Advancing the diagnosis of major depressive disorder: Integrating neuroimaging and machine learning

Abstract

Major depressive disorder (MDD), a psychiatric disorder characterized by functional brain deficits, poses considerable diagnostic and treatment challenges, especially in adolescents owing to varying clinical presentations. Biomarkers hold substantial clinical potential in the field of mental health, enabling objective assessments of physiological and pathological states, facilitating early diagnosis, and enhancing clinical decision-making and patient outcomes. Recent breakthroughs combine neuroimaging with machine learning (ML) to distinguish brain activity patterns between MDD patients and healthy controls, paving the way for diagnostic support and personalized treatment. However, the accuracy of the results depends on the selection of neuroimaging features and algorithms. Ensuring privacy protection, ML model accuracy, and fostering trust are essential steps prior to clinical implementation. Future research should prioritize the establishment of comprehensive legal frameworks and regulatory mechanisms for using ML in MDD diagnosis while safeguarding patient privacy and rights. By doing so, we can advance accuracy and personalized care for MDD.

Key Words: Major depressive disorder; Biomarkers; Neuroimaging; Machine learning; Personalized treatment; Resting-state functional magnetic resonance imaging; Functional connectivity; Model accuracy; Major depressive disorder diagnosis

Core Tip: Major depressive disorder (MDD), especially in adolescents, poses considerable diagnostic and therapeutic challenges owing to its heterogeneity and the subjective nature of traditional assessment methods. Recent advances in neuroimaging, combined with machine learning (ML) technologies, have led to the development of promising biomarkers and diagnostic tools for MDD. However, these challenges can be addressed through improved data privacy protection measures, advanced encryption and anonymization techniques, greater model transparency, stricter data quality control, and the establishment of clear ethical and legal frameworks. Such efforts are crucial to ensuring the safe, reliable, and compliant application of ML technologies in MDD diagnosis.

INTRODUCTION

Major depressive disorder (MDD) has become one of the leading causes of mental illness and physical disability worldwide[1]. Data suggest that over the past three decades, the prevalence of depression has considerably increased, with a trend toward affecting younger populations[2-4]. Adolescence is a pivotal period for individual development, marked by considerable physiological, emotional, and sociopsychological changes that considerably increase the susceptibility of adolescents to mental disorders[5]. Adolescent depression can lead to severe long-term consequences, including more severe depression in adulthood, lower educational attainment, unemployment, reduced social support, early marriage, and early childbearing—all of which are concerning psychosocial issues[6,7]. Moreover, depression is closely linked to substance abuse, nonsuicidal self-injury, academic difficulties, and physical health problems. The most severe consequence of untreated depression is the increased risk of suicide[8]. Therefore, early diagnosis and intervention are crucial.

Diagnosing MDD presents multifaceted challenges stemming from both the inherent heterogeneity of the disorder and the subjectivity embedded in current diagnostic approaches. MDD is a highly heterogeneous condition, marked by considerable interindividual variation in its etiology, pathophysiological mechanisms, symptomatology, and disease severity. Clinically, the assessment of mental health disorders predominantly depends on psychiatrists’ observations and semistructured interviews focused on symptom intensity, along with standardized self-assessment scales[9]. This diagnostic approach relies heavily on clinicians’ interpretation of individual symptom profiles and patient narratives, making it vulnerable to subjective biases. The convergence of heterogeneity and subjectivity not only complicates diagnosis but also increases the risk of misdiagnosis—a considerable concern because approximately one-quarter of individuals with bipolar disorder are mistakenly diagnosed with MDD[10,11]. Therefore, there is an urgent need to thoroughly examine individual differences and clarify the specific biological mechanisms underlying depressive disorders to develop more accurate diagnostic algorithms and personalized treatment strategies. The integration of objective biomarkers and the refinement of diagnostic protocols offer considerable potential to improve diagnostic accuracy and therapeutic effectiveness, ultimately enhancing patients’ overall prognostic outcomes.

This editorial comments on a study recently published in World Journal of Psychiatry titled “Resting-state functional magnetic resonance imaging and support vector machines for the diagnosis of major depressive disorder in adolescents”[12]. The study used functional magnetic resonance imaging (fMRI) and support vector machine (SVM) analysis to identify abnormal functional connectivity (FC) values in the right lingual gyrus as potential neuroimaging biomarkers for distinguishing MDD patients. The editorial summarizes the potential of neuroimaging and machine learning (ML) in the diagnosis of MDD and discusses the ethical, technical, and practical challenges, along with future application prospects.

