Abstract 35P
Background
Some tumour types carry specific genomic alterations, and genomic alterations across numerous different cancers may guide the inference of tumour origin. Therefore, we have explored the feasibility of identifying tumour type based on genomic alterations.
Methods
To predict tumour site of origin, 11867 cases were included in this study. We constructed a random forest classifier using a training cohort of 9493 patients representing 21 cancer types, and the best parameters were determined from 5-fold cross-validation of the training data. Genomic profiling of DNA was performed on formalin-fixed paraffin-embedded tumour samples through NGS with a panel of 381 cancer-related genes.
Results
In our test set of 2374 patients, we accurately predicted tumour type in 1163 cases (49.0% of cases) based on 5-fold cross-validation. The positive predictive value was highest in tumour types with distinctive molecular profiles, or larger sample size (e.g. lung cancer).
Conclusions
These results suggest that the application of artificial intelligence has the potential to predict tissue of origin for a tumour by using mutation data.
Clinical trial identification
Editorial acknowledgement
Legal entity responsible for the study
Shanghai Chest Hospital.
Funding
Has not received any funding.
Disclosure
All authors have declared no conflicts of interest.