Depression, a major mental health challenge, affects an estimated 300 million people worldwide with a lifetime risk of 5%-17% [1,2]. It is also anticipated to emerge as the leading cause of disability worldwide by 2030.2 The prevalence of depression in South Korea is 7.7%, with the country also having the highest suicide rate globally [3]. Thus, developing accurate predictive models for depression and suicide should be considered a public health priority to facilitate novel therapeutic approaches in South Korea.

Early detection and intervention are critical for enhancing clinical outcomes in depressive disorders. Despite this, a South Korean 2021 national-level survey revealed that only 28.2% of individuals with mental disorders had visited a medical center for treatment [4]. Among those who did not seek treatment, predominant impediments were stigma associated with mental disorders and fear of discrimination [5]. Currently, some limitations currently exist in accurately diagnosing and treating depression. First, most assessment scales rely on assessing depressive symptoms [6,7]. However, evaluating depression-related protective and vulnerability factors alongside surface depression symptoms can yield more comprehensive insights into depression’s chronicity and severity, thereby offering guidance for individualized treatment [8-11].

Despite the heterogeneity of depression, its diagnosis and treatment still hinge on depression-related symptom and sign assessment, while specific objective criteria—which could facilitate early diagnosis—have yet to be established [6,7]. Extant methods for depression and suicide prediction predominantly depend on subjective self-reported measures, potentially lacking in objectivity owing to individuals’ usual reluctance to fully disclose their thoughts [12]. The lack of objective biomarkers complicates accurate diagnosis and risks oversimplifying depression by overlooking individual symptoms and treatment response variations. Nonetheless, efforts have been made to incorporate objective indicators into depression diagnosis, with biomarkers reported to have the potential to identify treatment responsiveness and thereby improve understanding and management of depression. Meta-analytic findings indicate that cortisol is the only marker extensively studied for its ability to predict major depressive disorders’ (MDD) onset, relapse, and recurrence [13].

The pathogenesis of depression has been associated with hypothalamic-pituitary-adrenal (HPA) axis dysregulation, which leads to cortisol release as the end-product of a stress response [14,15]. Cortisol Awakening Response (CAR) includes a sharp increase in cortisol secretion after waking up, reaching its peak within 30-45 minutes and returning to baseline levels within 60 minutes after waking up [16]. Recently, CAR has been investigated as a biomarker of HPA axis function, reflecting responsiveness to external stressors upon awakening [17,18]. CAR is also useful to identify consistent stress and time conditions, rendering it useful as a diagnostic biomarker. Studies examining the CAR and depression have reported heterogeneous patterns [19,20], with depression associated with both increased and decreased CAR. These discrepancies may be attributed to differences in depression severity and differences in depressive states [21]. For example, in acute and mild-to-moderate depression groups, researchers observed an inverted U-shape with hyperactivity, whereas this inverted U-shape occurred for hypoactivity in the severe depression group [22]. A low CAR has also been associated with severe depression [23], being related to a more chronic course of the disorder [23-25].

To address these past limitations, we developed Minds.NAVI, a holistic assessment software program for depression screening that integrates psychological and biological biomarkers [26]. Particularly, it incorporates results from a psychological assessment battery, including assessments of current depressive symptoms, the PROtective and Vulnerable factors battEry (PROVE) test, and HPA axis function assessed via salivary cortisol and dehydroepiandrosterone (DHEA) levels.

Conventional depression clinical screening is insufficient to capture the holistic complexity of the disorder. Meanwhile, artificial intelligence (AI) techniques offer significant advantages through their ability to process large volumes of heterogeneous data. AI application may indeed allow a deeper understanding of depression, and assist mental health professionals in predictive decision-making [27]. AI-based predictions enable the early identification of high-risk medical conditions in patients, promoting early intervention adoption [28]. Particularly, machine learning techniques excel in feature selection and extraction across multiple data modalities (e.g., biomarkers and psychological data), which is particularly valuable when handling diverse data types [29,30]. They have demonstrated robust performance in psychiatric data analysis through advantages in pattern and relationship identification across multiple variables [29,30]. Such a systematic, multimodal approach to feature analysis might be critical for discovering meaningful patterns among various elements like salivary stress hormones, depressive symptoms, and protective-vulnerability factors. The outcome would be a more in-depth analysis of depression (e.g., discovering subtypes and allowing treatment personalization) compared with traditional assessment methods. Machine learning can also be advantageous for handling smaller datasets by leveraging cross-validation and regularization to prevent overfitting and improve model generalization [31].

In this research, we conducted a comparative analysis using four distinct models: two binary classification models (with and without salivary hormone data) to determine the presence or absence of depression according to Minds.NAVI results; two multiclass classification models (with and without physiological signals) aligned with the four-level classification of Minds.NAVI. The goal is delivering a comprehensive analysis of the impact of incorporating physiological signals alongside psychological factors in both binary and multiclass depression classification scenarios.

