Machine Learning for Lung Cancer Subtype Classification: Combining Clinical, Histopathological, and Biophysical...

Background/Objectives: Despite advances in diagnostic techniques, accurate classification of lung cancer subtypes remains crucial for treatment planning. Traditional methods like genomic studies face limitations such as high cost and complexity. This study investigates whether integrating atomic force microscopy (AFM) measurements with conventional clinical and histopathological data can improve lung cancer subtype classification. Methods: We developed and analyzed a novel dataset combining clinical, histopathological, and AFM-derived biophysical characteristics from 37 lung cancer patients. Various machine learning techniques were evaluated, with a focus on Bayesian Networks due to their ability to handle complex data with missing values. Leave-One-Out Cross-Validation was employed to assess model performance. Results: The integration of biophysical features improved classification accuracy from 86.49% to 89.19% using a data-driven Bayesian Network model, though this improvement was not statistically significant (p = 1.0). Four key features were identified as highly predictive: sex, vascular invasion, perineural invasion, and ALK mutation. A simplified model using only these features achieved identical performance with significantly reduced complexity (BIC 51.931 vs. 268.586). Conclusions: While AFM-derived measurements showed promise for enhancing lung cancer subtype classification, larger datasets are needed to fully validate their impact. Our findings demonstrate the feasibility of incorporating biophysical measurements into cancer classification frameworks and identify the most predictive features for accurate diagnosis. Further research with expanded datasets is needed to validate these findings.