Predicting Hospital Readmission Expense Using Cascading

10 min read

In this article by Michael Covert, author of the book Learning Cascading, we will look at a system that allows for health care providers to create complex predictive models that can assess who is most at risk for such readmission using Cascading.

(For more resources related to this topic, see here.)

Overview

Hospital readmission is an event that health care providers are attempting to reduce, and it is the primary target of new regulations of the Affordable Care Act, passed by the US government. A readmission is defined as any reentry to a hospital 30 days or less from a prior discharge. The financial impact of this is that US Medicare and Medicaid will either not pay or will reduce the payment made to hospitals for expenses incurred. By the end of 2014, over 2600 hospitals will incur these losses from a Medicare and Medicaid tab that is thought to exceed $24 billion annually.

Hospitals are seeking to find ways to predict when a patient is susceptible to readmission so that actions can be taken to fully treat the patient before discharge. Many of them are using big data and machine learning-based predictive analytics. One such predictive engine is MedPredict from Analytics Inside, a company based in Westerville, Ohio. MedPredict is the predictive modeling component of the MedMiner suite of health care products. These products use Concurrent Cascading products to perform nightly rescoring of inpatients using a highly customizable calculation known as LACE, which stands for the following:

Length of stay: This refers to the number of days a patient been in hospital
Acute admissions through emergency department: This refers to whether a patient has arrived through the ER
Comorbidities: A comorbidity refers to the presence of a two or more individual conditions in a patient. Each condition is designated by a diagnosis code. Diagnosis codes can also indicate complications and severity of a condition. In LACE, certain conditions are associated with the probability of readmission through statistical analysis. For instance, a diagnosis of AIDS, COPD, diabetes, and so on will each increase the probability of readmission. So, each diagnosis code is assigned points, with other points indicating “seriousness” of the condition.
Diagnosis codes: These refer to the International Classification of Disease codes. Version 9 (ICD-9) and now version 10 (ICD-10) standards are available as well.
Emergency visits: This refers to the number of emergency room visits the patient has made in a particular window of time.

The LACE engine looks at a patient’s history and computes a score that is a predictor of readmissions. In order to compute the comorbidity score, the Charlson Comorbidity Index (CCI) calculation is used. It is a statistical calculation that factors in the age and complexity of the patient’s condition.

Using Cascading to control predictive modeling

The full data workflow to compute the probability of readmissions is as follows:

Read all hospital records and reformat them into patient records, diagnosis records, and discharge records.
- Read all data related to patient diagnosis and diagnosis records, that is, ICD-9/10, date of diagnosis, complications, and so on.
  
  Read all tracked diagnosis records and join them with patient data to produce a diagnosis (comorbidity) score by summing up comorbidity “points”.
Read all data related to patient admissions, that is, records associated with admission and discharge, length of stay, hospital, admittance location, stay type, and so on.
- Read patient profile record, that is, age, race, gender, ethnicity, eye color, body mass indicator, and so on.
Compute all intermediate scores for age, emergency visits, and comorbidities.
Calculate the LACE score (refer to Figure 2). Assign a date and time to it.
Take all the patient information, as mentioned in the preceding points, and run it through MedPredict to produce these variety of metrics:
- Expected length of stay
- Expected expense
- Expected outcome
- Probability of readmission

Figure 1 – The data workflow

The Cascading LACE engine

The calculational aspects of computing LACE scores makes it ideal for Cascading as a series of reusable subassemblies. Firstly, the extraction, transformation, and loading (ETL) of patient data is complex and costly. Secondly, the calculations are data-intensive. The CCI alone has to examine a patient’s medical history and must find all matching diagnosis codes (such as ICD-9 or ICD-10) to assign a score. This score must be augmented by the patient’s age, and lastly, a patient’s inpatient discharge records must be examined for admittance to the ER as well as emergency room visits. Also, many hospitals desire to customize these calculations. The LACE engine supports and facilitates this since scores are adjustable at the diagnosis code level, and MedPredict automatically produces metrics about how significant an individual feature is to the resulting score.

Medical data is quite complex too. For instance, the particular diagnosis codes that represent cancer are many, and their meanings are quite nuanced. In some cases, metastasis (spreading of cancer to other locations in the body) may have occurred, and this is treated as a more severe situation. In other situations, measured values may be “bucketed”, so this implies that we track the number of emergency room visits over 1 year, 6 months, 90 days, and 30 days.

The Cascading LACE engine performs these calculations easily. It is customized through a set of hospital supplied parameters, and it has the capability to perform full calculations nightly due to its usage of Hadoop. Using this capability, a patient’s record can track the full history of the LACE index over time. Additionally, different sets of LACE indices can be computed simultaneously, maybe one used for diabetes, the other for Chronic Obstructive Pulmonary Disorder (COPD), and so on.