This study developed and evaluated machine learning models for classifying depression using both physiological and psychological indicator data from the Minds.NAVI assessment system. Our findings suggest the potential value of integrating salivary hormone data with psychological assessments for depression, thus addressing a key limitation in current depression assessment practices that rely primarily on subjective self-report measures. However, given the relatively small dataset (480 participants), these results should be interpreted cautiously and require validation in larger, more diverse populations before clinical implementation.

It is important to acknowledge a structural consideration regarding the relationship between the input features and the dependent variable. The primary objective of this study was to evaluate whether a machine learning approach could achieve comparable or superior predictive performance when applied to clinical variables used by the existing rule-based Minds.NAVI system. Importantly, the input variables were selected based on their clinical relevance and their independence from the outcome variable; specifically, only variables that could be considered independent of the final classification result were included as model inputs. Furthermore, the computational operations, weighting schemes, and decision rules embedded in the existing Minds.NAVI algorithm were not included in any form during the machine learning model’s training process. The RandomForest model learned non-linear, data-driven patterns across all features simultaneously, without being constrained by the predefined thresholds or decision hierarchy used in the rule-based system. Therefore, the machine learning model does not replicate or reproduce the internal algorithm; rather, it independently derives classification patterns from the raw feature values. We acknowledge that because the dependent variable (Minds.NAVI severity classification) is ultimately derived from the same domain of clinical indicators, a structural relationship between input and output inherently exists. However, this is methodologically distinct from direct circularity, in which the output computation is simply relearned. The fact that the multiclass model did not achieve perfect accuracy—particularly for intermediate severity categories (orange class sensitivity=0.40)—provides empirical evidence that the model does not merely reproduce the rule-based system but exhibits genuine variation in predictive capacity across severity levels. Nevertheless, we recognize that the current study design does not fully resolve the circularity concern, and future studies should evaluate model performance against independent clinical diagnoses (e.g., clinician-rated severity scales such as HAM-D or MADRS) as the primary outcome variable.

The multiclass classification model showed different performance levels according to salivary hormone data (85.9% vs. 76.1% accuracy), with a 9.8 percentage point improvement when physiological markers were included. This substantial enhancement suggests that physiological markers contribute meaningful information beyond what can be captured through psychological assessments alone. The improvement was particularly pronounced in distinguishing intermediate severity levels (yellow and orange categories), where the addition of salivary hormone data led to substantial increases in both sensitivity and specificity. This differential contribution across severity levels is clinically significant, as distinguishing between mild, moderate, and severe depression has important implications for treatment planning and resource allocation. The finding aligns with previous research indicating that HPA axis dysfunction [14,15], as measured through salivary hormones, may help differentiate various depression severity levels. Particularly, it reflects the heterogeneous nature of depressive disorders.

The near-perfect binary classification performance warrants careful interpretation. It is important to note that the existing rule-based Minds.NAVI system had already demonstrated high binary classification performance in a prior prospective confirmatory clinical trial: sensitivity of 97.2%, specificity of 95.2%, and accuracy of 95.8%. Given this established baseline, it is plausible that a machine learning model trained on refined data could achieve comparable or even higher performance. Additionally, the prior clinical trial evaluated performance using 120 participants, whereas the independent test set in the current study comprised 92 participants; therefore, direct comparison between the two results should be made with caution. To minimize the risk of data leakage, 20% of the total dataset— 92 samples with clinically confirmed ground truth—was separated as an independent test set entirely excluded from the model development process and used only for final performance evaluation. The contrast between high binary classification performance and reduced multiclass accuracy suggests that the high AUC in the binary task reflects the clinical distinctiveness of the two broad categories (normal vs. depression) rather than a ceiling effect from model overfitting. Meanwhile, the reduced sensitivity for intermediate severity categories (particularly orange, sensitivity=0.40) in the multiclass model indicates that genuine prediction uncertainty remains for clinically ambiguous boundary cases, further arguing against systematic overfitting or information leakage.

In contrast to the multiclass results, both models in the binary classification achieved an identical high accuracy of 97.8% (normal vs. depression). The lack of incremental benefit from salivary data in the binary task suggests a ceiling effect, where psychological indicators alone are sufficiently robust for identifying depression presence or absence. The discrepancy between the substantial improvement in multiclass classification and the equivalent performance in binary classification therefore highlights the specific utility of physiological markers: they appear essential for capturing nuanced biological gradations and distinguishing severity levels. This pattern likely reflects the complex, non-linear relationship between HPA axis function and depression severity reported previously [21,22]. Since our stress response system can exhibit either hyperactivity or hypoactivity depending on chronicity and severity, these physiological subtleties, while critical for multiclass differentiation, may not be apparent in a simple binary diagnostic framework.

