TS-Fingerprint

TS-Fingerprint outperformed compared methods on average rank across datasets and metrics. The method was reported to be faster at inference than Medformer. TS-Fingerprint represents each medical time series with a fixed set of latent tokens rather than a variable-length token sequence. Its architecture uses learnable queries and cross-attention to map patch sequences into Fingerprint Tokens. The fixed-rank bottleneck prevents representation length from scaling with input length.