Machine Learning Classification of Cryopathy Syndromes: A Comprehensive Comparative Study

Technical Summary: Machine Learning Classification of Cryopathy Syndromes

Problem Statement
The classification of cryopathy syndromes presents a significant diagnostic challenge due to the overlap of laboratory patterns across different diagnostic categories and the rarity of certain conditions. Routine interpretation of cryoglobulin-related tests often relies on expert judgment, as standard laboratory panels provide a limited view of the immunological state, with single profiles compatible with multiple diagnoses. This study addresses the specific computational challenge of automating the multiclass classification of cryopathy syndromes from routine laboratory data. The task is characterized by severe class imbalance (an imbalance ratio exceeding 80:1), a limited feature space, and genuine clinical overlap between 14 distinct diagnostic categories, including autoimmune, immunodeficiency-related, allergic, mediator-type, viral-associated, and control groups.

Methodology
The study utilized a dataset of 2,686 patient records from a clinical immunology laboratory in Lviv, Ukraine, spanning pre-COVID, COVID, and wartime periods. The data included demographic variables, spectrophotometric cryoglobulin levels, precipitation tests, and hemagglutinin/hemolysin titers.

• Preprocessing and Feature Engineering: Raw data underwent cleaning, encoding, and imputation. Crucially, seven clinically informed engineered features were constructed, including interactions between cryoglobulin levels and hemagglutinin titers, ratios of auto- to isohemagglutinin, and logarithmic transformations of cryoglobulin levels.
• Modeling Framework: All algorithms were implemented from scratch in NumPy to ensure transparency and reproducibility, avoiding library-specific defaults. The study evaluated 12 modeling strategies:
  • Base Classifiers: Decision Trees, Random Forest (RF), Gradient Boosted Trees (GBT), and Multi-Layer Perceptron (MLP).
  • Ensemble Methods: Soft-voting ensembles combining RF, GBT, and MLP.
  • Advanced Strategies: Hierarchical classification (pathogenetic type to specific diagnosis), period-aware models (separate models for different epidemiological periods), and targeted binary classifiers for specific diagnostic pairs.
  • Class Balancing: Synthetic Minority Over-sampling Technique (SMOTE) and aggressive augmentation with Gaussian noise were applied to address class imbalance.
  • Calibration: Platt scaling was applied to improve the reliability of probability estimates.
• Evaluation: Performance was assessed using stratified train-test splits and stratified 5-fold cross-validation. Metrics included macro-averaged F1 score, accuracy, Cohen's kappa, Top-3 accuracy, and Expected Calibration Error (ECE).

Key Results

• Model Performance: The best overall multiclass performance was achieved by a soft-voting ensemble of Random Forest and Gradient Boosted Trees (Macro F1: 0.1704). Tree-based methods consistently outperformed neural networks; the MLP models yielded the lowest performance, suggesting that for this small, heterogeneous tabular dataset, deep learning is not superior to ensemble baselines.
• Class Imbalance Trade-off: A distinct trade-off was observed: models with stronger augmentation improved overall accuracy but reduced macro F1 scores (sensitivity to minority classes), whereas SMOTE-based balancing improved macro F1 at the cost of overall accuracy.
• Feature Importance: Permutation analysis revealed that engineered features, particularly the logarithmically transformed cryoglobulin level and interaction terms involving hemagglutinin titers, were more informative than raw laboratory variables.
• Binary vs. Multiclass: Targeted binary classifiers for specific diagnostic pairs significantly outperformed the full multiclass models. The strongest results were observed for distinguishing Hepatitis C from Chronic Hepatitis (F1: 0.8000) and Cryovasculitis from Immunodeficiency-Related Disease (F1: 0.7126).
• Temporal Stability: Period-aware models did not outperform a global model trained on the combined dataset, indicating that the relationship between laboratory features and diagnosis remained stable across pre-COVID, COVID, and wartime periods.
• Calibration and Shortlisting: Probability calibration substantially improved reliability for the largest class (Immunodeficiency-Related Disease). While Top-1 accuracy was limited (0.2664), Top-3 accuracy reached 0.5310, indicating the correct diagnosis appeared in the top three predictions for over half the cases.

Significance and Claims
The authors position this work as an exploratory study that fills a gap in the literature, as no prior studies have addressed multiclass machine learning classification of cryopathy syndromes from routine cryoglobulin data. The study does not claim to have developed a clinically ready diagnostic tool for full automation. Instead, the significance lies in establishing a quantitative baseline and proposing a realistic, staged decision-support strategy.

The paper argues that the most practical near-term application is a hybrid approach: using a calibrated multiclass ensemble to generate a shortlist of plausible candidates (Top-3), followed by targeted binary classifiers for difficult differential diagnosis pairs. This approach aligns with clinical reasoning by narrowing the search space rather than replacing expert judgment. Additionally, the study highlights the value of domain-informed feature engineering in revealing diagnostically relevant structures that raw data alone cannot capture. The findings suggest that while full multiclass classification remains difficult due to clinical overlap, machine learning can capture diagnostically relevant structure in cryoglobulin data when integrated with domain knowledge and appropriate evaluation metrics.