Classify the fNIRS signals of first-episode drug-naive MDD patients with or without suicidal ideation using machine learning.

backgroundMajor Depressive Disorder (MDD) has a high suicide risk, and current diagnosis of suicidal ideation (SI) mainly relies on subjective tools.Neuroimaging techniques, including functional near-infrared spectroscopy (fNIRS), offer potential for identifying objective biomarkers. fNIRS, with its advantages of non-invasiveness, portability, and tolerance of mild movement, provides a feasible approach for clinical research. However, previous fNIRS studies on MDD and suicidal ideation have inconsistent results due to patient and methodological differences.Traditional machine learning in fNIRS data analysis has limitations, while deep - learning methods like one-dimensional convolutional neural network (CNN) are under-explored. This study aims to use fNIRS to explore prefrontal function in first-episode drug-naive MDD patients with suicidal ideation and evaluate fNIRS as a diagnostic tool via deep learning.

methodsA total of 91 first-episode drug-naive MDD patients were included and categorized into two groups based on their scores on the suicidal item of the 17-item Hamilton Depression Rating Scale (HAMD-17): 40 patients with suicidal ideation (SIs) and 51 patients without suicidal ideation (NSIs). Concurrently, 39 healthy controls (HCs) were recruited. We utilized fNIRS to measure the hemodynamic responses in the prefrontal cortex of each group during the verbal fluency task (VFT). A Kruskal-Wallis test was conducted to analyze the changes in oxyhemoglobin concentration among the three groups, and receiver operating characteristic (ROC) curves were generated for each region of interest.

resultsCompared to HCs, NSIs exhibited significantly reduced activation in the left dorsolateral prefrontal cortex (lDLPFC), frontopolar prefrontal cortex (FPC), orbitofrontal cortex (OFC), and ventrolateral prefrontal cortex (VLPFC), while SIs showed significantly decreased activation in the entire prefrontal cortex. The activation values of SIs in DLPFC, FPC, and OFC were significantly lower than those of NSIs. The highest accuracy for the three-class classification was observed in the lFPC, reaching 69.80%. The SIs group had the largest area under the ROC curve (AUC = 0.88) in the rFPC, while the NSIs group had the largest area under the ROC curve (AUC = 0.88) in the rDLPFC. The HCs group exhibited the largest area under the ROC curve (AUC = 0.92) in the rDLPFC and rVLPFC.

conclusionDLPFC, FPC, and OFC may serve as biomarker brain regions for identifying suicidal ideation in first-episode drug-naive MDD patients. The fNIRS-VFT task can be utilized clinically as an auxiliary diagnostic tool for mental disorders.