Predicting Postoperative Delirium

The goal of this experiment is to be a model for a larger, more dynamic study, with the aim to be able to predict delirium. Delirium is a relatively common postoperative condition in patients and lacks any objective screening tool, or effective diagnosis technique. Therefore, in designing an accurate algorithm using historical and dynamic patient data, it is our goal to recognize and identify the possibility of delirium in patients. **1. Pre-Experimental Research** Before beginning the module, we studied heavily into delirium. We began by researching past studies on delirium in the public domain and their findings. We also conducted our own personal research using retrospective analysis on historical data from various hospitals and patients. Comparing the two sets of results, we gathered a set of risk factors and warning signs for delirium that we found to have a significant research backing. **2. Data Extraction** Data was extracted from a database using SQL queries and Excel tables. **3. Data Specification** The experiment contains 8 patient data categories in which data can be inputted, including: 1. Diastolic blood pressure 2. MAP blood pressure 3. Glucose level 4. Oxygen Saturation 5. P/A of renal failure 6. P/A of depression 7. P/A of pyschosis 8. P/A of alcohol abuse These categories were chosen out of 66 total possible fields and are the risk factors and warning signs that we have found can accurately identify delirium. **4. Creating Experiment** This experiment utilizes a two-class boosted decision tree, which we have found produces the most accurate results. The data is first inserted and preprocessed using the "Clean Missing Data" module. Then, columns from the 8 chosen categories listed above are selected and uniquely included in the experiment. Eight trees were then created off of that module to represent the categories, each of them containing a split condition chosen based on our research. In splitting the data, it is possible to see the results of the condition and if it has accurate results. This is how we knew to exclude data that did not prove to be relevant to the experiment.