TECHNOLOGY AND POTENTIAL OF NEUROIMAGING AND ML IN MDD DIAGNOSIS

Gray and white matter are essential for regulating various complex behaviors and cognitive functions, including emotions, perception, attention, and social interactions[13,14]. Disruptions in these brain structures may underlie psychiatric disorders, with deficits in areas like the prefronto-limbic system and corpus callosum often seen in MDD[15-17]. Identifying measurable brain activity markers for MDD is critical for optimizing diagnosis. Advances in neuroimaging technology, particularly fMRI, have enabled the detection of brain activity and structural changes in MDD patients.

FMRI, based on blood-oxygen-level-dependent signals, is widely used to monitor brain activity by detecting changes in blood flow[18]. It includes resting-state functional MRI (rs-fMRI) and task functional MRI (task-fMRI). Studies have shown that patients with MDD predominantly exhibit lower task-related activation in task-fMRI[19]. However, there are also cases of enhanced activation in certain brain regions, such as the prefrontal and anterior temporal areas, indicating the difficulties experienced by patients with MDD in executive functioning and emotional regulation[20]. Unlike task-fMRI, which requires a task design, rs-fMRI offers the convenience of not requiring specific tasks and is highly sensitive to intrinsic brain network activity, making it particularly suitable for patients who cannot comply with tasks or for special populations. The current analysis methods related to rs-fMRI data mainly include regional homogeneity (ReHo), amplitude of low-frequency fluctuations (ALFF), fractional ALFF (fALFF), and FC analysis. ReHo analysis revealed significant increases in the left dorsal medial prefrontal gyrus and the left anterior lobe of the cerebellum, whereas significant decreases were observed in the frontal, temporal, parietal, and occipital cortices[21-23]. Similarly, ALFF and fALFF analyses revealed abnormalities in the intensity of spontaneous neural activity in the brains of patients with MDD, such as decreases in ALFF in the bilateral cerebellum, bilateral precuneus, and left occipital cortex, whereas increases in ALFF occurred in the right superior frontal gyrus, bilateral insula extending into the striatum, and left supramarginal gyrus[24,25]. Additionally, fALFF analysis showed that fALFF in the left frontal region of patients with MDD was significantly lower than that in healthy controls[26].

In contrast to ReHo, ALFF, and fALFF, which reflect spontaneous neural brain activity, FC is quantified by calculating the Pearson correlation coefficients between different brain regions, providing a direct measure of their functional interactions[27]. In patients with MDD, FC is often abnormally enhanced between the default mode network (DMN), central executive network, and salience network, which is closely associated with the pathological characteristics of the disease[28]. Existing research shows a trend toward diversification and combination in neuroimaging feature selection, including the use of FC features and dynamic functional network connectivity, along with low-frequency oscillation features (e.g., ALFF, fALFF), ReHo, and network properties (e.g., degree centrality) to describe the complex variations in brain activity and network structure[29-31]. The integration of multiple features, combined with ML techniques, enables the processing of vast amounts of information and facilitates the classification or prediction of data based on specific characteristics. This approach significantly enhances the diagnostic potential for MDD.

Integrating ML with clinical decision-making helps clinicians identify MDD patients and optimize treatment by providing personalized recommendations. ML methods like SVM and deep learning (DL) are trained to recognize disease-specific imaging features, distinguishing MDD patients from healthy controls or different disease subtypes[32-34]. These methods can also predict treatment outcomes, improving early risk identification and therapeutic efficacy[35].

Moreover, traditional ML is still the dominant approach in existing studies. Some existing studies have also explored other methods, such as logistic regression (LR), random forest, Extreme Gradient Boosting (XGBoost), AdaBoost, sparse LR, maximum margin clustering (MMC)-based classifiers, and linear discriminant analysis[36-40]. DL algorithms have also been introduced to enhance classification performance[40]. These studies also employed different performance evaluation metrics, such as accuracy, area under the curve, and consistency. Cross-validation techniques, such as leave-one-out cross-validation and k-fold cross-validation, were employed to ensure the generalizability and robustness of the models[36-40].

