ATR-FTIR spectroscopic analysis and two-stage machine learning for the differentiation of bipolar and major depressive disorder.

In this study, we aimed to develop a two-stage hierarchical diagnostic model based on machine learning with attenuated total reflectance-Fourier transform infrared (ATR-FTIR) spectroscopy in order to distinguish patients with bipolar disorder (BD) and major depressive disorder (MDD) from healthy controls, and to distinguish between the two psychiatric disorders. A total of 102 age- and sex-matched individuals (BD = 34, MDD = 34, Control = 34) were included in the study. In the first stage, psychiatric cases and healthy controls were classified with an accuracy rate of 87.2% using the Random Forest algorithm. The most decisive spectral wavenumbers in this distinction were 3006 cm-1 and 1090 cm-1. In the second stage, the Radial Based Support Vector Machines (RBF-SVM) model was applied only on patient groups, and a Kappa value of 0.506 was obtained with a cross-validation accuracy of 75.1%. In this distinction, the 1148 cm-1 and 1541 cm-1 bands were determined as the strongest biomarkers. In regression analyses, a significant correlation was found between Beck Depression Inventory scores and FTIR spectra in the depression group, while no similar relationship was found with Young Mania Rating Scale scores in the bipolar group. The findings show that symptom severity in MDD creates state-dependent changes at the biochemical level, while the spectral profile provides trait-dependent information in BD. In conclusion, the integration of ATR-FTIR spectroscopy and machine learning stands out as an innovative and powerful analytical approach in the diagnosis of psychiatric disorders and the objective assessment of symptom severity.