Our choice of RandomForestClassifier, a traditional machine learning approach, rather than deep learning methods, was deliberate and based on several considerations. First, our relatively small dataset size (480 participants) makes traditional machine learning methods more appropriate than deep learning approaches, as the latter typically require thousands of samples to achieve optimal performance and avoid overfitting. Second, RandomForest models offer better interpretability through feature importance analysis, which is crucial for clinical applications where understanding the variables driving the predictions is essential for building trust and clinical acceptance. Third, the ensemble nature of RandomForest, which combines multiple decision trees, provides inherent robustness and reduces variance compared to single models. Therefore, our implementation of comprehensive hyperparameter tuning via GridSearchCV, combined with a 5-fold cross-validation and class weight balancing, represents a principled approach to model development that addresses the specific challenges posed by our dataset. Deep learning methods can be explored in future studies with larger datasets, while the traditional machine learning approach adopted here aligns well with current recommendations for medical AI applications emphasizing interpretability, validation rigor, and appropriate matching of method complexity to data availability.

Feature importance analysis revealed several key insights. While psychological indicators, particularly the depression total score, remained the strongest predictors, the cortisol- DHEA ratio emerged as the third most important feature in our models. This high ranking of the cortisol-DHEA ratio corroborates recent research highlighting the importance of the HPA axis function in depression. Indeed, salivary cortisol has shown potential as a biomarker for depression, which is supported by our findings that depict its practical utility in a machine learning context.

Additionally, the cortisol-DHEA ratio’s significant contribution to model performance aligns with the growing body of evidence suggesting that HPA axis dysfunction is crucial for depression pathophysiology [14,15,34]. The integration of both psychological symptoms and physiological markers in our model contributed to model performance, which reflected the complex, multifaceted nature of depression, offering a more robust framework for depression screening and severity assessment. The improved performance for depression severity identification when including salivary hormone data suggests that physiological markers might be valuable in capturing subtle variations in depression manifestations that are hard to grasp through psychological assessments alone.

This study has several limitations. First, the relatively small dataset (480 participants) limits statistical power and increases the risk of overfitting despite our use of 5-fold cross-validation, class weight balancing, and GridSearchCV parameter optimization. The class imbalance in the dataset (particularly the small number of green category cases) suggests that performance metrics should be interpreted cautiously. The absence of an external validation cohort further limits the generalizability of our findings; future studies should incorporate independent datasets from different clinical settings to validate model performance, and the lack of extensive clinical validation data means that model performance in real-world clinical settings requires further verification. Second, the demographic composition of the sample was skewed toward younger (mean age: 35.17 years) and predominantly female participants (69.3%), which may limit the applicability of our findings to broader populations, particularly older adults or male-dominant clinical groups. Third, the performance comparison between models with and without salivary hormone data was descriptive rather than supported by formal statistical testing (e.g., paired permutation tests or bootstrapped confidence intervals). The existing Minds.NAVI system’s prospective confirmatory clinical trial was evaluated using approximately 120 clinical data samples; in a retrospective analysis context such as the present study, a substantially larger sample size would be required to achieve statistically stable significance. We estimate that the overall dataset would need to be expanded to at least approximately 1,000 samples to enable adequate train/validation/test splits and reliable statistical verification. Fourth, missing data were handled using mean imputation, which may have artificially reduced variance in the resilience-related variables (res_sc, res_ir, res_po). Although the proportion of missing data was approximately 14% (68 out of 480 samples) and confined to a single domain, more complex imputation methods in a limited sample environment could introduce additional model assumptions or alternative sources of bias. The impact of different imputation strategies on model performance should be examined in future work with sufficiently expanded datasets. Fifth, feature importance was derived using the MDI method, which can be biased toward features with high cardinality or wide value ranges. Permutation-based importance and other alternative methods should be explored in subsequent analyses to ensure the robustness of feature ranking. Sixth, the term “physiological indicators” was used broadly in the original manuscript but refers specifically to salivary cortisol and DHEA measures; we have clarified this terminology throughout the revised manuscript. Finally, although the input variables were selected based on their independence from the outcome variable and the existing algorithm’s computational rules were not included in model training, the dependent variable (Minds.NAVI classification) is ultimately derived from the same domain of clinical indicators. The reported performance metrics should therefore be interpreted with awareness of this structural relationship, and future research should adopt independent clinical outcomes, such as clinician-rated severity assessments (e.g., HAM-D, MADRS), as the primary dependent variable to more rigorously evaluate the model’s true predictive capacity.

Future research should prioritize the collection of larger, multi-site datasets to enable formal statistical comparisons between models and to establish external validation cohorts. In summary, this study has significance in confirming the potential for AI-based approaches to complement or replace existing rule-based systems, even in a data-limited environment. Future studies with larger-scale data collection and external validation are planned to further verify the generalization performance and clinical utility of the model. There are also important implications from a clinical perspective. The ability to distinguish depression severity levels with reasonable accuracy could facilitate more targeted interventions and efficient mental health resource allocation. For instance, individuals in the yellow category (attention needed) might benefit from preventive interventions or watchful waiting, while those in the red category (severe) would require immediate intensive treatment. The implementation of physiological markers can help identify individuals potentially at a higher risk of progression to more severe depression, even when their psychological symptoms appear relatively mild. The non-invasive nature of salivary hormone collection also makes this approach potentially scalable for routine clinical screening, though practical considerations such as the need for standardized collection procedures and the added cost of laboratory analysis must be carefully evaluated.