Table 1.
Performance of CUP classification model published previously.
| Refs. | Year | Data type | Method |
Performance of external validation dataset |
||
|---|---|---|---|---|---|---|
| Accuracy | Validation tumor | # of tumor types | ||||
| [21] | 2011 | RT-PCR | K-nearest neighbor | 83 % (187) | P + M | 28 |
| [33] | 2011 | Microarray | Machine learning | 88.5 % (462) | P + M | 15 |
| [19] | 2016 | DNA methylation microarray | Random forest | 94 % (534) | M | 21 |
| [32] | 2020 | Targeted DNA sequencing | Random forest | 74.1 % (11644) | P + M | 22 |
| [22] | 2020 | DNA methylation microarray | deep neural network | not measured (581) | P+M | 10 |
| [17] | 2020 | Gene expression | 1d-inception | 86.96 % (23) / 72.46 % (69) | M | 6 / 18 |
| [18] | 2020 | Whole genome sequencing | deep neural network | 88 %P / 83 %M (2120) | P + M | 16 |
| [11] | 2021 | Whole slide image | multitask neural network | 79.9 % (682) / 61 % CUP (317) | M + CUP | 17 |
| [20] | 2022 | Whole genome sequencing | Random forest | 58 % CUP (141) | CUP | - |
| Our model | 2023 | DNA methylation microarray | Vision transformer | 96.4 %P (693) / 94.4 %M (302) | P + M | 14 |
RT-PCR: Reverse transcription polymerase chain reaction; P: primary tumor; M: metastatic tumor; CUP: cancer unknown primary