SVM Model with Hybrid Parameter Tuning Strategy for Colon Cancer Classification in Nairobi County

Abstract

Colon cancer remains one of the leading causes of cancer-related deaths worldwide, with recent evidence pointing to a rising incidence in Nairobi County, Kenya. This study designs and evaluates a region-specific Support Vector Machine (SVM) classification model with a hybrid parameter-tuning algorithm for accurate colon cancer diagnosis in Kenyan hospitals. The work addresses diagnostic delays driven by a shortage of pathologists, lengthy manual slide reviews, and limited access to advanced procedures. Using the HIPAA-compliant LC25000 histopathological image dataset, we applied feature selection, normalization, and image augmentation to improve model robustness. Our two-stage hybrid hyperparameter-tuning strategy first performed a coarse Grid Search (C ∈ {0.1, 1, 10, 100}, γ ∈ {10⁻⁶, 10⁻⁵, 10⁻⁴}) followed by a Random Search (C ∈ [0.01, 100], γ ∈ [10⁻⁷, 10⁻⁵]) on the intersection of the top 10% configurations, enabling dense sampling of promising ranges. This approach leveraged SVM strengths in handling small, high-dimensional datasets, applying robust regularization, and weighting classes to address imbalance while reducing cross-validation variance by 15–20%. Compared to standalone grid and Random Search methods, the hybrid model achieved 83.0% accuracy (improvements of 1.5% and 2.0%), an F1-score of 82.4%, malignant recall of 81.0%, and benign precision of 84.0%. In practice, these gains translate to dozens more correctly classified slides per 4,000, representing an impactful improvement for resource-limited clinics in Nairobi. At Kenyatta National Hospital (KNH), the SVM model correctly identified 73 of 77 positive cases, yielding 94.81% sensitivity (recall) and minimizing missed colon cancer diagnoses. These findings demonstrate that tailored, efficient machine learning models can strengthen diagnostic capacity where expert resources are scarce. Future work includes collecting 200 balanced KNH images (100 benign, 100 malignant) for proper validation, implementing Bayesian optimization with an expected improvement acquisition function, and training on combined histopathology and clinical data.