Please download to get full document.

View again

of 86
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Information Report



Views: 4 | Pages: 86

Extension: PDF | Download: 0

Related documents
A STRATEGY FOR REDUCING CONGESTIVE HEART FAILURE READMISSIONS THROUGH THE USE OF INTERVENTIONS TARGETED BY MACHINE LEARNING PREDICTION A Dissertation Presented to The Graduate Faculty at the University of Akron In Partial Fulfillment of the Requirements for the Degree Doctor of Philosophy James Natale May, 2015 i A STRATEGY FOR REDUCING CONGESTIVE HEART FAILURE READMISSIONS THROUGH THE USE OF INTERVENTIONS TARGETED BY MACHINE LEARNING PREDICTION James Natale Dissertation Approved: Advisor Dr. Shengyong Wang Committee Member Dr. Chien-Chung Chan Committee Member Dr. Richard Einsporn Committee Member Dr. Ping Yi Accepted: Department Chair Dr. Sergio Felicelli Dean of the College Dr. George K. Haritos Interim Dean of the Graduate School Dr. Rex D. Ramsier Date Committee Member Dr. Jiang Zhe ii ABSTRACT Hospitals are faced with a financial penalty when patients are readmitted to the hospital within thirty days of discharge. However preventing readmissions is a difficulty for hospitals. There are a wide variety of interventions which are aimed towards improved patient outcomes, as well as preventing patient readmissions. However these interventions are expensive in terms of financial outlay, as well as in the time spent by staff. However, if interventions are applied indiscriminately, these methods would cost more than they would have saved due to readmissions prevention. We describe an all encompassing strategy which would allow for hospitals to reduce their readmissions in a cost-effective manner. We apply an analytical approach to all aspects of the problem. By creating a predictive model with machine learning methods on hospital records, we can determine the risks of patients being readmitted. We detail how the literature on intervention strategies can be condensed and utilized to determine prospective strategies which may be of interest. Utilizing the risk predicted by the model, as well as published literature on interventions we determine the optimal solution of which patients will receive which interventions through a genetic algorithm heuristic search. Only by combining the three aspects together can we formulate an analytics driven approach to reducing readmissions in a cost-effective manner. iii TABLE OF CONTENTS Page LIST OF TABLES... vii LIST OF FIGURES...ix CHAPTER I. INTRODUCTION Research Motivation Research Objectives Dissertation Overview...2 II. RESEARCH BACKGROUND AND LITERATURE REVIEW Rise of Chronic Disease and the Readmission Crisis Problem Justification Optimization Problem Literature Review Reducing Readmissions Predicting Readmission Risk...11 III. RISK PREDICTION Data Collection Data Cleaning Data Mining Initial Investigation...21 iv Feature Selection Feature Selection Discussion with Domain Experts Cost Sensitive Learning Selection of Modeling Approach Calculating Patient Risk Finalization of Modeling Approach The Class Imbalance Problem Learning Algorithm Feature Selection...42 IV. OBTAINING INTERVENTION OPTIONS FROM LITERATURE Efficacy Methodology Followed Example Example Example Summarized Literature on CHF Readmission Interventions Formulation of Estimates for Experimentation Discussion of Results...57 V. OPTIMIZATION OF IMPLEMENTATION THROUGH HUERISTICS Understanding the Genetic Algorithm Enterprise Economic Concerns Results of Experimentation Base-Line Measures...69 v Results Interpretation Sensitivity Analysis Comparison to Other Solution Heuristics...84 VI. FUTURE RESEARCH Method for Applying Approach in Practice Intervention Options Sampling Missing Data Imputation Drop Features Drop Patients Imputation Portability of Research Framework to Other Hospitals A Question of Root Cause Differentiation Between Necessary and Unnecessary Readmissions Summary of Results...98 BIBLIOGRAPHY vi LIST OF TABLES Table Page 1 Overview of parameters used in data mining Data Mining Algorithms trialed on the CHF readmissions data set 20 3 Feature selection algorithms types Common filter feature selection methods Univariate Feature Selection for features often found in literature Univariate statistical test feature selection for features found in the care coordination admission assessment Univariate statistical test feature selection for features obtained from Crimson care registry Univariate statistical test feature selection from procedural and diagnosis coding 30 9 A Comparison of Features Selected in Each Method Relevant features which must be answerable by an intervention article Outcomes of Literature, Utilizing Patient Interviews On Reducing Readmissions Amount of Literature by Group A synopsis of the relevant features covered in an article A synopsis of the relevant features covered in an article A synopsis of the relevant features covered in an article Amount of Literature after Validity Assessment vii 17 Finalized List of Eight Programs Considered for