#16 · ML / AI methods

Input selection and fusion strategies govern biomedical ML prognostic performance

The dominant thread is how inputs are structured, selected and combined before learning. In prostate cancer, adding weak histology, radiology and clinical modalities degrades biochemical recurrence-free survival accuracy, so additional inputs are not uniformly beneficial. Copula fusion of clinical and gene-expression risk scores yields no AUC gain over the clinical model for breast cancer mortality, reinforcing this caution. On the whole-slide side, TDAM-CRC links MIL features from colorectal cancer images to the multi-omics hub gene MRPL37. Single-cell items address related questions: a ZINB-based denoiser with InfoVAE-GAN latent factors, a fusion of four feature types across six strategies for cell typing, and a benchmark of 16 deep learning integration methods. What is new is explicit evaluation of input composition and fusion strategy against baseline models, rather than architecture alone.