A sparse reverse nearest neighbors discriminant analytic approach for major depressive disorder identification

Major Depressive Disorder (MDD), a severe mental illness that significantly impairs human quality of life, currently affects approximately 300 million individuals worldwide. Its clinical heterogeneity and neurobiological complexity pose substantial challenges to research and early prediction. Conventional depression screening methods predominantly rely on patient self-reports or clinical scales, which are susceptible to emotional fluctuations, memory biases, and clinicians’ subjective judgments, frequently resulting in misdiagnosis or underdiagnosis. Consequently, there exists an urgent demand for objective automated detection approaches based on physiological biomarkers. In recent years, functional near-infrared spectroscopy (fNIRS) has emerged as a promising non-invasive cerebral optical imaging technology, demonstrating considerable potential in affective computing and depression identification due to its portability and capacity for long-term monitoring. However, existing fNIRS-based depression studies are constrained by limited sample sizes and the high heterogeneity in patient symptom manifestations and pathological mechanisms, which hinder the identification of unified biomarkers for fNIRS-based diagnosis. Furthermore, conventional machine learning-derived features exhibit insufficient discriminative power and sensitivity to noisy data. To address these limitations, this study proposes a Sparse Reverse Nearest Neighbor Local Fisher Discriminant Analysis (SRLFDA) algorithm for feature extraction from fNIRS data. The algorithm incorporates reverse nearest neighbors to reduce outlier noise interference in scatter matrix estimation, enhancing boundary-aware learning. Coupled with sparse projection matrices derived from extended Bregman iteration, it significantly boosts robustness and discrimination. In experimental validation, we categorized task-state datasets into positive, neutral, and negative affective states, subsequently evaluating SRLFDA performance across three distinct datasets. Ultimately, under negative auditory stimuli, our algorithm in conjunction with the Support Vector Machine classifier yielded encouraging results, with an accuracy of 91.7%. The results demonstrate that our algorithm achieves superior accuracy and robustness in depression screening applications.