Can Statistical Process Control be Applied to Identify and Classify Depressive Disorders from Peripheral Biosignatures?
Coradduzza, Donatella
Sanna, Andrea
De Miglio, Maria Rosaria
Carru, Ciriaco
Medici, Serenella
Grosso, Massimiliano
Pdf

How to Cite

Coradduzza D., Sanna A., De Miglio M.R., Carru C., Medici S., Grosso M., 2026, Can Statistical Process Control be Applied to Identify and Classify Depressive Disorders from Peripheral Biosignatures?, Chemical Engineering Transactions, 128, 139-144.
Pdf

Abstract

Clinical studies commonly rely on supervised classification methods to distinguish healthy and diseased subjects. However, these approaches require predefined class labels and may not fully account for the intrinsic variability of biological systems. This work investigates an alternative unsupervised framework inspired by Multivariate Statistical Process Control (MSPC). Following the process-monitoring analogy, healthy individuals are considered to represent an "in-control" state, whereas pathological subjects correspond to an "out-of-control" state. A Principal Component Analysis (PCA) model was built using 42 blood-derived variables collected from healthy controls, including haematological parameters, inflammatory indices and trace elements. The resulting model was used as a reference to evaluate new observations with Hotelling's T² and Squared Prediction Error (SPE) statistics. Independent healthy controls, patients with Major Depressive Disorder (MDD), and patients with Bipolar Disorder (BD) were projected onto the PCA model without recalibration. The results show that healthy subjects are largely classified as in-control observations, whereas both BD and MDD patients exhibit significant deviations from the healthy reference state. SPE was particularly effective in detecting disease-related deviations, yielding AUC values of 0.991 for HC vs MDD and 0.954 for HC vs BD, whereas discrimination between BD and MDD was more limited. These findings suggest that MSPC can provide an effective unsupervised framework for detecting abnormal peripheral biosignatures, while differential diagnosis may require additional modelling or complementary clinical information.
Pdf