Precision Diagnosis of Depression Levels via Distributed Classification and SHAP Analysis

This article presents a novel classification service designed for medical datasets with limited sample sizes, specifically focusing on depression assessment based on blood pressure oscillograms. The service employs multiple binary classification components, each trained to differentiate specific class separations. A key feature of the method is the use of both individual input features and correlated feature products, enhancing classification accuracy. The system achieved notable performance on a small dataset with accuracy, recall, and precision values of 0.9172, 0.9341, and 0.9811, respectively. While precision is high, recall is relatively lower, indicating a slight tendency to underestimate depression levels. Class 4, representing the highest depression level, demonstrated the most classification mismatches due to the small sample size. The SHAP analysis identified five key features—ULF_70-100, ULF_per_100-70, O_L2_pos, ULF_100-70, and ULF_per_total—as the most influential, most of which are associated with ultra-low-frequency (ULF) oscillations. The developed classification service proves its applicability in healthcare, offering potential for further adaptation to broader medical domains, enabling personalized medical diagnostics and remote healthcare solutions.