Non-invasive screening of lung-cancer in the early stages through breath detectors would demand light-gas overlaps and excellent pattern-recognition capabilities of the sensors. We introduce a hybrid diagnostic system that combines hollow-core photonic crystal fiber (HC-PCF) spectroscopic sensor with intelligent classifiers. HC-PCF (core 2 μm, Λ = 3 μm, d = 1.5 μm, 4 rings) in COMSOL is used to measure changes in the effective index, mode confinement, and changes in transmission at controlled volatile organic compound (VOC) concentrations. Synthetic and simulation-derived spectra (wavelength vs. intensity, VOC label, concentration) train two complementary models: a feature-engineered Random Forest (RF) and an end-to-end 1D CNN. With a 70/15/15 train/val/test split, the CNN (AUC = 0.96; sensitivity = 91.8% specificity = 95.7) and the RF (AUC = 0.94) both achieve accuracy rates of 94.1% and 92.4% respectively. Five-fold CV indicates a variance of ≤ ±1.4% and a reduction in performance of less than 3% when under noise stress, which indicates robustness. Proposed method is competitive compared to recent breath-VOC baselines that report pooled sensitivity/specificity 85–86 and AUC 0.93, and allows compact, fiber-based optics to be used in the point-of-care setting. The contributions include an HC-PCF sensing model (COMSOL validated) of VOCs, a dual-path ML pipeline (engineered features + raw spectra) and an empirical performance and robustness analysis that satisfies clinical screening requirements.