Heuristics Finalized List of Eight Programs Considered for Heuristics 56 viii LIST OF FIGURES Figure Page 1 Example Receiver Operator Characteristic curve Example of the Rapid Miner Process creation Representation of the k-fold cross-validation which is often performed Rapid Miner process for predicting CHF readmissions through c comparing ROC curves Sub-process of the compare ROC's operation, which gives the c modeling algorithms and necessary transforms Initial performance in readmission prediction Entirety of a random forest tree for CHF readmission prediction C simply guesses 'no' for every patient Representation of the SMOTE algorithm creating synthetic points ROC curve obtained through data post feature selection and C oversampling with an optimistic bias ROC curve obtained through data post feature selection and c oversampling with an pessimistic bias ROC curve for the prediction of CHF patient readmissions c under univariate feature selection ROC Curve for patient readmissions under the features selected c by the tree-based feature selection Performance of the random forest on the classification of CHF c readmissions with SMOTE Performance of the random forest on the classification of CHF c readmissions without SMOTE ix 15 Final leaf nodes of an example patient Genetic algorithm which was implemented to determine the c optimal solution for applied interventions Representation of the two point crossover mechanism which was utilized Patient risk is laid out sequentially, with patients at highest risk first Representation of the solution vector format used A view of the population characteristics as generations increase Showing model performance evaluation as the generation increases A plot of the difference in the fitness values as the generation increases Plot which shows the starting point and the length for the vector c of application of program 1 to patients Plot which shows the starting point and the length for the vector c of application of program 2 to patients Plot which shows the starting point and the length for the vector c of application of program 3 to patients Plot which shows the starting point and the length for the vector c of application of program 4 to patients Plot which shows the starting point and the length for the vector c of application of program 5 to patients Plot which shows the starting point and the length for the vector c of application of program 6 to patients Which shows the starting point and the length for the vector of c application of program 7 to patients Which shows the starting point and the length for the vector of c application of program 8 to patients The performance under the 'inflated' rate The performance under the best estimated rate The performance under the 'deflated' rate x 34 Performance comparisons of several machine learning algorithms c and the changes in performance due to sample size xi CHAPTER I INTRODUCTION 1.1. Research Motivation The purpose of this research is the formulation of a holistic strategy for hospital systems to follow in order to financially optimize their efforts to reduce hospital readmissions. For delivering health care, health care organizations are ultimately concerned with the readmissions of patients. The largest diagnosis group which impact hospital readmissions is Congestive Heart Failure (CHF) patients, which often have a readmit rate as high as 25%. In order to reduce readmissions, a hospital system must engage strategies which prevent readmissions. However there are strict limitations on the amount of expenditure which a hospital system may use to reach their goals. In order to prevent readmissions effectively and efficiently, we treat this as an optimization problem, in which we are required to complete all of the components outlined below. Prediction of readmission risk for Congestive Heart Failure patients Methods of interventions o Expected reduction in readmissions o Cost of Interventions Financial impact of Readmissions Optimization techniques to select the appropriate methods, and individual patients 1 1.2. Research Objectives The objective of this research is to apply an analytics driven approach which combines several commonly used algorithms to each part of the readmission reduction problem. While the case considered is for CHF, it is trivial to apply this to other diagnosis groups. By giving the case of CHF, we are able to show that the framework developed is applicable in a real-world situation. We use a local community hospital system's records to provide the proof of concept Dissertation Overview The dissertation is divided into six chapters, beginning with the initial introduction, and providing background information on the subject area. As this research lies in the intersection of many topic areas, we strive to address each, including the current state and trends of healthcare, and the description of the problem. Chapter 2 will include a summary of the problem as well as provide background information which is required for problem understanding it will also include a