States is concerned by the burgeoning healthcare expenses (Shipeng Yua, 2015). One of the main factors contributing to the healthcare cost is the avoidable patient readmission. Unplanned patient readmission has been a significant measure of care quality. However, the Affordable Care Act of 2010 introduced the Readmission Reduction Program which became effective on October 1, 2012. According to the School of Public Health, Veterans Administration can save $2,140 per patient by managing patients prone to readmission (Kathleen Carey, 2016). Moreover, studies have shown that 15 to 25 percent of discharged patients are readmitted in less than 30 days. According to the Agency for Healthcare Research and Quality, about 1.8 million patients were readmitted (Anika and Hines, 2014). Fierce Healthcare reported that in 2011, hospitals spend $41.3 billion to treat unplanned readmitted patients (Shinkman, 2014). A study published by Harvard Business Review stated that prioritized and effective communication with the patient and complying to evidence-based care standards could check patient readmission rate by 5 percent (Claire Senot, 2015). However, fostering desired communication within a hospital is arduous due to the complexity of the system. Our study focuses on predicting patient readmission. Individuals with a high risk of readmission can be provided with alternative preventive measures such as intensive post-discharge care or home care (Davood Golmohammadi, 2015). We define patient readmission as the readmission caused due to poor discharge planning resulting in reoccurrence of the treated disease and worsening health condition. When an individual requires readmission within 90 days’ post-discharge for the same cause for which she or he was admitted to a hospital in the very first place is termed as the patient readmission. The reason behind considering readmission within 90 days is since the patients during the first three-month post-discharge are susceptible to the diseases and have suicidal behavior among individuals who have a mental disorder (Appleby, 2013). Alarming Hospital Discharge Concerns This section classifies the three most crucial poor patient discharge issues that encourages patient readmission. Early Patient Discharge The fundamental decision healthcare providers need to take is whether an individual has recovered enough to leave the hospital independently. Poor decision making at this instance hinders patient safety, resulting in emergency © 2018 Avishek Choudhury and Dr. Christopher M. Greene. This open access article is distributed under a Creative Commons Attribution (CC-BY) 3.0 license. Avishek Choudhury and Dr. Christopher M. Greene / American Journal of Engineering and Applied Sciences 2018, 11 (4): 1320.1331 DOI: 10.3844/ajeassp.2018.1320.1331 readmission or sometimes death. “A man died after a hospital failed to treat sepsis” and discharge the patient before time (Ombudsman, 2003). According to Homeless Link, more than 70% of underprivileged people were discharged without any housing and addressing underlying health conditions (IHSMHL, 2012). 2005). The literature search process comprised of the following three steps: (a) Systematic literature search using electronic database search, ‘snowballing’ (Greenhalgh, 2005), (b) identify relevant papers based on their title and (c) article selection based on their abstract. Poor Patient Assessment and Consulting Prior Discharge The literature survey was limited to the following database: ACM Digital Library, ASME Digital Collection, BIOSIS Citation Index, CINDAS Microelectronics Packaging Material Database, CiteSeer, Computer Database, Emerald Library, Energy and Power Source, Engineering Village, IEEE Xplore, MEDLINE, OSA Publishing, PubMed, Safari Books Online, ScienceDirect, Sci-Finder, SPIE Digital Library and Springer. The search conducted had no constraint of time zone and the following material type was considered: Articles, Newspaper Articles, Dissertations, Conference Proceedings, Statistics Data Sets, Technical Reports and Websites. The search keywords used were “patient readmission,” “readmission” risk, “readmission survey,” “readmission prediction” and “prediction models.” All the papers that contained any of these words anywhere in the article were selected. Then based on their title, 104 papers where shortlisted. Finally, after reading the abstract 33 peers reviewed articles were finalized as the reference for this study. Often patients physically fit enough are not mentally capable of coping at home. These patients after discharge often fail to continue medications and lose mental health which in turn enhances the plausibility of readmission. Such conditions are common among older adults who are not capable of independently maintaining their health either due to cognitive or financial constraints. According to King’s Fund, “being discharged without proper support is an invitation to relapse, worsening of the condition and readmission” (Maguire, 2015). Absence of Home Care Plans Insufficient communication and coordination between hospitals and community healthcare providers is another concern that needs attention. Due to insufficient domestic healthcare facilities, discharged patients with health care requirements are left alone at home which leaves the patient susceptible to health deterioration and emergency readmission. During 2002 and 2012, 3,225 suicides were recorded by The National Confidential Inquiry into Suicide and Homicide by People with Mental Illness, 2014. To minimize such occurrences, NHS recommends hospitals to follow up with their discharged patients within 7-2002 1day post discharge and ensure availability of crisis support (Assessment, 2013). Problem Statement