Hotchkiss Brain Institute

Hyperspectral imaging on blood samples produces a noisy, high-dimensional time-series. Two failure modes were always within reach. Data leakage between train and evaluation could produce a model that looked spectacular and meant nothing. Outliers could quietly drive whichever architecture happened to be sensitive to them, again producing accuracy that would not survive a real clinical setting.

Anything we shipped had to be defensible against both.

We built the pipeline with Gaussian smoothing to reduce sensor noise, MinMax scaling for normalisation, and MiniRocket-style feature extraction to lift the signal out of the time-series. Splits were constructed to make leakage structurally hard, not just unlikely. Outliers were inspected, not silently absorbed.

On the modelling side, we benchmarked GRU, RNN, LSTM, and HIVE-COTE 2.0 against XGBoost, LightGBM, and Random Forest, on the same splits. Random Forest came out on top, both on raw accuracy and on stability across folds. We picked it on the evidence.

The Random Forest model exceeded 98% accuracy on hyperspectral blood-sample classification, with the leakage and outlier controls in place. Equally important, the pipeline is reproducible end-to-end, so future researchers can re-run, extend, and audit the result rather than taking it on trust.