Dalikmata-Ultima: Revolutionizing Medical AI Diagnosis for Skin Cancer and Pneumonia with Hierarchical Classification, Model Hybridization and OOD Evaluation

Global healthcare systems face critical workforce and diagnostic gaps, necessitating scalable AI solutions. This article presents Dalikmata-Ultima, a framework employing two distinct methodologies for dermatological and radiological diagnosis grounded in the Universal Approximation Theorem and VC bounds for structural reliability. For dermatological diagnosis, the system utilizes a hybrid CNN-MLP architecture to integrate EfficientNetV2 image features with patient metadata. Trained and evaluated using the ISIC 2019 dataset, the framework employs the Forced Dominance Transform (FDT) to synthesize these inputs while maintaining a clinical hierarchy that separates primary malignancy screening from subtype classification. For pneumonia detection, a vision-only CNN pipeline was implemented using the Mendeley Chest X-ray dataset. Due to the absence of structured metadata, hybridization was omitted to maintain a data-dependent design. This model underwent out-of-distribution (OOD) validation using clinical data from Pasig City Children's Hospital (PCCH), addressing a critical gap in Philippine medical AI research and proving the feasibility of deploying contextually relevant diagnostic support in real-world settings.