There exist several possible causes responsible for unplanned patient readmission. However, our study does not focus on identifying the responsible cause, but it provides with an efficient prediction model that can be deployed to a clinical scenario and help healthcare units to be prepared for the unavoidable readmissions and provide alternative care to preventable readmissions. The proposed model provides healthcare providers with a decision support system to identify individuals prone to readmission and thus minimize early discharge and ensure follow up with the discharge patients. Systematic Literature Review Design The study conducts a systematic review of methods and models used in predicting unplanned patient readmission to meet our research motive (Dixon Woods, Search Strategy and Inclusion Criteria Findings From the Literature Review Miller et al. (1984) used multiple regression to develop a five-year prediction model for patient readmission. This paper was an indication that multiple regression for predictor variable analysis was a viable option. Hodgson et al. (2001) estimated readmission rates for all psychiatric admissions in North Staffordshire. Survival analysis was used to find the ones that predicted readmission. It used the Survival Analysis Log-Rank Test and Cox Regression for this purpose. Betihavas et al. (2012) extended the scope of the study done. It included non-clinical factors as well such as social instability. It also called out the need for predictive analysis and the lack of such tools for clinicians to use in risk assessment of readmission. Allison (2012) studied variables that could potentially predict readmission chances for patients previously admitted in pulmonary rehabilitation. Apart from eliminating excessive pain and unnecessary illness, it sought to reduce health care cost, in general, using discriminant and predictive analysis. Zheng et al. (2015) used metaheuristic and data mining approaches. A neural networks algorithm and SVM classifiers were used. These models could perform risk measure with higher sensitivity and F measure. Bakal et al. (2014) generated a prediction as well as risk evaluation model for the rehospitalization of Heart Failure (HF) patients. It was found using five years of follow up half the patients had returned for 1321 Avishek Choudhury and Dr. Christopher M. Greene / American Journal of Engineering and Applied Sciences 2018, 11 (4): 1320.1331 DOI: 10.3844/ajeassp.2018.1320.1331 rehospitalization. It concluded that trying to elongate the gap between hospitalizations should be an essential goal for evaluating the quality of treatment. Ajorlou et al. (2014) proposed a risk prediction model based on hierarchical nonlinear mixed effect to recognize patients with high likelihood of discharging, non-compliances to decrease Medicare costs and improve quality of care provided by hospitals. It applied stepwise variable selection in the mixed-effect framework and extended the (typical) random frailty model for Weibull hazard function with incorporated patient factors. Wang et al. (2014) validated the use of the LACE index when studying readmission risk of patients with CHF. Bayati et al. (2014) evaluated the cost-effectiveness and efficiency of the methodology that combines prediction and decision making. Machine learning classifiers were used with the patient data to perform the cost analysis. Inouye et al. (2015) was one of the few studies that used patient self-reports as means for risk assessment of readmission. An automated multicall follows up system was implemented. Amarasingham et al. (2015) in the same year focused on Electronic Medical Record (EMR) models to access readmission. Kang et al. (2015) used Retrospective Analysis and multivariate analysis to determine readmission risk factors. Futoma et al. (2015) compared several predictive models for predicting early readmissions. Deep learning is used to analyze the five conditions that CMS uses to penalize hospitals. It used Logistic Regression, Penalized Logistic Regression, Random Forests and Support Vector Machines and Neural Network deep learning methods. A framework for assessing patient readmission risk was developed. It found random forests, penalized regressions and deep neural networks to be the best predictors. Shams et al. (2015) developed a new metric for evaluating possible avoidable readmission. A tree-based classification method was proposed that factored in the previous history of the patient’s readmissions and the various risk factors that were identified by the researchers of the paper. Pack et al. (2016) focused on readmission prediction for patients with heart valve surgery specifically. It used a generalized predictive equation for predicting readmission. Turgeman and May (2016) developed a predictive model for hospital readmissions using a boosted C5.0 tree and Support Vector Machine as base and secondary classifiers respectively. It tried to balance the readmission classification problem. Lewis et al. (2016), compared the accuracy of two different risk prediction models, The Hospital Readmission Reduction Program (HRRP) and the Risk-Standardized Readmission Rate (RSRR) models. (Tong et al., 2016) compared several existing models on an all-cause non-elective basis. LACE, LASSO logistic, AdaBoost, STEPWISE logistic are compared with varying sample sizes. Golmohammadi used neural networks, classification and regression models and chi-square automatic interaction detection for analysis (Golmohammadi and Radnia, 2016). All models had an overall accuracy of over 80%. The latter two gave the user the ability to select misclassification costs additionally. C5.0 was used to find any recurring patterns using patient history. Low