Compressing ICD9 codes for predictive modeling

There have been many attempts to rank the quality of healthcare providers, particularly hospitals. The definition of quality must take into consideration the severity of a patient's condition. Sicker patients will be more at risk for complications, and mortality. Without considering patient severity, hospitals might only admit low risk patients to improve quality rankings, while transferring their more severe patients. The problem is to define what patient severity really means. It is the purpose of this module to discuss the development of patient risk adjustment. We will discuss typical adjustment methods that rely upon a specific list of patient diagnoses. Next, we will discuss a method that uses all possible combinations of diagnoses. Risk adjustment is based upon the diagnoses that define the patient's condition. A patient with diabetes and congestive heart failure, for example, will be at much higher risk for any procedure compared to a patient with no severe diagnoses. There has to be some way of choosing which diagnoses to use for risk adjustment, and which should be discarded. The Charlson Index, for example, uses a list of 17 major diagnoses and assigns weights of 1 to 3 to each diagnosis. The severity score for one patient is a sum of the weights of the 17 diagnoses that the patient has. Patients in a higher risk category should have a higher predicted mortality rate, and higher treatment costs. Otherwise, the risk adjustment has questionable results. Usually, risk is defined for each individual procedure as well, since some procedures have higher risk compared to others. Once the risk adjustment is determined, regression is used to predict outcomes, either mortality, morbidity, or cost. Providers are ranked based upon the difference between the actual outcome and the predicted outcome. Problems occur because different providers do not code diagnoses uniformly. Over-coders are rewarded because they successfully shift patients into higher risk categories, reducing the difference between actual and predicted outcomes. It would be preferable to develop an index that uses all of the possible patient diagnoses, and all possible combinations of diagnoses. However, there has to be some method used to compress all of these different combinations into a meaningful index of severity so that the adjusted risk can be used to predict patient outcomes. In this module, we will demonstrate such a technique, and discuss its superiority as a tool of risk adjustment.