Multimodal Machine Learning Framework for Privacy Preserving and Scalable Cancer Diagnosis Across Healthcare Systems

Cancer remains a major global health challenge, where early risk identification plays a crucial role in improving patient outcomes and supporting clinical decision-making. This study presents a comprehensive machine learning framework for cancer prediction using a structured clinical dataset comprising 1500 patient records with demographic, physiological, lifestyle, and genetic risk factors. The dataset includes both continuous and categorical variables, requiring appropriate preprocessing and encoding to ensure compatibility across diverse classification models.A systematic evaluation of ten machine learning algorithms is conducted, including Logistic Regression, Decision Tree, Random Forest, Support Vector Machine, k-Nearest Neighbors, Naive Bayes, Gradient Boosting, Extreme Gradient Boosting, Light Gradient Boosting Machine, and CatBoost. Model performance is assessed using five-fold cross-validation to examine stability, followed by evaluation on an independent held-out test set using accuracy, precision, recall, and F1-score.The experimental results demonstrate that ensemble-based boosting methods consistently outperform traditional and linear models. CatBoost achieves the best overall performance, with a cross-validation accuracy of 0.944 ± 0.006 and a test accuracy of 0.9667, along with high precision and recall, indicating strong generalization and reliable detection capability. Comparative analysis using confusion matrices further highlights that tree-based ensemble models reduce false negative rates, which is critical for medical diagnosis. In contrast, distance-based and probabilistic models such as k-Nearest Neighbors and Naive Bayes show reduced sensitivity to positive cases.Correlation analysis indicates that cancer history, genetic risk, smoking behavior, and demographic factors contribute meaningfully to prediction, while physical activity exhibits a negative association with cancer occurrence. Despite the strong predictive performance, the study acknowledges that the dataset is synthetically generated, which may limit direct clinical applicability.