et al. (2016) also compared the results with LACE index. After retrospective cohort analysis, like Kang et al. (2015), it grouped the predictors into categories. Wang et al. (2016), aimed to find the accuracy of Severity of Illness (SOI) and Risk of Mortality (ROM) individually, in predicting readmissions. Similar to Hogarth’s work (Mahajan et al., 2016; 2017) created a regularized logistic regression model for risk prediction on a thirty-day basis. This was yet another study in the heart related patient readmission domain and was limited to risk prediction and comparison of risk prediction models specifically (Mahajan et al., 2016). Kroeger et al. (2018) determines whether Pediatric Early Warning Score before transfer may serve as a predictor of unplanned readmission to the cardiac intensive care unit. Jiang et al. (2018) utilized feature selection algorithms and machine learning models to develop a risk prediction system that is dynamic and accurate. Several studies have implemented diverse modeling methods to determine the factors that influences hospital patient readmission rate (Betihavas et al., 2012; Davison et al., 2016; Golas et al., 2018; Hebert et al., 2014; Lum et al., 2012; SHAMEER et al., 2017; Shams et al., 2015; Wasfy et al., 2013; Yu et al., 2015). Methodology Our study does not involve the participation of any patient. All analysis is based on anonymized data and ensures confidentiality. The dataset consists of 55 attributes and a sample size of 100,000 instances and represents 10 years of data collected from 130 US hospitals (Avishek, 2018; Strack et al., 2014). Table 1 below shows the data distribution. The original database contains curtailed, superfluous and noisy information as expected in most of the real-world data (Strack et al., 2014). There were some attributes that could not be treated directly since they had a high percentage of missing values. These features were “weight” (97% values missing), “payer code” (40%) and “medical specialty” (47%). “Weight” attribute was too sparse to be considered and was not included in further analysis. “Payer code” was neglected since it had a high percentage of missing values and it was not considered relevant to the outcome. “Medical specialty” attribute was accounted for analysis, adding the value “missing” in order to account for missing values. Large percentage of missing values of the “weight” attribute can be explained by the fact that prior to the HITECH legislation of the American Reinvestment and Recovery Act in 2009 hospitals and clinics were not required to capture it in a structured format (Strack et al., 2014). 1322 Avishek Choudhury and Dr. Christopher M. Greene / American Journal of Engineering and Applied Sciences 2018, 11 (4): 1320.1331 DOI: 10.3844/ajeassp.2018.1320.1331 Table 1: Data description Predictors HbA1c No test was performed Result was high and the diabetic medication was changed Result was high, but the diabetic medication was not changed Normal result of the test Gender Female Male Discharge disposition Home Otherwise Admission source Emergency room Referrals Otherwise Specialty of the admitting physician Internal medicine Cardiology Surgery General practice Missing values Other Primary diagnosis Circulatory system Diabetes Respiratory system Digestive system Injury and poisoning Musculoskeletal and connective tissue problem Genitourinary system disease Neoplasm Other Race African American Caucasian Other Missing Age Less than or equal to 30yrs. 30-60 yrs. Older than 60yrs. Age Age in years Time in hospital Days between admission and discharge Number of encounters % of population Readmitted -----------------------------------------Number of encounters % in group 57,080 4,071 81.6% 5.8% 5,324 361 9.4% 8.9% 2,196 3.1% 166 7.6% 6,637 9.5% 590 8.9% 37,234 32,750 53.2% 46.8% 3,462 2,997 9.3% 9.2% 44,339 25,645 63.4% 36.6% 3,184 3,275 7.2% 12.8% 37,277 22,800 9,907 53.3% 32.6% 14.2% 3,563 2,032 846 9.6% 8.9% 8.5% 10,642 4,213 3,541 4,984 33,641 12,963 15.2% 6.0% 5.1% 7.1% 48.1% 18.5% 1,044 309 284 492 3,237 1,093 9.8% 7.3% 8.0% 9.9% 9.6% 8.4% 21,411 5,747 9,490 6,485 4,697 4,076 3,435 2,536 12,107 18.5% 8.2% 13.6% 9.3% 6.7% 5.8% 4.9% 3.6% 17.3% 1,093 529 710 532 524 354 313 239 1,129 8.4% 9.2% 7.5% 8.2% 11.2% 8.7% 9.1% 9.4% 9.3% 12,626 52,300 3,138 1,920 18.0% 74.7% 4.5% 2.7% 1,116 4,943 256 144 8.8% 9.5% 8.2% 7.5% 1,808 21,871 46,305 mean 64.9 2.6% 31.3% 66.2% median 67 112 1,614 4,733 1st Qu 55 6.2% 7.4% 10.2% 2nd Qu 77 4.3 3 2 6 The primary dataset contained numerous inpatient visits for some patients and the observations could not be considered as statistically independent, an assumption of the logistic regression model. We thus used only one encounter per patient; in particular, we measured only the first encounter for each patient as the primary admission and determined whether or not they were readmitted within 90 days. Furthermore, we detached all encounters that resulted in either discharge to a hospice or patient death. After filtering out the data, we were left with 69,984 encounters that constituted the final dataset for analysis. The methodology employed in this study can be broadly categorized into the following sections: Data preprocessing, implementation of predictive models and model optimization. 