Methods of Knowledge Discovery in Healthcare
Task 1
Use your methodology of choice to answer the following research questions:
1. What is the most common disease for each age group? What is the prevalence of the top three diseases for each age group? Consider that each patient has up to two diagnosis codes.
2. Compare (i) the in-hospital mortality of men and women and (ii) the in-hospital mortality for each of top three diseases between men and women.
3. (i) Are there any demographic factors that are found to be different between long and short hospital stays? (ii) What is the most common long stay primary diagnosis and which is the most common short stay primary diagnosis
4. What is the effect of the length of stay to (i) the total cost (ii) the in-hospital mortality?
5. Investigate the relationship between the Discharge Destination and the Age Group.
Task 2
1. Observe the data and design an appropriate relational schema. Then create the schema on the DBMS system of your choice (preferably MySQL server). The schema should have the following merits:
a. Normalization principles should be applied to avoid duplicates
b. Appropriate Data types should be defined
2. Import the data into your newly developed schema
3. Create the following queries
a. An appropriate query which returns a result which is identical to the given csv file. This way you are demonstrating how one can extract data from an Electronic Medical Record database, to use for data analysis.
b. Queries which return the Coverage Ratio (Coverage Ratio = COVERED_CHARGES/TOTAL_CHARGES) of patients who stayed in the hospital for a period of time longer than 5 days. How does this compare to the Coverage Ratio of patients with a Long Stay?
c. Is there any variation in the average Length of Stay of patients admitted to the hospital in different days of the week (DAY_OF_ADMISSION)? Showcase this with an appropriate SQL query and design an appropriate graph comparing the average Length of Stay between Friday admissions (DAY_OF_ADMISSION=6) and Monday admissions (DAY_OF_ADMISSION=2). Discuss possible reasons that may contribute to what you have found.
d. Using an appropriate method, explore a possible relationship between the DRG_PRICE and the TOTAL_CHARGES. Is there a linear relationship between these two properties?
Task 3
Using the Weka implementation of k-means, please find out:
(a) The appropriate number of clusters which are required to adequately ‘describe' the discharge characteristics of the patients (discharge destination, discharge status, stay indicator). Use the elbow method to define the number, by evaluating the ‘within cluster sum of squared errors' you get as a result in your Weka output. Draw an appropriate graph to explain your answer.
(b) Based on the number of clusters you specified in the above step, please calculate those clusters.
(c) Briefly discuss two interesting (in your opinion) profile groups you have just found.
(d) Is the method we have followed a supervised or an unsupervised data mining technique? Please explain your answer.
Task 4
We need to use data mining to predict whether the Diagnosis Related Code (DRG) price of an admission is going to be higher or lower than $80,000. Make sure your features are not in string data type.
(a) Use any appropriate method to modify the class attribute values to be only of two values, either zero (DRG price less than $80,000) or one (DRG price more than $80,000) so that the problem will be binary classification. Integrate the new attribute (DRG_PRICE_BINARY) into your dataset.
(b) Observe the available features and explain how the data are acquired in a temporal manner during the healthcare procedure in the real hospital. Specifically, define what do clinicians/administrators already know:
1. At the time when the patient enters the hospital
2. At the time when the patient is discharged from the hospital
(c) Answer, for the two above scenarios, the following questions:
1. Why did we exclude the attributes 7, 10 and 11?
2. Undergo the feature selection process CfsSubsetEval to select the appropriate features for each of the two scenarios. This way, you will be able to know which features will be included in your classification later on.
3. Use the classifiers (i) Naïve Bayes and (ii) Logistic Regression to classify the DRG_PRICE_BINARY, for each scenario, by using the features you found to be useful during feature selection.
4. Discuss the accuracy of the classification for the two classifiers in each scenario, in terms of:
4a. the overall accuracy
4b. the accuracy prediction of expensive (>$80,000) DRG costs
4c. the accuracy prediction of not so expensive (less than $80,000) DRG costs.
Attachment:- Dataset.csv