Speaker
Description
The adoption of artificial intelligence in healthcare is increasing rapidly, yet only a small number of applications have reached routine clinical use. A central obstacle is the lack of transparency, as many machine learning models operate as black boxes that produce predictions without interpretable reasoning. This contribution explores an alternative approach based on geometric data representations. Patient trajectories are embedded into a low-dimensional manifold that is learnt directly from the clinical measurements using the UMAP approximation algorithm \cite{mcinnes_umap_2020}. This replaces the opaque model parameters with a structure that remains interpretable within the original feature space. The method is designed for time-dependent intensive care unit data and introduces a semi-metric on trajectories inspired by the discrete Fréchet distance \cite{Eiter1994ComputingDF}, which measures similarity between a new patient trajectory and a disease-specific manifold representation. This allows both disease membership testing, which evaluates whether a trajectory is consistent with a given condition, and outcome prediction by comparing trajectories to manifolds trained on surviving and non surviving patient groups.
Bibliography
@misc{mcinnes_umap_2020,
title = {{UMAP}: {Uniform} {Manifold} {Approximation} and {Projection} for {Dimension} {Reduction}},
shorttitle = {{UMAP}},
doi = {10.48550/arXiv.1802.03426},
publisher = {arXiv},
author = {McInnes, Leland and Healy, John and Melville, James},
month = sep,
year = {2020},
}
@inproceedings{Eiter1994ComputingDF,
title={Computing Discrete Fr{\'e}chet Distance},
author={Thomas Eiter and Heikki Mannila},
year={1994},
url={https://api.semanticscholar.org/CorpusID:16010565}
}