1323 Avishek Choudhury and Dr. Christopher M. Greene / American Journal of Engineering and Applied Sciences 2018, 11 (4): 1320.1331 DOI: 10.3844/ajeassp.2018.1320.1331 Data Preprocessing Handling Outliers Data preprocessing included three steps: Feature selection, handling outliers, data balancing and data partitioning. Median Absolute Deviation Method was used to address outliers (Avishek and Greene C., 2018) and (Leys, 2013). All missing values were replaced by the column mean. Feature Selection Data Balancing Large datasets hinder the speed of algorithms and even deteriorate classification accuracy (Kohavi and John, 1997). The concern raised due to data size is termed as the minimal-optimal problem (Nilsson, 2007). Our study employs Boruta algorithm and stepwise regression to determine the best features within the dataset. Boruta algorithm is a wrapper developed on random forest classification algorithm (Liaw and Wiener, 2002). In this algorithm, the relevance of any attribute is retrieved as the loss of classification accuracy caused due to permutation of attribute values among objects. It calculates the shuffled correlations between the response and the attributes. It also computes the Z-score to determine attributes’ relevance by dividing the mean accuracy loss by its standard deviation. In addition to Boruta, stepwise regression was also implemented. Stepwise regression is designed as an automatic computational procedure in which the performance of the regression increases with increase in the input variable (Barnett et al., 1975; Campolongo et al., 2000). Stepwise regression is a different version of the forward selection in which after every step a variable is added, all selected attributes in the model are analyzed to determine any loss in relevance. If an irrelevant variable is found, it is blocked from the model. Stepwise regression mandates two significance levels: One for adding attributes and one for eliminating attributes. The cutoff plausibility for adding an attribute must be less than the cutoff probability for eliminating attributes to avoid an infinite loop trap (Mengchao Wang, 2013). The entire data was randomly shuffled. Oversampling, undersampling and rose sampling methods were performed to balance the response variable (Al-Wesabi et al., 2018). Figure 1 shows the comparison of different data sampling methods. Missing Value Imputation Median Absolute Deviation Method was used to address missing values (Avishek, 2018;Leys, 2013). This method helps to avoid any outliers within the dataset. The same concern can also be addressed by scaling and normalizing the dataset between 0 and 1. The Equations (1) used in this study to perform data normalization are given below: Normalized ( Ni ) = ( Ni − Emin ) / ( Emax − Emin ) Where: Emin = The minimum value for variable E Emax = The maximum value for variable E *Note: If Emax equals Emin then Ni equals 0.5 Data Partitioning The balanced data was partitioned into training (70%) and testing (30%). All predictive models were fitted on the training data and the testing accuracy was measured for model evaluation. Implementation of Prediction Models Algorithms such as random forest, support vector machine, Recursive Partitioning and Regression Tree, Gradient Boosting Method and General Linear Model were used on the training data to predict the readmission risk. Random Forest Random Forests (RF) is a type of decision tree that employs modified tree learning algorithm. RF at each split of the input variables during the learning phase randomly selects a subset of features. This process is also termed called "feature bagging." and helps in determining the few highly correlated attributes that significantly influences the predictors for yielding the best-fit target output with high accuracy. Usually, for any classification problem having ‘x’ features, √x (rounded down) features are used in each split. Whereas for regression type analysis it is recommended having x/3 (rounded down) splits and at least node size of 5. Support Vector Machine A Support Vector Machine (SVM) is a discriminative classifier that uses a hyperplane to segregate different classes either linearly or radially. In other words, in supervised learning, the algorithm produces an optimal hyperplane which classifies output into specific categories. For SVM Linear the prediction Equation (2 shown below) for input is a dot product between the input (x) and each support vector (xi): (1) f ( x ) = Bo + ∑ ( ( ai ) * ( x, xi ) ) (2) Equation 2 computes the inner products of an input vector (x) with all support vectors in training data. The coefficients Bo and ai are assessed from the training data by the acquiring algorithm. 1324 Avishek Choudhury and Dr. Christopher M. Greene / American Journal of Engineering and Applied Sciences 2018, 11 (4): 1320.1331 DOI: 10.3844/ajeassp.2018.1320.1331 1.00 0.75 y-axis Model Original Over Rose Under 0.50 0.25 0.00 F1 Precision Recall x Sensitivity Specificity X-axis Fig. 1: Comparing data sampling methods. In this figure the x-axis represents the performance measure such as F1, precision, recall, sensitivity, specificity. The y-axis in the figure shows their respective value ranging from 0 and 1. Higher y value signifies better performance SVM polynomial kernel and exponential kernel are expressed as the respective Equations (3 and 4) given below: By implementing gradient descent, the minimum MSE can be found using the Equation (6) given below: y p i = y p i – α * 2 * ∑ ( yi − y p i ) K ( x, xi ) = 1 + ∑ ( x * xi ) d (6) (3) where, is learning rate and Σ(yi-yip) is sum of residuals. K ( x, xi ) = e ( −γ * ∑ (( x − xi 2)) (4) General Linear Model The general equation for the General Linear Model (GLM) is defined as the Equation (7) given below: Gradient Boosting Method Gradient boosting method is a classification and regression technique that ensembles several weak prediction models and produces a decision tree. GBM like any other supervised learning defines and minimizes the loss function. The Equation (5) below shows the Loss function: Loss = MSE = ∑ ( yi − y p i ) 2 y = βo + β1 X 1 + β2 X 2 + … + βn −1 X n −1 + βn X n (7) The βs in the given GLM