Healthcare providers are always looking for ways to improve the quality of patient care and identify best practices to optimize the revenue cycle. A smarter and more efficient way to do this is to use predictive analytics to find patterns in large quantities of data collected in clinical information systems, electronic medical records and revenue cycle information systems.

Jeff Kaplan

These insights can help providers proactively identify opportunities to improve clinical and financial outcomes, reduce medical errors, and increase patient satisfaction. A few examples include:

Reducing medical errors—flag potentially critical clinical errors or drug combinations and deviations from clinical best practices

Denial management—predict in advance claims likely to be denied, before they are sent to payers

Underpayment prediction—identify claims likely to be underpaid by a payer, based on specific claim attributes

Capacity prediction—predict when divert situations are likely and accurately forecast demand for services

Financial impact of clinical practices—identify key correlations between clinical practices and best financial outcomes

Construct "what if" scenarios—model scenarios based on demand shifts, updated payer contracts, and even clinical process changes

Infection prevention—model, predict, and find sources of hospital-acquired infections

Payer negotiation—identify key drivers for underperforming payers

Uninsured patient payment prediction—prioritize self-pay claims most likely to pay in a timely fashion.

Predictive analytics implementation should follow a six-step process:

Step 1—Define success

First, define the goal of the data mining project, whether it be to boost cash flow through decreased denials, or to proactively identify medical errors. Next, determine how you are going to measure the project's success, which may be directly related to the impact of the predictive model (e.g. divert situations reduced by 15 percent).

Finally, define the outputs of the model (e.g. Excel reports, XML interface to integrate with a business office work-driver) and determine how frequently the model should be refreshed.

Step 2—Identify required resources

Determine the best data mining approach to reach project goals, including the required data sources and potential modeling approaches. Then, map business requirements to the data sources collected by the patient accounting system, clinical information system, or whichever data is necessary to support the goals. The most important question to ask is, "Does the data collected support the predictive model needed to achieve our project's goals and success criteria?"

The result of the first two phases should be a clearly defined project roadmap, which includes detailed descriptions of the data sources, timeline for acquiring and integrating data, potential modeling approaches, the automation plan, and ROI measurement.

Step 3—Format and integrate

Phase three is about acquisition, analysis, and cleaning of the data. Your data gurus should work closely with the statistical modeling resources to audit the data and identify transformations required to support optimal data mining. Having clean data sources to support data mining is paramount to success.

Hospital data tends to be collected in disparate silos, and frequently contains incomplete patient information, misspellings, and out-of-range values. There is no magic bullet for data cleaning, other than close analysis of the data to ensure it meets the expectations outlined in the requirements.

Step 4—Design and develop

The actual predictive modeling begins: take an iterative approach to building models, by regularly sharing results with key project stakeholders to ensure the analysis are meeting project goals. It is also important to choose the right algorithms for the application: "tuning" algorithms for optimal performance will maximize the accuracy and performance of your data mining models.

Also, consider algorithm performance in this phase, as depending on the automation processing requirements it may be necessary to trade-off increases in accuracy for a boost in performance. There are various model evaluation techniques which can help the team measure the accuracy and efficacy of the model, the most common being a technique called cross-fold validation, which is a method for estimating the predictive accuracy when applying a given model to new data, based upon multiple sampling and testing on a given dataset.

Step 5—Validate and review

Identify the most accurate and relevant data mining model(s) and run tests to ensure it is accurate over existing data. It's essential to pre-test your models with a sub-set of your patient base to ensure the models are properly scoring and the system is running in a timely manner. As results and project goals are reviewed, the team may also identify special organization-specific business logic to be added on top of the models.

Step 6—Deploy and test

The core value of most data mining applications is automating the process of updating the models continuously with new data; then reusing and leveraging the process for future data mining goals. The automation plan designed in Step Two can be carried out by the database and system architects. Additionally, the output of the models, whether a report or integration with an existing information system, should be finalized, tested, and deployed.

The automation of data mining models can be made efficient through the use of technology solutions that integrate data import and load functions, relational databases, and predictive algorithms. Rather than a piecemeal approach, an integrated solution offers a consistent platform for the data analysts, statistical modelers, and database developers to collaborate.