Although DL applications in MDD are less common than traditional ML methods, they have demonstrated promising results. For instance, an unsupervised ML method based on MMC effectively segmented voxels in the perigenual cingulate cortex into two subregions, leveraging distinct functional connection patterns in the resting state to successfully distinguish patients with MDD from healthy controls[40]. Convolutional neural networks have been employed for whole-brain FC analysis, revealing the key role of visual and sensorimotor networks in MDD through SHapley additive exPlanations (SHAP) feature exploration[41]. Additionally, spectral graph convolutional networks, which integrate effective connectivity with nonimaging phenotypic data, have proven effective in analyzing complex neuroimaging data[42].

As a potential auxiliary diagnostic tool, the effectiveness of ML classification methods in the diagnosis of MDD is influenced by various factors, such as the selection of neuroimaging features and classifiers[31,36,43-45]. Studies have demonstrated that using data lacking discriminative features can result in poor classification performance. For instance, a study involving 360 patients with MDD and 360 healthy controls demonstrated that using FC as a feature for SVM classification achieved accuracies close to chance levels (52.5%–52.8%) for both binary and ternary classifications[36]. Another study demonstrated that without feature selection, the SVM classifier’s accuracy remained at random levels in a cohort of 360 patients with MDD and 360 healthy controls[46]. Alternatively, a study using five distinct features and four feature selection methods achieved a higher accuracy of 90.2%. This highlights the considerable influence of the methodology and feature selection on the classification performance[29]. In contrast, incorporating highly discriminative features can considerably enhance the classification performance. For example, a combination of MiR-9 levels, Childhood Trauma Questionnaire scores, and FC in the amygdala–prefrontal–limbic regions yielded an optimal accuracy of 85.1%. This accuracy is considerably higher than that achieved using nondiscriminative features[44]. Another study also reported strong discriminative features in the prefrontal and DMN regions, showing considerable differences between patients with MDD and healthy controls, with classification accuracies reaching 90%[47]. Furthermore, the selection of classification algorithms also impacts accuracy. Although SVM has been widely used, other classifiers can yield comparable results. A study comparing six classifiers [SVM, LR, decision tree (DT), ridge regression, least absolute shrinkage and selection operator, and XGBoost] using FC data from 1021 patients with MDD and 1100 healthy controls demonstrated that XGBoost achieved the highest accuracy of 72.8%[45]. In another study, researchers improved the model’s ability to differentiate between MDD and healthy controls using various feature extraction algoristhms and optimizing SVM parameters, achieving an accuracy of 66%[31].

Although the early results of ML methods in MDD diagnosis are encouraging, their full potential lies in optimizing feature selection and classification algorithms. This will lead to earlier and more accurate diagnoses. Future ML models that integrate multimodal neuroimaging features with individualized data could pave the way for personalized treatment plans and considerably improve early intervention capabilities for MDD in clinical settings[48].

RISKS AND CHALLENGES

The use of ML in healthcare presents significant challenges, particularly concerning informed consent and privacy protection. ML algorithms require large patient datasets for training, raising concerns about how patient data will be used. Obtaining informed consent is complicated because patients may not fully understand terms like "ML models" and "data algorithms", which can lead to the "fallacy of consent"[49]. The current legal framework fails to clearly define data ownership, and the resulting ambiguity complicates the challenge of balancing research accessibility with patient privacy protection[50,51]. Although continuous accumulation of medical data is crucial for improving ML algorithms, the repeated use of data without proper consent may not align with patient expectations[52]. Patients often lose control over their personal information once it is integrated into databases, and even if data use appears harmless or patients are unaware of it, the erosion of ownership and control, along with the lack of transparency regarding data usage, can still infringe upon individual privacy. Additionally, some medical data are uploaded and stored online, considerably increasing the risk of data breaches. Patient information uploaded on the internet may be vulnerable to hacker attacks or database vulnerabilities. Notably, long-term sensitive medical data, such as genetic and specific disease information, can lead to persistent privacy violations and potential discrimination if leaked[53]. The increasing aggregation and analysis of data increases the risk of reidentification in large online databases[54]. Despite traditional deidentification methods, health data remain susceptible to reidentification, raising concerns about privacy breaches[55]. Balancing data use while protecting patient privacy is a critical challenge in medical ML. However, data scarcity in adolescents with MDD can lead to overfitting and reduced diversity. This can compromise the model’s ability to generalize and perform well in real-world settings[56]. Algorithmic bias is also a significant ethical issue in healthcare, as imbalanced or unrepresentative training data can lead to unfair treatment of certain groups, such as women and people of different races and ethnicities[57]. For example, using data from predominantly white populations to predict cardiovascular risk in nonwhite populations may result in overestimating or underestimating the risk, which can affect diagnostic and treatment decisions[58].