equation are coefficients or weights dispensed to the input or predictor variables, i.e., the X’s on the right-hand side of the prediction equation. Recursive Partitioning and Regression Tree (5) where, yi = ith target value, yip = ith prediction, L(yi,yip) is the loss function. The Recursive Partitioning and Regression Tree (Rpart) algorithm splits the dataset recursively. In other words, the split continues till a given termination 1325 Avishek Choudhury and Dr. Christopher M. Greene / American Journal of Engineering and Applied Sciences 2018, 11 (4): 1320.1331 DOI: 10.3844/ajeassp.2018.1320.1331 criterion is attained. It is crucial to observe that the algorithm makes the best decision at each splitting stage, without any contemplation of optimality in the upcoming stages. In other words, this approach ensures local optimality. Due to this approach, deep trees are prone to overfitting. However, overfitting can be checked by developing shallower trees by terminating the algorithm at an ideal point or by pruning the deep tree to the desired criterion. Rpart follows the later technique to minimize overfitting. Overfitting minimization is achieved by the following Equation (8 and 9) shown below: min ( Cα (T ) ) (8) Equation 8 minimizes the cost Cα(T) assigned to each variable which is the linear combination (see Equation 9 below) of error R(T) and the number of leaf nodes in the tree |T|: Cα (T ) = R (T ) + α T (9) Performance Measures To analyze and compare each model, we considered accuracy, sensitivity and specificity (Choudhury, 2018). Accuracy (ACC) is determined as the number of correct prediction upon the total number of the dataset. The accuracy can vary from 0 to 1. Where 1 is the best possible result. Sensitivity (SN) is the count of true (correct) positive predictions divided by the total number of positives. The value of SN varies from 0 to 1, where 1 is the best possible sensitivity. Specificity (SP) is the number of true (correct) negative predictions divided by the total number of negatives. The value of SP varies from 0 to 1, where 1 is the best possible specificity. Table 2: Parameters for genetic algorithm GA Characteristics Genes’ alley in chromosome Real values Chromosome length = No. of variables Population Random (uniform) population of real values Size = 15 Selection Strategy Linear-rank selection Crossover Method Local arithmetic crossover Crossover probability = 0.8 Mutation Method Uniform random mutation Mutation Probability = 0.1 Replacement Strategy Elitism by %5 Termination Strategy No. of generations = 15 Constraint Handling Constraints repair mechanism bound of [-10,10] Model Optimization Genetic Algorithm and Greedy Ensemble algorithm were implemented to obtain the best fit model. Genetic Algorithm (GA) is a random but global search in solution space that is inspired by natural behavior of chromosomes in transmitting characteristics (genes) from one generation after another in which genes will be updated randomly with the aim of crossovers and mutations strategies through generations to produce a chromosome which is the best representative of optimal solution (Rowe, 2015). This is study considered GA for tuning a set of classifiers which provided the best performance among applied data mining methods in the previous section due to enhance the predictive accuracy and investigate the effect of classifiers’ parameters on its performance measurement. Applied GA characteristics have been introduced in Table 2. Results Boruta algorithm performed 18 iterations in 1.32 h and identified 10 important attributes as shown in the Fig. 2. However, “number of inpatients”, “number of emergencies”, “number of diagnoses”, “diabetes med”, “number of outpatients”, “number of procedures” and “number of medications” were identified as the top seven influential factors by stepwise regression as shown in Fig. 3. The important attributes identified by both Boruta and stepwise regression were (a) number of medications, (b) number of procedures, (c) number of emergencies, (d) number of outpatients, (e) number of inpatients and (f) number of diagnosis. For further analysis we employed interaction effect study. Interaction effects study is the analysis of how multiple predictor variables, when considered together, have an impact on the main variable for analysis. It helps establish a relationship, not just between the predictors and the main variable, but also between the predictor variables themselves as shown in Fig. 4. The blue and red lines mark the low and high levels of one variable when it is being considered along with another variable. For example, in the interaction plot of the number of outpatients*number of inpatients, in a higher value setting of the number of inpatients (21), the mean of readmitted is decreasing (approaching zero) for some outpatients increases. This signifies that if every time a patient is an inpatient, he or she also becomes an outpatient, it becomes less likely that he or she will be readmitted. Similarly, there seems to be a change in behavior when the number of inpatient and number of diagnoses is considered, wherein the number of diagnoses is the changing variable. When the number of diagnoses setting is higher, it naturally means that chances of 1326 Avishek Choudhury and Dr. Christopher M. Greene / American Journal of Engineering and Applied Sciences 2018, 11 (4): 1320.1331 DOI: 10.3844/ajeassp.2018.1320.1331 readmission will remain high with increasing number of inpatients for a particular patient. However, even in a low setting, the mean of readmitted increases and crosses over the high setting line. This may signify that of a particular patient is an inpatient again and again, but there are no new diagnoses, the