SmellSense AI: A Proposed Hybrid Deep Learning Framework for Non-Invasive Disease Diagnosis via Breath-Based Volatile Organic Compound Analysis

Early detection of lung cancer, diabetes mellitus, and chronic respiratory disorders still depends on biopsies, imaging, and blood-based assays that are invasive, costly, and slow. Exhaled breath carries volatile organic compounds (VOCs) that reflect metabolic and pathological change, and offers a non-invasive alternative. This paper proposes SmellSense AI, a hybrid framework that couples a three-element metal-oxide semiconductor (MOS) sensor array with a Deep Belief Network (DBN) and a Bidirectional Gated Recurrent Unit (Bi-GRU), with hyperparameters tuned by a Sine-Cosine Marine Predator (SCMP) optimizer. The paper specifies the acquisition hardware and a fixed 150 mL/min sampling flow, a baseline-corrected preprocessing pipeline, a 50-dimensional feature vector built from IMFCC, BFCC, and DWT streams, and the SCMP search space and fitness function. It also defines a pilot protocol for 30 participants across healthy, diabetic, and respiratory/oncological cohorts. The proposed framework will be benchmarked against conventional single-model machine learning classifiers, namely Support Vector Machines (SVM) and Random Forest, using macro F1-score and ROC-AUC. The contribution at this stage is methodological: a complete, reproducible specification to guide ethics review, data collection, and empirical validation.