Ensemble ML beats neural nets for cryopathy syndrome classification

Cryopathy syndromes are notoriously difficult to diagnose because laboratory patterns overlap across diagnostic categories and some conditions are rare, forcing heavy reliance on expert judgment. In a new preprint on arXiv, researchers led by Nataliya Shakhovska analyzed 2,686 patient records spanning 14 diagnostic labels, including cryoglobulin measurements, precipitation tests, hemagglutinin and hemolysin titers, and demographic data. They preprocessed the data with cleaning, encoding, imputation, normalization, and clinically informed interaction features, then evaluated 12 modeling strategies: Random Forest, Gradient Boosted Trees, Multi-Layer Perceptron, soft-voting ensembles, SMOTE balancing, hierarchical classifiers, period-aware models, targeted binary classifiers, and probability calibration.

Performance was assessed via stratified train-test splits and 5-fold cross-validation. The overall task was difficult due to marked class imbalance and clinical overlap, but the best multiclass results came from a soft-voting ensemble of Random Forest and Gradient Boosted Trees. Tree-based methods consistently outperformed the neural network, and feature engineering—especially cryoglobulin-based interaction features—significantly improved discrimination. Cross-validation confirmed stable performance for the balanced Random Forest model. The findings suggest that ensemble, tree-based approaches with crafted features are a practical, interpretable strategy for clinical decision support in cryopathy classification.