hospital may be failing concerning accuracy of judgment, which is why chances if readmission still increase. Boruta Feature Selection 20 15 0 Importance 10 0 0 5 0 No. In-Patient No. Emergency No. out-Patient No. Lab Procedures No. Medications Time in hospital No. Procedures No. Diagnosis Age DIABETES MED Shadow Max Shadow Mean Shadow Min -5 Fig. 2: Boruta feature selection 12 10 80 8 60 6 40 4 Fig. 3: Stepwise regression variable importance 1327 num_medications 7 num_procedures 6 diabetesMed 12 0 number_emergency 9 0 number_outpatient 8 20 number_diag0ses 11 2 number_inpatient 10 nsubsets 100 nsubsets sqrt qcv sqrt rss Avishek Choudhury and Dr. Christopher M. Greene / American Journal of Engineering and Applied Sciences 2018, 11 (4): 1320.1331 DOI: 10.3844/ajeassp.2018.1320.1331 Interaction plots for readmitted Describes how readmitted changes if you change the settings of two X variables. num_procedur’ num_medicati num_medicati 1 81 0 -2 -4 num_procedur’ number_outpa num_medicati’ number_outpa Mean of readmitted 0 -2 -4 num_ procedur’ number_inpat num_medicati’ number_inpat number_outpa’ number_inpat number_ inpat 0 21 0 -2 -4 num_procedur’ number_diag0 num_medicati’ number_diag0 number_outpa’ number_diag0 number_inpat’ number_diag0 number_diag0 1 9 0 -2 -4 0.0 2.5 5.0 0 40 80 0 20 40 0 10 20 Fig. 4: Interaction effect analysis Table 3: Model performance after data processing Algorithm Gradient boosting method General linear model Support vector machine Recursive partitioning and regression trees Accuracy (%) 96.12 96.35 97.00 96.90 Table 4: Model performance after implementing genetic algorithm Gradient General boosting method linear model Optimized accuracy (%) 97.05 97.05 Optimized classifiers’ parameters -0.69, -2.36 1.32, 1.64 We implemented under-sampling method to reduce the bias of the response variable. Table 3 shows the model performance and testing accuracy of all the selected algorithms. To further enhance the performance of the models, the Genetic Algorithm was used to obtain the optimized performance as shown in the following Table 4. Since using Genetic Optimization gave two best models with same output, we performed greedy ensemble to find the best one model. Gradient Boosting Method was found to be the best after both Genetic Optimizations as well as Greedy Ensemble Method and we recommend GBM with 98.50% prediction accuracy as the best fit model for this dataset. Moreover, “number of inpatients” was found to be the most influencing factor that determines patient readmission risk. Sensitivity 0.94 0.94 0.98 1.00 Support vector machine 97.04 -4.46, 3.30 Specificity 0.97 0.96 0.96 0.96 Recursive partitioning and Regression Tree 97.04 0.50, 5.67 However, patient age, diabetes, time spent in the hospital, number of lab procedures, number of outpatients and the number of emergencies had significant relevance. Discussion and Conclusion Readmission rate is a quality evaluation metric customarily used to extrapolate the quality of life index of patient population and the quality of healthcare delivery (Shameer et al., 2017). Irrespective of the developments in biomedical and healthcare research practices, hospital quality control offices still use traditional predefined sets of variables to infer the probability patient readmission (Shameer et al., 2017). However, predictive analytics could provide evidences to improve the quality of healthcare delivery. Uniting 1328 Avishek Choudhury and Dr. Christopher M. Greene / American Journal of Engineering and Applied Sciences 2018, 11 (4): 1320.1331 DOI: 10.3844/ajeassp.2018.1320.1331 predictive analytics with preventive measures would involve patients, physicians and payers to contribute proactively in taming the health and wellness In this study, we implement a predictive analytical approach to identify patients prone to readmission and thus, systematically reduce the number of avoidable readmissions mainly caused by patient non-compliances to medication instruction or early discharge from hospital. Our proposal has the capability of capturing both patient and population-based variations of hospital readmissions. It incorporates patient with diverse health concerns across 130 US hospitals. The novelty of our method is to directly incorporate patients’ history of readmissions into modeling framework along with other demographic and clinical characteristics. We also verify the effectiveness of the proposed approach by validating training accuracy. Some contributions made in this paper are (i) applying Boruta algorithm and stepwise variable selection and (ii) implementing genetic and greedy ensemble algorithm to optimize our predictive models. Our study recommended optimized gradient boosting method for identifying patient most likely to get readmitted. Furthermore, the study also emphasizes on the effectiveness of data preprocessing. It measures the influence of data balancing, removing outliers and imputing missing values on the classification accuracy. Our study also produces highest readmission prediction accuracy. Some research directions can be sought by trying different variable selection techniques such as LASSO or Nonnegative Garrote for better subset regressions. Also, in presence of high right censored data, it is interesting to consider some health care cost measures from which it may be possible to statistically estimate the mean population cost for readmission. Author’s Contributions Avishek Choudhury: Research, data collection, analysis, data interpretation, figure formation, coding and writing manuscript. Dr. Christopher M. Greene: Manuscript writing and formatting. Ethics All data was collected with the permission of the organization and patient’s medical and personal information