Correlation Finder

Below is the summary of this research and full paper. 

Introduction 

A great number of medical decisions are made in hospitals everyday and a lot of medical errors are also made. It is reported that one of the major factors contributing to medical errors is inexperienced and young clinicians.  To support such inexperienced clinicians and decrease the probability of medical errors, we introduce Correlation Finder, a clinical decision support system that uses a statistical method to find implicit correlations between mutually exclusive medical activities. 

In general, healthcare processes are made up of a series of medical activities. Clinicians oftentimes face situations where multiple medical activities are mutually exclusive and a medical activity has to be chosen over the others. However, in many cases, inexperienced clinicians are not sure which option he or she has to select. For such situations, Correlation Finder runs a T-Test, a statistical method to decide whether two sets of data are significantly different, for all medical variables of medical activities in question to find implicit or hidden correlation between multiple medical activities, which inexperienced clinicians would otherwise not notice, thus it would result in a medical error. In this way we believe that Correlation Finder can help inexperienced clinicians make right medical decisions and feel confident in their decision. 

Correlation Finder

Fig1. Healthcare processes are composed of several medical activities(indicated as MA) and some of them are mutually exclusive. 

Fig2. Once a medical activity has been chosen, a corresponding medical action is taken. In this example, the medical action is making a diagnosis and giving a corresponding treatment. 

T-Test

T-Test, also known as Student’s T-Test, is one of the most widely used statistical methods used to see if two sets of data differ significantly such as whether men and women have different average heights. Significantly means that two sets of data are not likely to be different by accident. 

Use Cases

• Correlation between Subset of Historic Cases and Current Medical Activities: 

In case of Correlation between Previous and Current Medical Activities, if the result of the previous medical activity is similar, thus gives no hint in selecting medical options, Correlation Finder considers the subset of historic cases preceding the previous medical activity. For example, as described in Figure 3, if the result of MA9 is always around 8g (in this case, tumor weight), and thus lacks in its discrimination function, Correlation Finder considers the subset of preceding historic cases (MA7 and MA8) and how many of each case went though each of current medical activities (MA10 and MA11). 

92% of MA7 and 13% of MA8 led to MA10.  

8% of MA7 and 87% of MA8 led to MA11. 

Fig3. Correlation Finder considers how the subset of historic cases (MA7 and MA8) affects the current medical activities (MA10 and MA11). 

• Use Case for Diagnosis-Treatment

Fig4. Correlation Finder displays detailed information on the selected treatment option. 

Implementation

Fig5. The implementation sequence of Correlation Finder

Fig6. Left top: excel data consists of several sheets which represent each medical activity and every medical activity contains patient information, diagnosis and treatment history. Right top: data structure is modified so that it is easier to find relations between medical activities. Bottom: from the modified data structure, the overall flow is drawn. 

Conclusion

We introduced Correlation Finder, a clinical decision support system that uses a statistical method to detect implicit or invisible correlations between mutually exclusive medical activities which inexperienced clinicians may otherwise overlook. Correlation Finder runs T-Test for all medical variables and finds if any of them have significantly different values between multiple medical activities. Once a medical activity has been selected, a diagnosis and corresponding treatment are made. As a single diagnosis has multiple treatment options, Correlation Finder displays possible treatment options that are suitable to a specific diagnosis for a specific patient condition. That said, current Correlation Finder has a limit. It only works for two mutually exclusive medical activities. A possible way to solve this problem can be to split a set of mutually exclusive medical activities into two subsets with all possible combinations and run T-Test for the two subsets. In this way Correlation Finder finds two subsets that have significantly different values. Also, collecting and accessing historic data is important to Correlation Finder. In order to achieve this, leveraging billing process can be considered. It is essential to document diagnosis and treatment history for hospital records as well as health insurance claims. In general, ICD (Inter- national Classification of Diseases) code is widely use for diagnosis and OPS (Operationen- und Prozedurenschluessel) code is normally used in Germany for treatment. These codes are key information for billing processes. Using these codes, accessing historic data will be easily achieved.