wtf is readmission_project?

This project tries to predict which diabetic patients are likely to be readmitted to the hospital within 30 days of being discharged, using patient records instead of guesswork. Hospital readmissions are expensive and often a sign that a patient's care did not go as planned, so the goal is to flag high risk patients at discharge time so care teams can follow up with them first. The data comes from a public dataset of over 100,000 inpatient hospital visits across 130 US hospitals between 1999 and 2008. After removing duplicate visits and patients who died or entered hospice care, the working dataset has about 69,650 unique patient encounters, of which about 7 percent were readmitted within 30 days. The project runs as a series of steps. First it cleans the data and builds new features, such as turning age ranges into numbers and grouping diagnosis codes into broader categories. Next it loads the cleaned data into a small SQL database and runs queries to check patterns, like how readmission rates differ by age or by how many times a patient was hospitalized before. Then it trains three different prediction models, Logistic Regression, Random Forest, and XGBoost, and picks the best one based on a standard accuracy measure. Finally, a Streamlit dashboard lets anyone explore the cohort data, compare the models, and calculate an individual patient's risk score interactively. The best model, XGBoost, reached a score of about 0.68 on this measure, which the README says lines up with other published research on this same dataset, since predicting 30-day readmission is a genuinely hard problem given the data available. The models were tuned to catch as many true readmission cases as possible, since missing a high risk patient matters more than a false alarm. The biggest factors driving risk were prior hospital visits, length of the current stay, where the patient was discharged to, and their main diagnosis category.