Prediction of deaths due to colorectal cancer in turkiye using machine learning

The primary objective of this study is to forecast region-based colon cancer mortality figures for the period 2025-2030 by utilizing mortality data from 2015-2024 in Turkey with the XGBoost machine learning algorithm. Colon cancer stands as a significant public health issue, being one of the most common cancers with high mortality rates both globally and in Turkey. Accurate forecasting of the future mortality burden is critical for developing health policies, allocating resources effectively, and targeting preventive strategies. Within the scope of this thesis, a multivariate dataset including the variables olumtarihi (date of death), cinsiyet (gender), yas_grup (age group), and bolge_adi (region name) was used. Considering the limitations of traditional time series models in modeling complex and non-linear relationships, the eXtreme Gradient Boosting (XGBoost) algorithm, known for its high predictive accuracy and flexibility, was preferred in this study. The time series data were transformed into a supervised learning problem using feature engineering techniques. In this process, lag features, rolling window statistics, and time-based (calendar) features were derived from historical data to enable the model to learn temporal dependencies, seasonality, and trends. Categorical variables (cinsiyet, yas_grup, bolge_adi) were converted into a numerical format using the One-Hot Encoding method. A recursive forecasting strategy was adopted to solve the multi-step forecasting problem. In this strategy, the model, trained to make single-step predictions, sequentially generates future values by using its own predictions as input for the subsequent step. The model's performance was rigorously evaluated using Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE) metrics with time-series cross-validation. The resulting forecasts reveal the expected colon cancer mortality burden in different geographical regions of Turkey between 2025 and 2030. These results, by highlighting regional differences and potential increases in risk, provide evidence-based, strategic insights for public health planners and policymakers.