Figure 2 – The LACE subassembly

MedPredict tracking

The Lace engine metrics feed into MedPredict along with many other variables cited previously. These records are rescored nightly and the patient history is updated. This patient history is then used to analyze trends and generate alerts when the patient is showing an increased likelihood of variance to the desired metric values.

What Cascading does for us

We chose Cascading to help reduce the complexity of our development efforts. MapReduce provided us with the scalability that we desired, but we found that we were developing massive amounts of code to do so. Reusability was difficult, and the Java code library was becoming large. By shifting to Cascading, we found that we could encapsulate our code better and achieve significantly greater reusability. Additionally, we reduced complexity as well. The Cascading API provides simplification and understandability, which accelerates our development velocity metrics and also reduces bugs and maintenance cycles.

We allow Cascading to control the end-to-end workflow of these nightly calculations. It handles preprocessing and formatting of data. Then, it handles running these calculations in parallel, allowing high speed hash joins to be performed, and also for each leg of the calculation to be split into a parallel pipe. Next, all these calculations are merged and the final score is produced. The last step is to analyze the patient trends and generate alerts where potential problems are likely to occur.

Cascading has allowed us to produce a reusable assembly that is highly parameterized, thereby allowing hospitals to customize their usage. Not only can thresholds, scores, and bucket sizes be varied, but if it’s desired, additional information could be included for things, such as medical procedures performed on the patient. The local mode of Cascading allows for easy testing, and it also provides a scaled down version that can be run against a small number of patients. However, by using Cascading in the Hadoop mode, massive scalability can be achieved against very large patient populations and ICD-9/10 code sets.

Concurrent also provides an excellent framework for predictive modeling using machine learning through its Pattern component. MedPredict uses this to integrate its predictive engine, which is written using Cascading, MapReduce, and Mahout. Pattern provides an interface for the integration of other external analysis products through the exchange of Predictive Model Markup Language (PMML), an XML dialect that allows many of the MedPredict proprietary machine learning algorithms to be directly incorporated into the full Cascading LACE workflow. MedPredict then produces a variety of predictive metrics in a single pass of the data. The LACE scores (current and historical trends) are used as features for these predictions. Additionally, Concurrent provides a product called Driven that greatly reduces the development cycle time for such large, complex applications. Their lingual product provides seamless integration with relational databases, which is also key to enterprise integration.

Results

Numerous studies have now been performed using LACE risk estimates. Many hospitals have shown the ability to reduce readmission rates by 5-10 percent due to early intervention and specific guidance given to a patient as a result of an elevated LACE score. Other studies are examining the efficacy of additional metrics, and of segmentation of the patients into better identifying groups, such as heart failure, cancer, diabetes, and so on. Additional effort is being put in to study the ability of modifying the values of the comorbidity scores, taking into account combinations and complications. In some cases, even more dramatic improvements have taken place using these techniques. For up-to-date information, search for LACE readmissions, which will provide current information about implementations and results.

Analytics Inside LLC

Analytics Inside is based in Westerville, Ohio. It was founded in 2005 and specializes in advanced analytical solutions and services. Analytics Inside produces the RelMiner family of relationship mining systems. These systems are based on machine learning, big data, graph theories, data visualizations, and Natural Language Processing (NLP). For further information, visit our website at http://www.AnalyticsInside.us, or e-mail us at info@AnalyticsInside.us.

MedMiner Advanced Analytics for Health Care is an integrated software system designed to help an organization or patient care team in the following ways:

Predicting the outcomes of patient cases and tracking these predictions over time
Generating alerts based on patient case trends that will help direct remediation
Complying better with ARRA value-based purchasing and meaningful use guidelines
Providing management dashboards that can be used to set guidelines and track performance
Tracking performance of drug usage, interactions, potentials for drug diversion, and pharmaceutical fraud
Extracting medical information contained within text documents
Designating data security is a key design point
PHI can be hidden through external linkages, so data exchange is not required
If PHI is required, it is kept safe through heavy encryption, virus scanning, and data isolation
Using both cloud-based and on premise capabilities to meet client needs

Concurrent Inc.

Concurrent Inc. is the leader in big data application infrastructure, delivering products that help enterprises create, deploy, run, and manage data applications at scale. The company’s flagship enterprise solution, Driven, was designed to accelerate the development and management of enterprise data applications. Concurrent is the team behind Cascading, the most widely deployed technology for data applications with more than 175,000 user downloads a month. Used by thousands of businesses, including eBay, Etsy, The Climate Corporation, and Twitter, Cascading is the defacto standard in open source application infrastructure technology. Concurrent is headquartered in San Francisco and can be found online at http://concurrentinc.com.

Summary

Hospital readmission is an event that health care providers are attempting to reduce, and it is a primary target of new regulation from the Affordable Care Act, passed by the US government. This article describes a system that allows for health care providers to create complex predictive models that can assess who is most at risk for such readmission using Cascading.