Another important consideration is the inconsistency of research results, which is influenced by data quality and volume, as well as model performance. Despite the expansion of datasets, high-quality data cannot be guaranteed because of the variations in imaging preprocessing, acquisition techniques, and patient demographics across different institutions. Factors like the number of head coil channels, inconsistent subject positioning, varying image contrast, differences in magnetic resonance scanner manufacturers and field strengths, and physiological factors like heart rate, cardiopulmonary function, age, and gender, contribute to the variability in MRI quality[59-62]. This heterogeneity complicates data integration and analysis[63]. ML models require sufficient sample sizes for effective training. However, data scarcity in adolescents with MDD can lead to overfitting and reduced diversity. This can compromise the model’s ability to generalize and perform well in real-world settings[56]. Additionally, the effectiveness of these models heavily depends on the selection of neuroimaging features and the choice of algorithms. Features used for training must be representative and discriminative. Otherwise, classification accuracy may be close to random, considerably diminishing the model’s practical value and its ability to provide meaningful diagnostic assistance in clinical settings[36]. The choice of algorithm plays a crucial role in evaluating model performance. An unsuitable algorithm can compromise the stability and accuracy of the results, leading to unreliable outcomes or limited applicability to new datasets[31]. Moreover, studies using the same algorithm, such as the SVM classifier, have reported considerably varied results, highlighting the instability of models across different datasets and feature selection criteria[31,44,47].

The limitations in the integration of ML into clinical practice are also important, especially in terms of model explainability, decision-making authority, and practical application challenges. The most critical issue is model explainability, particularly the enigmatic “black-box” characteristic of certain ML models. The primary concern lies in the inherent complexity of these models. DL architectures, which have extensive parameter counts and intricate multilayered nonlinear configurations, render decision-making processes opaque and challenging to elucidate. The unpredictability of the training process poses another major challenge because these models are trained on heterogeneous and complex datasets, with the optimization procedure navigating a high-dimensional solution space to find the optimal outcome. This complexity makes it difficult to predict and interpret the patterns and decisions that the model learns, and the lack of transparency may reduce the confidence of healthcare practitioners, potentially affecting clinical decision-making[64-66]. A further concern is the potential weakening of the decision-making authority of clinicians. Overreliance on ML-driven diagnostic and therapeutic recommendations may weaken the trust-based doctor-patient relationship, reduce clinicians’ decision-making authority, and lead to negative consequences, such as diminished trust, dehumanized decision-making, and decreased job satisfaction for doctors[66]. Furthermore, this reliance blurs the lines of accountability. When errors or misjudgments occur, it becomes unclear whether the doctor, model developer, or medical institution should be held responsible. This ambiguity poses considerable challenges to accountability and hinders the widespread adoption of ML models in medical settings[49]. Finally, the clinical feasibility of ML for the diagnosis of MDD faces resource and technology limitations, such as hardware, data support, and physician training, especially in low-resource areas[67]. These challenges may hinder the widespread adoption and effective use of ML technologies. Therefore, finding a balance between ML’s supportive role and clinicians’ decision-making authority, ensuring clear accountability, and addressing resource constraints are crucial steps for advancing the broader application of ML in MDD diagnosis.

PROSPECTS

To overcome the challenges hindering the seamless application of medical ML, future efforts should prioritize developing a comprehensive framework that guarantees data compliance, enhances model accuracy, and fosters trust (Figure 1).

Figure 1 Core enhancing factors, risks, challenges, and future directions in medical machine learning. ALEF: Amplitude of low-frequency fluctuations; DFNC: Dynamic functional network connectivity; DT: Decision tree; FC: Functional connectivity; LASSO: Least absolute shrinkage and selection operator; MDD: Major depressive disorder; ReHo: Regional homogeneity; RF: Random forest; Rs-fMRI: Resting-state functional magnetic resonance imaging; SVM: Support vector machine; XGBoost: Extreme Gradient Boosting.