was secured. Data Statement The data used in this study can be retrieved from DOI: 10.17632/nntck7ddgt.1. URL: https://data.mendeley.com/datasets/nntck7ddgt/2 References Ajorlou, S., I. Shams and K. Yang, 2014. Predicting patient risk of readmission with frailty models in the Department of Veteran Affairs. Proceedings of the IEEE International Conference on Automation Science and Engineering, Aug. 18-22, IEEE Xplore Press, Taipei, Taiwan, pp: 576-581. DOI: 10.1109/CoASE.2014.6899384 Allison, N., 2012. Examining the likelihood of readmission to inpatient pulmonary rehabilitation using a variety of predictors. ProQuest Dissertations and Thesis, ProQuest Dissertations Publishing. Al-Wesabi, Y.M.S., A. Choudhury and D. Won, 2018. Classification of cervical cancer dataset. Amarasingham, R., F. Velasco, B. Xie, C. Clark and Y. Ma et al., 2015. Electronic medical record-based multicondition models to predict the risk of 30-day readmission or death among adult medicine patients: Validation and comparison to existing models. BMC Med. Inform. Dec. Mak., 15: 39-39. DOI: 10.1186/s12911-015-0162-6 Anika, L. and M.L. Hines, 2014. Conditions with the largest number of adult hospital readmissions by payer. HCUP Statistical Brief# 172. Rockville: Agency for Healthcare Research and Quality. https://www.hcupus.ahrq.gov/reports/statbriefs/sb17 2-Conditions-Readmissions-Payer.pdf Appleby, L.K.N., 2013. The national confidential inquiry into suicide and homicide by people with mental illness. Healthwatch England. Avishek, C. and C.M. Greene, 2018. Prognosticating autism spectrum disorder using artificial neural network: Levenberg-Marquardt algorithm. J. Bioinform. Syst. Biol., 1: 001-010. DOI: 10.26502/fjbsb001 Avishek, C.K., 2018. Decision support system for renal transplantation. Proceedings of the IISE Annual Conference, (AC’ 18), IISE, Orlando. Bakal, J.A., F.A. Mcalister, W. Liu and J.A. Ezekowitz, 2014. Heart failure readmission: Measuring the ever-shortening gap between repeat heart failure hospitalizations. PLoS ONE, 9: 106494-106494. DOI: 10.1371/journal.pone.0106494 Bayati, M., M. Braverman, M. Gillam, K.M. Mack and G. Ruiz et al., 2014. Data-driven decisions for reducing readmissions for heart failure: General methodology and case study. PloS one, 9: e109264e109264. DOI: 10.1371/journal.pone.0109264 Betihavas, V., P.M. Davidson, P.J. Newton, S.A. Frost and P.S. Macdonald et al., 2012. What are the factors in risk prediction models for rehospitalisation for adults with chronic heart failure? Australian Critical Care, 25: 31-40. DOI: 10.1016/j.aucc.2011.07.004 1329 Avishek Choudhury and Dr. Christopher M. Greene / American Journal of Engineering and Applied Sciences 2018, 11 (4): 1320.1331 DOI: 10.3844/ajeassp.2018.1320.1331 Campolongo, F., S. Tarantola and A. Saltelli, 2000. Sensitivity anaysis as an ingredient of modeling. Stat. Sci. DOI: 10.1214/ss/1009213004 Choudhury, C., 2018. Identification of cancer: Mesothelioma’s desease using logistic regression and association rule. Am. J. Eng. Applied Sci. Claire Senot, A.C., 2015. What has the biggest impact on hospital readmission rates. Harvard Business Review. Davison, B.A., M. Metra, S. Senger, C. Edwards and O. Milo et al., 2016. Patient journey after admission for acute heart failure: length of stay, 30-day readmission and 90-day mortality. Eur. J. Heart Failure. DOI: 10.1002/ejhf.540 Davood Golmohammadi, N.R., 2015. Prediction modeling and pattern recognition for patient readmission. Int. J. Product. Econom., 171: 151-161. DOI: 10.1016/j.ijpe.2015.09.027 Dixon Woods, M.A., 2005. Synthesizing qualitative and quantitative evidence: A review of possible methods. J. Health Services Res. Policy, 10: 45-53. DOI: 10.1177/135581960501000110 Futoma, J., J. Morris and J. Lucas, 2015. A comparison of models for predicting early hospital readmissions. J. Biomed. Inform., 56: 229-238. DOI: 10.1016/j.jbi.2015.05.016 Golas, S.B., T. Shibahara, S. Agboola, H. Otaki and J. Sato et al., 2018. A machine learning model to predict the risk of 30-day readmissions in patients with heart failure: A retrospective analysis of electronic medical records data. BMC Med. Inform. Dec. Mak. DOI: 10.1186/s12911-018-0620-z Golmohammadi, D. and N. Radnia, 2016. Prediction modeling and pattern recognition for patient readmission. Int. J. Product. Econom. DOI: 10.1016/j.ijpe.2015.09.027 Greenhalgh, 2005. Effectiveness and efficiency of search methods in systematic reviews of complex evidence: Audit of primary sources. BMC, 331: 1064-1065. DOI: 10.1136/bmj.38636.593461.68 Hebert, C., C. Shivade, R. Foraker, J. Wasserman and C. Roth et al., 2014. Diagnosis-specific readmission risk prediction using electronic health data: A retrospective cohort study. BMC Med. Inform. Dec. Mak. DOI: 10.1186/1472-6947-14-65 Hodgson, R.E., M. Lewis and A.P. Boardman, 2001. Prediction of readmission to acute psychiatric units. Soc. Psychiatry Psychiatric Epidemiol., 36: 304309. DOI: 10.1007/s001270170049 IHSMHL, 2012. Improving hospital admission and discharge for people who are homeless. Inclusion Health St Mungo's Homeless Link, Healthwatch England. Inouye, S., V. Bouras, E. Shouldis, A. Johnstone and Z. Silverzweig et al., 2015. Predicting readmission of heart failure patients using automated follow-up calls. BMC Med. Inform. Dec. Mak., 15: 22-22. DOI: 10.1186/s12911-015-0144-8 Jiang, S., K.S. Chin, G. Qu and K.L. Tsui, 2018. An integrated machine learning framework for hospital readmission prediction. Knowledge-Based Syst., 146: 73-90. DOI: 10.1016/j.knosys.2018.01.027 Kang, C., K. Kim, J.H. Lee, Y.H. Jo and J.W. Park et al., 2015. Predictors of revisit and admission after discharge from an emergency department in acute pyelonephritis. Hong Kong J. Emergency Med., 22: 154-162. DOI: 10.1177/102490791502200304 Kathleen Carey, T.S., 2016. The Cost of hospital readmissions: Evidence from the VA. Health Care Manage. Sci., 19: 241-248. DOI: 10.1007/s10729014-9316-9 Kohavi, R. and G.H. John, 1997. Wrappers for feature subset selection. Artificial Intell., 97: 273-324. DOI: 10.1016/S0004-3702(97)00043-X Kroeger, A.R., J. Morrison and A.H. Smith, 2018. Predicting unplanned readmissions to a pediatric cardiac intensive care unit using pre-discharge Pediatric Early Warning Scores. Congenital Heart Dis., 13: 98-104. DOI: 10.1111/chd.12525 Lewis, C.M., Z.L. Cox, P. Lai and D.J. Lenihan, 2016. Evaluation of two different models to predict heart failure readmissions. Heart Lung: J. Acute Critical Care, 45: 374-374. DOI: 10.1016/j.hrtlng.2016.05.008 Leys, C., 2013. Detecting outliers: Do not use standard deviation around the mean, use absolute deviation around the mean. J. Exp. Soc. Psychol., 49: 764-766. DOI: 10.1016/j.jesp.2013.03.013 Liaw, A. and M. Wiener, 2002. Classification and Regression by random Forest. R News. Lum, H.D., S.A. Studenski, H.B. Degenholtz and S.E. Hardy, 2012. Early hospital readmission is a predictor of one-year mortality in communitydwelling older medicare beneficiaries. J. General Internal Med. DOI: 10.1007/s11606-012-2116-3 Maguire, D., 2015. The king's fund. Premature discharge: Is going home early really a Christmas gift? kingsfund.org.uk Mahajan, S., P. Burman and M. Hogarth, 2016. Analyzing 30-day readmission rate for heart failure using different predictive models. Nurs. Inform., 2016: 225: 143-147. PMID: 27332179 Mahajan, S.M., P. Burman, A. Newton and P.A. Heidenreich, 2017. A validated risk model for 30day readmission for heart failure. Stud. Health Technol. Inform., 245: 506-510. PMID: 29295146 1330 Avishek Choudhury and Dr. Christopher M. Greene / American Journal of Engineering and Applied Sciences 2018, 11 (4): 1320.1331 DOI: 10.3844/ajeassp.2018.1320.1331 Mengchao Wang, J.W., 2013. A comparison of approaches to stepwise regression for global sensitivity analysis used with evolution optimization. Proceedings of the 13th Conference of International Building Performance Simulation Association, (PSA’ 13), France, pp: 2551-2558. Miller, D.J., N.C. Beck and C. Fraps, 1984. Predicting rehospitalization at a community mental health center: A „double‐crossed” validation. J. Clin. Psychol., 40: 35-39. DOI: 10.1002/1097-4679(198401)40:1<35::AIDJCLP2270400106>3.0.CO;2-8 Nilsson, R., 2007. Consistent feature selection for pattern recognition in polynomial time. J. Mach. Learn. Res. Ombudsman, T.P.O., 2003. The parliamentary ombudsman and the health service ombudsman website. Parliamentary Ombudsman and the Health Service Ombudsman. Pack, Q.R., A. Priya, T. Lagu, P.S. Pekow and R. Engelman et al., 2016. Development and validation of a predictive model for short‐and medium‐term hospital readmission following heart valve surgery. J. Am. Heart Assoc., 5: e003544-e003544. DOI: 10.1161/JAHA.116.003544 Rowe, J.E., 2015. Genetic Algorithms. In: Springer Handbook of Computational Intelligence, Kacprzyk, J. and W. Pedrycz (Eds.), Springer, Berlin, pp: 825-844. Shameer, K., K.W. Johnson, A. Yahi, R. Miotto and L. Li et al., 2017. Predictive modeling of hospital readmission rates using electronic medical recordwide machine learning: A case-study using mount sinai heart failure cohort. Biocomputing. Shams, I., S. Ajorlou and K. Yang, 2015. A predictive analytics approach to reducing 30- day avoidable readmissions among patients with heart failure, acute myocardial infarction, pneumonia, or COPD. Health Care Manage. Sci., 18: 19-34. DOI: 10.1007/s10729-014-9278-y Shinkman, R., 2014. Questex. Fierce Healthcare. https://www.fiercehealthcare.com/finance/readmissi ons-lead-to-41-3b-additional-hospital-costs Shipeng Yua, F.F., 2015. Predicting Readmission Risk with Institution-Specific Prediction Models. Artificial Intell. Med., 65: 89-96. DOI: 10.1016/j.artmed.2015.08.005 Strack, B., J.P. Deshazo, C. Gennings, J.L. Olmo and S. Ventura et al., 2014. Impact of HbA1c measurement on hospital readmission rates: Analysis of 70,000 clinical database patient records. BioMed Res. Int. DOI: 10.1155/2014/781670 Tong, L., C. Erdmann, M. Daldalian, J. Li and T. Esposito, 2016. Comparison of predictive modeling approaches for 30-day all-cause non-elective readmission risk. BMC Med. Res. Methodol. DOI: 10.1186/s12874-016-0128-0 Turgeman, L. and J.H. May, 2016. A mixed-ensemble model for hospital readmission. Artificial Intell. Med., 72: 72-82. DOI: 10.1016/j.artmed.2016.08.005 Wang, H., C. Johnson, R.D. Robinson, V.A. Nejtek and C.D. Schrader et al., 2016. Roles of disease severity and post-discharge outpatient visits as predictors of hospital readmissions. BMC Health Services Res., 16: 564-564. DOI: 10.1186/s12913-016-1814-7 Wang, H., R.D. Robinson, C. Johnson, N.R. Zenarosa and R.D. Jayswal et al., 2014. Using the LACE index to predict hospital readmissions in congestive heart failure patients. BMC Cardiovascular Disorders, 14: 97-97. DOI: 10.1186/1471-2261-14-97 Wasfy, J.H., K. Rosenfield, K. Zelevinsky, R. Sakhuja and A. Lovett et al., 2013. A prediction model to identify patients at high risk for 30-day readmission after percutaneous coronary intervention. Circulation: Cardiovascular Quality and Outcomes. Yu, S., F. Farooq, A. van Esbroeck, G. Fung and V. Anand et al., 2015. Predicting readmission risk with institution-specific prediction models. Artificial Intell. Med. DOI: 10.1016/j.artmed.2015.08.005 Zheng, B., J. Zhang, S.W. Yoon, S.S. Lam and M. Khasawneh et al., 2015. Predictive modeling of hospital readmissions using metaheuristics and data mining. Expert Syst. Applic., 42: 7110-7120. DOI: 10.1016/j.eswa.2015.04.066 1331