Explainable AI Identifies Cortical Asymmetry to Detect Major Depression

A deep learning model using fNIRS data achieves 81% accuracy by mapping hemodynamic imbalances in the dorsolateral prefrontal cortex.

For Doctors in a Hurry

Clinicians lack explainable diagnostic tools for identifying major depressive disorder using functional near-infrared spectroscopy brain imaging data.
The researchers analyzed data from 48 patients and 68 healthy controls during a verbal fluency task using a convolutional neural network.
The model achieved 81.17 percent accuracy, 79.5 percent sensitivity, and 82.38 percent specificity in identifying major depressive disorder.
The authors concluded that their model effectively visualizes decision-making processes by highlighting critical activity in the dorsolateral prefrontal cortices.
This diagnostic tool may assist clinicians by providing objective, interpretable evidence of interhemispheric asymmetry in patients with major depressive disorder.

The Search for Objective Diagnostic Tools in Psychiatry

Major depressive disorder remains a primary driver of global disability, yet its diagnosis continues to rely heavily on subjective clinical interviews and patient self-reports [1, 2]. While functional near-infrared spectroscopy (fNIRS), a portable, non-invasive tool for measuring cortical hemodynamic responses by tracking blood flow changes in the brain, has emerged as a potential diagnostic aid, its transition into routine clinical practice has been slowed by methodological inconsistencies [3, 4]. In some regions, such as Japan, optical topography is already utilized as an auxiliary laboratory test to help differentiate between various depressive states [5]. However, the complexity of interpreting these hemodynamic waveforms often requires specialized expertise, limiting the utility of the raw data for the average practitioner [1]. A new study investigates how deep learning can bridge this gap by providing automated, interpretable diagnostic insights that translate complex cortical signals into actionable clinical data.

Deep Learning and the Challenge of Interpretability

To address the limitations of traditional diagnostic methods, the researchers implemented an explainable artificial intelligence (XAI) model based on a convolutional neural network (CNN). A convolutional neural network is a sophisticated deep learning architecture specifically designed to process grid-like data, such as images or time-series signals, by identifying spatial and temporal patterns. While standard deep learning models often function as opaque systems that provide an output without a clear rationale, the use of an explainable artificial intelligence (XAI) model ensures that the logic behind a diagnosis is accessible to the clinician. This transparency is critical for computer-aided diagnosis (CAD) in a psychiatric setting, as it allows physicians to verify that the model is focusing on relevant physiological markers rather than statistical noise. By making the "black box" of AI transparent, clinicians can better integrate these findings with their own bedside observations.

Measuring Hemodynamic Responses During Cognitive Stress

To evaluate the diagnostic utility of the explainable artificial intelligence model, the researchers recruited a study population consisting of 48 patients with major depressive disorder and 68 healthy controls. This cohort provided the necessary data to train and validate the convolutional neural network, ensuring the model could distinguish between pathological and typical cortical activity. The researchers utilized functional near-infrared spectroscopy (fNIRS) to monitor brain activity while participants performed a verbal fluency task. This specific cognitive challenge serves as a provocative test, requiring the patient to generate as many words as possible from a given category within a set timeframe. By measuring changes in oxygenated hemoglobin during this period of cognitive stress, the fNIRS system captured the dynamic hemodynamic responses of the prefrontal cortex in real time, providing a window into the brain's functional capacity under load.

Diagnostic Accuracy and Validation

The researchers evaluated the diagnostic utility of the convolutional neural network by measuring its ability to correctly identify patients with major depressive disorder within the study population. The model achieved an average accuracy of 81.17 percent, indicating a high level of precision in distinguishing between clinical depression and healthy controls based on hemodynamic data. In terms of clinical reliability, the system demonstrated a sensitivity of 79.5 percent, which reflects its capacity to correctly identify individuals who actually have the disorder. Furthermore, the model showed a specificity of 82.38 percent, a metric that quantifies its ability to correctly identify those without the condition, thereby minimizing the risk of false positive diagnoses in a clinical setting. To ensure the reliability of these performance metrics and to confirm that the findings were not the result of statistical noise, the researchers employed ten-fold cross-validation, a statistical technique that involves partitioning the total dataset into ten distinct subsets where the model is iteratively trained on nine and tested on the tenth. This process ensures that the accuracy remains consistent across different patient samples and is not skewed by outliers.

Mapping the Dorsolateral Prefrontal Cortex

To understand the physiological basis of the model's diagnostic decisions, the researchers utilized layer-wise relevance propagation (LRP), a technique that identifies the specific contribution of individual input data points to the final prediction by redistributing the prediction score backward through the neural network layers. This analysis revealed that channels within both the right and left dorsolateral prefrontal cortices (DLPFC) were critical for classification. By localizing the decision-making process to these specific regions, the study provides a biological rationale for the model's accuracy, as the dorsolateral prefrontal cortex is a primary hub for executive function and emotional regulation, both of which are frequently impaired in patients with major depressive disorder. The researchers further validated the model's internal logic by comparing its findings to traditional statistical methods, observing that the relevance score distribution in the right dorsolateral prefrontal cortex closely aligned with the Fisher score distribution, a standard statistical metric used to rank the importance of specific features in distinguishing between two groups. A detailed analysis of symmetrical channel pairs with high relevance scores revealed distinct interhemispheric asymmetry in the waveforms of patients with major depressive disorder. While healthy controls typically demonstrate balanced hemodynamic activity across both hemispheres during cognitive stress, the patients with depression exhibited a significant imbalance in oxygenated hemoglobin levels between the left and right sides of the brain. For the practicing physician, these findings offer a visual map of cortical dysfunction, providing an objective physiological marker that correlates with the clinical presentation of depression and supports the use of functional near-infrared spectroscopy as a diagnostic aid.

Study Info

Functional near-infrared spectroscopy-based computer-aided diagnosis of major depressive disorder using explainable artificial intelligence: Comparison with conventional machine learning

Kyeonggu Lee, M. Chun, Jinuk Kwon, JongKwan Choi, et al.

Journal Journal of Affective Disorders

Published December 19, 2025