Application of functional near-infrared spectroscopy and machine learning to predict treatment response after six months in major depressive disorder.

Abstract

Depression treatment responses vary widely among individuals. Identifying objective biomarkers with predictive accuracy for therapeutic outcomes can enhance treatment efficiency and avoid ineffective therapies. This study investigates whether functional near-infrared spectroscopy (fNIRS) and clinical assessment information can predict treatment response in major depressive disorder (MDD) through machine-learning techniques. Seventy patients with MDD were included in this 6-month longitudinal study, with the primary treatment outcome measured by changes in the Hamilton Depression Rating Scale (HAM-D) scores. fNIRS and clinical information were strictly evaluated using nested cross-validation to predict responders and non-responders based on machine-learning models, including support vector machine, random forest, XGBoost, discriminant analysis, Naïve Bayes, and transformers. The task change of total haemoglobin (HbT), defined as the difference between pre-task and post-task average HbT concentrations, in the dorsolateral prefrontal cortex (dlPFC) is significantly correlated with treatment response (p < 0.005). Leveraging a Naïve Bayes model, inner cross-validation performance (bAcc = 70% [SD = 4], AUC = 0.77 [SD = 0.04]) and outer cross-validation results (bAcc = 73% [SD = 3], AUC = 0.77 [SD = 0.02]) were yielded for predicting response using solely fNIRS data. The bimodal model combining fNIRS and clinical data showed inferior performance in outer cross-validation (bAcc = 68%, AUC = 0.70) compared to the fNIRS-only model. Collectively, fNIRS holds potential as a scalable neuroimaging modality for predicting treatment response in MDD.