justification for this investigation and a literature review of other approaches which attempt to meet the same end goals. Chapter 3 will include an introduction into the data mining process which was used in the formulation of the risk model. As has already been discussed the creation of the risk model remains a central component to overall efficacy of this readmission reduction framework. We will discuss topics, which include the available data sources, the formats which data were received in data preprocessing steps, including the creation of new variables and transformations of previously existing variables. As well, information on the current modeling approach which is been shown to be successful will also be explained including data sampling methodologies prediction model, formulations and 2 classification results. With the prediction model explained Chapter 4 will focus on the information which would be available to other hospital systems or healthcare organizations for the development of their intervention programs. Within this proposal, the scope will be limited to defining the information available and explaining how this information could be utilized by healthcare systems. However, the end-use of this, as well as the formal definitions and methodology which should be used in transferring published literature into defined intervention strategies and cost assessments as well as assessments of overall efficacy will be left for future research. Chapter 5 will include information related to search heuristics in the overall development of the final intervention strategy for the company to utilize. We have already discussed how risk is assessed and predicted and also how literature provides insight into possible intervention strategies. This chapter will focus on how these two ideas can be condensed down and utilized in an analytic fashion to define an optimum strategy. The focus will also include those aspects of the organization structure, which would be important in the formulation of constraints as well as in the functions which might be considered or excluded. The remaining chapters will be focused on results, conclusions, and discussion related to the implementation of this readmission reduction framework. 3 CHAPTER II RESEARCH BACKGROUND AND LITERATURE REVIEW 2.1. Rise of Chronic Disease and the Readmission Crisis At present, it is estimated that 75% of healthcare expenditures are focused on the treatment of chronic diseases.[1][2] A chronic disease, is one for which there is no known cure, which patients will have for the rest of their lives. While these conditions may be managed, reducing or eliminating the side-effects of the disease, there is no cure. There are several factors at work which have lead to this. Improvements in healthcare can now prevent mortality in situations which would have previously been terminal. Improved chronic disease management programs lower the mortality risk of patients, and many other additional factors. However the explanation for this rise is of little interest to this investigation. We are merely concerned that this is the case, and what our aims should be since it is. Having a growing proportion of the patient population living with one or more chronic disease has large effects on the healthcare system. One of the largest impacts deals with readmission rates. When a patient enters the hospital as an inpatient they are considered admitted to the hospital. Once the patients treatment has completed, and they are stable, the patient is discharged. At this point, the goal is that the patient is no longer requiring hospital inpatient services, and should not require them for a long period of time. One metric for the quality of care that a hospital is providing is called the readmit 4 rate. Essentially a readmission occurs when the patient is once again admitted to the hospital after they have been discharged. Typically, the length of time which is of interest is 30 days. This means that a patient who is admitted to the hospital within 30 days of their last discharge is considered a readmission. However we should note that there are a variety of lengths which are often tracked, such as 7, 15, 60, 90, 365. The logic behind using this rate as a metric for hospital care delivery quality is that hospital systems which often have patients returning to the hospital after a discharge must be delivering inferior care as compared to those which are able to prevent a larger percentage of readmissions. This also enables the elimination of a conflict of interest, whereby a hospital system would have an incentive to deliver poor care as that would generate additional patient volume from repeated admissions from patients. The importance of this issue has been magnified recently due to an addition of section 1886(q) to the Social Security Act, an addition as part of the Affordable Care Act.