First, to protect data privacy, stricter regulations and secure storage solutions must be implemented to ensure the safe custody and transmission of medical data, minimizing the risk of breaches and misuse[68]. Moreover, continuous research and development of advanced encryption technologies will provide additional technical support to safeguard data privacy[69]. With the increasing adoption of ML, it is essential to develop simplified and transparent informed consent mechanisms. Using plain language and interactive models can help patients understand how their data will be used, thus narrowing the knowledge gap between healthcare providers and patients[70]. Exploring dynamic and continuously updated consent models is essential to keep patients informed and actively participate in decisions about the use of their data during the process[71]. It is also essential to recognize the complexity of “ownership” in clinical data because it is not generated or controlled by patients but is the result of contributions from multiple stakeholders[51]. Therefore, the solution requires the collaborative involvement of governments, healthcare institutions, technology platforms, healthcare providers, and patients, to build a flexible and comprehensive data security governance framework that ensures balanced protection of all parties’ interests.

The issues related to heterogeneity in data sources and sample size can be addressed using the following methods. First, standardizing data acquisition processes and unifying scanning protocols can help reduce confounding factors; using standardized software packages, preprocessing settings, task contrasts, seed masks, etc., can help overcome heterogeneity in fMRI data during preprocessing, feature extraction, and statistical testing. Second, applying statistical methods, such as stratified analysis to adjust for physiological variables can control for potential biological interference and improve data accuracy and consistency. Additionally, conducting large-scale multi-center studies with diverse participant groups can expand sample sizes and help identify homogeneous subgroups, thereby improving model generalizability and data consistency. Currently, the ENIGMA MDD consortium, through global collaboration and data sharing, has facilitated the integration of large-scale datasets in the field of MDD, effectively improving the reliability and reproducibility of research and laying the foundation for understanding the neurobiological mechanisms of MDD and its clinical applications[72].

To improve model accuracy, integrating data from diverse populations and using techniques, such as data augmentation, are crucial strategies[73]. Additionally, selecting representative biomarkers and using suitable algorithms while considering clinical factors, such as disease severity, medications, and comorbidities, will improve the model’s accuracy and reliability[31,44,45,47,74]. The key to improving model performance lies in feature selection, model training, and selecting suitable classification algorithms. By using cross-validation and suitable algorithms, such as SVM or neural networks, the model’s generalization ability can be optimized, while incorporating clinical factors enhances accuracy[56,74-76]. Finally, a comprehensive model evaluation and optimization can be achieved using different evaluation metrics[45,47,56]. Moreover, standardized data collection and preprocessing protocols are essential to ensure data consistency and comparability. This enhances the scalability and stability of ML models across different institutional settings[77].

Building trust requires enhancing algorithmic explainability. For simpler models like regression models and DTs, their natural transparency and simplicity make them inherently interpretable, directly reflecting the decision-making logic. However, for more complex models, such as DL, their inherent complexity often requires post-hoc explainability techniques, such as local interpretable model-agnostic explanations and SHAP, to improve the oversight of decision-making processes[78,79]. This strengthens their oversight of the decision-making process[80]. Furthermore, a collaborative human–machine decision-making framework should be established to minimize the overreliance on ML models. This framework should clearly define the supportive role of models in clinical decision-making. To address accountability, healthcare institutions and regulatory agencies must establish clear guidelines outlining the responsibilities of model developers, healthcare providers, and institutions. This ensures the safe and transparent implementation of medical ML systems[49].

To address accessibility issues, data sharing platforms and synthetic data generation techniques can help overcome data scarcity. Second, online learning and various training methods can equip doctors and healthcare professionals with fundamental ML knowledge and its application in diagnosis while simplifying tools to make them easier to use. Additionally, policy support and financial investment can promote the widespread adoption of these technologies.

Future studies should focus on strengthening patient data privacy protection, optimizing model reliability, and fostering trust among healthcare providers. Establishing a standardized and regulated ecosystem for ML applications in healthcare is crucial to facilitate its widespread and safe adoption in clinical settings.

CONCLUSION

The convergence of neuroimaging and ML presents immense possibilities for the early detection of MDD, paving the way for more accurate diagnostic tools and individualized treatment plans. However, integrating ML models with neuroimaging also poses challenges related to data privacy, model reliability, and physician trust. To mitigate these challenges, systematic advancements from legal and policy standpoints are crucial. Robust legal frameworks and regulatory mechanisms must be established to ensure the progress of medical artificial intelligence technologies while safeguarding patient privacy and rights. These measures will create a solid foundation for the widespread adoption and sustained development of medical ML technologies.