Clinical Translatable and Transparency of Neural Network for Blood-Based Personalized Cancer Risk

To support cancer screening and identify precancerous conditions, such as atypical hyperplasia, to improve cure rates and reduce mortality, by analyzing the performance of a previously confirmed procedure.

We established 204 short-term blood-derived cell lines from patients with cancer between December 1, 2013 and December 31, 2022. A dataset of phenotypic patterns, cytopathological variables, and proliferation profiles was used to train a neural network model. Comparative analysis of standard optical, functional, and machine learning supported diagnosis was performed to verify reproducibility and clinical translatability.

Tumor heterogeneity was classified into 7 phenotypic patterns (Pn1-Pn7); the Pn6-7 group showed an overall survival of 8 months (95% CI, 6-9). Among variables Vc1-Vc8, Vc3 (area under the curve=1, P<.001) and Vc5-6 (area under the curve=0.838, P<.001) were identified as discriminant for cell atypia (rho=0.5 correlation with biopsy). Proliferative ranges were 0%-30% healthy, 30%-35% hyperplasia, and >35% cancer. Artificial intelligence-supported cytology showed positive and negative predictive values of 0.99±0.015 and 1±0 compared with histopathological specimens (0.94±0.1 and 0.85±0.04). The algorithm achieved a 1±0 sensitivity and a 0.98±0.04 specificity, with respect to traditional diagnosis (specificity [0.875-0.923] and sensitivity [0.926-0.984]).

The model demonstrated fast adaptive performance in predicting cancer risk and primary source assessment. The results suggest that this screening model is sufficient to detect atypical hyperplasia compared with models based on oligoanalysis for single or double mutations.