[3] This act establishes several things. One is that hospitals with disproportionately high readmissions will face reductions in reimbursements. It does this by viewing the readmissions for patients with the conditions of Acute Myocardial Infarction (AMI), Heart Failure (HF) and Pneumonia (PN) and determining the excess readmission ratio, a calculation based on a national average and adjusting per the hospital's patient population, including aspects such as comorbidities, frailty, and demographic characteristics. In future, the conditions being targeted are likely to be expanded to include acute exacerbation of chronic obstructive pulmonary disease (COPD); and 5 patients admitted for elective total hip arthroplasty (THA) and total knee arthroplasty (TKA). The specific definitions for calculations are shown below.[4] Formulas to Calculate the Readmission Adjustment Factor Excess readmission ratio = risk adjusted predicted readmissions risk adjusted expected readmissions Aggregate payments for excess readmissions = (excess readmit ratio for AMI 1) base operating DRG payments for AMI + (excess readmit ratio for HF 1) base operating DRG payments for HF + (excess readmit ratio for PN 1) base operating DRG payments forpn Aggregate payments for all dischrages = base operating DRG payments for all dischrages 6 Aggregate payments for all discharges = 1 Aggregate payments for excess readmissions Aggregate payments for all discharges Readmission Adjustment Factor = FY 2013: the higher of the Ratio or 0.99 (1% reduction) FY 2014: the higher of the Ratio or 0.98 (2% reduction) Formulas to Compute the Readmission Payment Adjustment Amount Wage adjusted DRG operating amount = DRG weight [(labor share wage index) + (nonlabor share cola) Base Operating DRG Payment Amount = Wage adjusted DRG operating amount + new technology payment Readmissions Payment Adjustment Amount = [Base operating DRG payment amount readmissions adjustment factor] base operating DRG payment amount 7 2.2. Problem Justification As the above changes show, if a hospital system cannot reduce their readmission rate to a level which can avoid penalties, it will face a financial burden. However, whatever the strategy to reduce the readmission rate, a reasonable cost must be maintained, while still remaining effective. Otherwise penalties will still be applied, along with the cost of readmission reduction strategies. We see that this can pose a difficult problem to solve. Exacerbating the problem is a secondary issue which is prompted by the necessity of managing costs associated with readmission reduction programs. As the readmissions account for only a small percentage of the patients, the vast majority of patients require no interventions. As such, an initiative which targets all patients being discharged would be particularly wasteful. However that still leaves the issue of being able to identify those patients which would be likely to be readmitted. This is a difficult aspect of the problem as there are a multitude for reasons or characteristics which may be responsible for hospital readmissions among the patients, and quite likely differing ones for each of the separate diagnoses Optimization Problem So we have a situation in which determining the optimum solution represents a nontrivial problem. We are faced with determining all of the following; What interventions to use How large should our intervention recipient group be Which specific individuals should be targeted What is the financial outcome 8 All of which are subject to several constraints in determining the solution for the optimum readmission reduction strategy. Consider the following constraints; Effectiveness of interventions (and changes based on reductions) Cost of interventions Additional staff required Staff duty reallocation Reimbursement reduction Improvements made by cohort hospitals Budget constraints All of this must happen in an environment in which the use of experimentation to determine the strategy is largely imprudent. The outcomes of experimental trials would be on actual patients. Beyond this ethical aspect are the more practical problems associated with running experiments in which the variation between individual subjects is high, and the length of time required to generate sample sizes large enough to show significant results. Moreover the managerial concerns of constructing an intervention program, even on a trial basis would be large. Also consider that a choice like this might very well require additional staff to accomplish. It would be problematic to go through a hiring process to implement a trial program, as the times to fill positions, especially skilled ones, can be quite long. Put simply, the readmission reduction strategy cannot be of the guess and check variety, but must be well thought out, analytics driven strategy, as 9 the timeliness of implementation is significant, and financial impact of these decisions is high. What we propose in the solution of the problem represents a method whereby hospitals may leverage the existing literature from implemented readmission reduction initiatives, information which cannot be achi
View more...
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks