Figure 1 is a graphical representation of our methodology. We acquired data from the Depresjon dataset and conducted exploratory data analysis to understand the nature of the data. We observed a class imbalance, which is addressed using sampling techniques, followed by normalization of data. We then applied feature engineering to extract features from the data. This was followed by designing models for binary (depressed vs nondepressed) and multiclass (severity levels) classification tasks. After model implementation, we used accuracy, precision, recall, specificity, MCC, and F₁-score as performance evaluation criteria. Finally, we applied the SHAP and LIME methods to explain the decision-making process of the selected model.

The dataset adopted in this work is the “Depresjon” (Norwegian word for depression) dataset, available at Simula repository [36]. The dataset consists of motor activity of individuals recorded per minute for several days by using an actigraphy watch. The dataset consists of 2 parts: The first part comprises measurements of activity levels related to each individual in the condition and control group. The files consist of timestamp, date, and activity columns. The second part comprises the demographic information related to each participant. The dataset is enriched with several variables: an identification number (number), days with measurements (days), gender (1 for female and 2 for male), and age group of the participant (age). It also records the type of affective disorder (afftype). The presence of melancholia is noted with 1 for melancholia and 2 for lack of melancholia (melanch). The dataset differentiates between inpatients ‘1’ and outpatients ‘2’ (inpatient). Education level is categorized by years completed (education). Marital status is recorded as 1 for married or cohabiting and 2 for single (marriage). Work status is coded as 1 for working or studying and 2 for unemployed or sick leave or pension (work). Finally, the dataset includes MADRS scores before (madrs1) and after (madrs2) the inspection.

The original dataset comprised 2 folders (condition and control); it was important to merge the data into a single coherent set. The resultant dataset after the merge operation has 1440 data points. For binary classification (depressed vs nondepressed), the “condition” group is labeled as 1, and the “control” group as 0. For multiclass classification (assessing the severity levels of depression), the labels were assigned as follows: 0 for normal, 1 for mild depression, and 2 for moderate depression. Participants were categorized based on their MADRS score: 0‐6 for “Normal” (no depression), 7‐19 for “Mild” depression, and 20‐34 for “Moderate” depression. Note that although the standard MADRS scale includes a “Severe” category for scores >=35, our dataset contained no cases with scores in the severe range, so no “Severe” class was included.

We observed that the data is imbalanced, which means that some output classes are more prevalent than others. Imbalance data can skew the performance of ML algorithms and require redressal. We used ADASYN for class imbalance problems. The intuition behind ADASYN is the usage of a weighted distribution for the less frequent instances in accordance with their difficulty of learning [37]. For the working of ADASYN, we refer the reader to He et al [37].

The dataset consists of activity measurements recorded by an actigraph watch, resulting in a continuous stream of data collected at regular intervals for each patient. However, using these raw time series values is not an effective approach, as it can lead to an unnecessarily large number of interconnected features and fails to capture the nuanced behaviors and physiological rhythms indicative of mental health status. In the context of depression and related mood disorders, subtle variations in daily activity patterns, energy distribution across frequencies, and nonlinear fluctuations are particularly revealing. To address this, we summarized the continuous time-series data into key statistical features, thereby translating raw sensor data into indicators more closely tied to psychological and clinical insights. This feature engineering approach is a standard mechanism for managing continuous data streams while preserving essential patterns and variability [38], effectively reducing data dimensionality and computational requirements.

To capture the multifaceted nature of human activity and its potential connections to depressive symptoms, we systematically extracted features from 5 broad categories: (1) time domain, (2) frequency domain, (3) circadian rhythm analysis, and (4) transition features. Each category addresses different aspects of activity that can reflect mood changes or disruptions in daily routines often associated with mental health conditions.

Time-domain features describe statistical properties and variability in the activity signal. Disruptions or heightened variability in these measures can be linked to changes in energy levels and daily routines often observed in individuals experiencing mood disturbances. Time-domain features (such as mean activity, variability, higher-order moments, autocorrelation, and entropy) have long anchored actigraphy research. Historic studies show these measures consistently distinguish depressive from nondepressive motor patterns, providing a compact, behaviorally interpretable window on psychomotor change [25-27,35,38].

The following time domain features are extracted from the continuous stream of data.

Analyzing the distribution of activity power across different frequencies can illuminate underlying periodicities. Alterations in these periodic components often correspond to disrupted sleep-wake cycles or abnormal energy distribution, which are common in depression. Multiple actigraphy studies link these spectral abnormalities to symptom severity, confirming their diagnostic value for mood inference [25,29,35,38].

The features extracted are:

Mood disorders are often accompanied by disruptions in circadian rhythms. Examining the timing of key transitions can yield valuable clues about sleep-wake irregularities [40] (Activity Onset and Offset: identify the start and end of the primary daily activity period. Shifts in these values can reveal delayed or advanced phases of sleep, which are frequently observed in individuals with depression or other affective disorders).

Activity does not always change smoothly throughout the day; abrupt transitions or steady patterns can be particularly telling of mental health status. Transition entropy captures the uncertainty or unpredictability inherent in the sequence of activity state transitions [41]. Transition entropy quantifies the randomness in the transitions between different activity states by evaluating the distribution of transition probabilities. Elevated transition entropy may reflect instability in daily routines or mood fluctuations, which are often observed in individuals with certain mental health conditions.

Alongside the time- and frequency-domain metrics, we also tested a set of nonlinear descriptors (notably Fractal Dimension and the Hurst Exponent) to better capture the self-similar structure and long-range dependencies often observed in human activity signals. We implemented several published algorithms and tuned their parameters across multiple window lengths. Despite these efforts, the extraction pipeline consistently produced a large fraction of undefined or zero values, signaling numerical instability with our data resolution. After troubleshooting and sensitivity checks, the issue persisted and threatened to compromise both model robustness and interpretability. To safeguard the overall validity of the study, we therefore documented the attempted nonlinear analyses and excluded these features from the final feature set. Figure 2 is a pictorial representation of the dataset after feature engineering.

After feature engineering, we performed data normalization using z-score normalization technique as shown in Eq. (1).

Note that xi and zi represent the individual feature score before and after standardization, whereas μ and σ represents the mean and standard deviation of the activity level scores.

In this study, we use accuracy, recall, specificity, precision, F₁-score, and MCC as performance evaluation criteria for ML models. These performance evaluation measures are based on the confusion matrix, which represents a summary of accurate and misclassified instances for a classification problem. A graphical representation of the confusion matrix is given in Figure 3. A confusion matrix has 4 key elements called true positive (TP), false positive (FP), false negative (FN), and true negative (TN), defined as follows [51].

TP is an instance when the model classifies a positive instance correctly. FP is an instance when the model classifies a negative instance as a positive instance. FN is an instance when the model classifies a positive instance as a negative instance. TN is an instance when the model classifies a negative instance as a negative instance. Based on the definitions of various elements of the confusion matrix, in Table 1, we define our performance evaluation metrics.

Table 2 highlights the performance of various ML models on the binary classification task. Among the evaluated models, XGBoost achieved the highest accuracy of 84.94%, significantly outperforming all other models. In contrast, LR was the least effective, with an accuracy of 70.68%. This ranking based on accuracy is consistently reflected across other metrics: XGBoost also leads in precision (83.40%), recall (87.45%), F₁-score (85.29%), and MCC (0.7012), while LR scores the lowest in each of these categories. The remaining models, including SVM, RF, and neural networks, show a progressive improvement in performance, with SVM outperforming LR and RF and neural networks closely trailing XGBoost. This uniform trend across all evaluation metrics underscores the robustness of XGBoost as the superior model for this classification task.

Furthermore, XGBoost exhibited an enhancement over the average performance of the other models across all evaluation metrics. Specifically, XGBoost surpassed the average accuracy, precision, recall, F₁-score, and MCC by 7.2%, 6.42%, 6.83%, and 17.47%, respectively. This improvement underscores XGBoost’s superior capability in capturing complex patterns and interactions within the data, making it the most effective model in our study. The findings demonstrate the effectiveness of ensemble methods like XGBoost in outperforming traditional and other ML models, thereby offering enhanced predictive accuracy and reliability.

Table 3 presents the performance of various ML models on the multiclass classification task. XGBoost achieved the highest accuracy of 85.91%, along with precision, recall, and F₁-scores exceeding 85%, and a MCC of 0.7895. In contrast, SVM and LR recorded the lowest accuracies of 53.22% and 54.12%, respectively, and underperformed across other metrics as well. RF and neural networks demonstrated strong performance with accuracies of 82.85% and 83.52%, respectively, and consistently higher precision, recall, F₁-score, and MCC. These results indicate that XGBoost surpassed the average accuracy, precision, recall, F₁-score, and MCC by 20.35%, 20.13%, 20.34%, 20.41%, and 33.07%, respectively.

To verify that the superior performance of XGBoost was not due to random chance, we conducted rigorous statistical significance testing, as detailed below. We ran a 10×3 repeated stratified cross-validation protocol (30 folds in total), preserving the original class distribution in every split. Within each fold, the model was fit to the training indices and its binary predictions on the held-out set were scored with the F₁-metric, which is appropriate for moderately imbalanced problems. We then applied paired 2-tailed t tests between XGB and each rival model, followed by Holm correction to control the family-wise error rate at α=.05.

For binary classification, across the 30 cross-validation folds, XGB achieved the highest mean F₁-score (0.851 ± 0.034) while its nearest competitor RF averaged 0.832; SVM, LR, and NN trailed at 0.659, 0.723, and 0.687, respectively. Paired significance testing confirmed that these differences are not due to sampling noise: the Holm-adjusted P values for XGB versus RF, LR, SVM, and NN were 3.6×10⁻⁴, 2.3×10⁻¹⁶, 2.7×10⁻¹⁷, and 1.7×10⁻²⁰, all well below 0.05, confirming that the superior performance of XGBoost is not by chance. Hence, XGB’s performance advantage is statistically significant against every other baseline in the study, supporting its selection as the primary model for downstream explainability analysis. The same trend was observed for multiclass classification.

In this section, we critically interpret XAI techniques (SHAP and LIME) to reveal the feature patterns that drive the depression-classification models, compare these explanations across XGBoost and neural-network architectures, and relate the findings to existing clinical evidence. One of the key critiques of ML models is their “black-box” nature [52]. ML models evolved over the last decade to become more complex, resulting in improved performance by virtue of modeling intricate relationships among the input features. However, the increased complexity also meant lower interpretability and explainability powers. In many use cases such as health care, the lack of transparency in decision making is critical, and the reasoning behind a model’s outcome is crucial for trust, accountability, and decision-making [52]. A variety of techniques is proposed in XAI to enhance the explainability of complex ML models. SHAP is a type of XAI framework derived from the concept of coalition game where a prediction is regarded as the “payoff” and the worthiness of the feature is assumed as the “player” [53]. It calculates the Shapley values and explains the prediction by showing how the features influenced the output. The importance of each feature that influenced the effectiveness of the ML algorithm is calculated as given in Eq. (2).

Where: f represents the prediction model, x is the input feature vector, N is the set of features, S is a subset of features excluding the feature i that is, feature under XAI analysis, xs is the feature vector with features in subset S replaced by baseline values and xi is the value of feature i in the input vector x. For XAI analysis, we restrict to the Xgboost model only as it was identified as the best performing model.

LIME is a popular framework in XAI that helps make ML predictions easier to understand. LIME explains a model’s predictions by simplifying it in the local context of a specific instance. It does this by slightly altering the input data around the instance and analyzing how these changes affect the model’s predictions. This process highlights which features are most influential in the local neighborhood. To determine the importance of each feature, LIME builds a simple, interpretable model (such as a linear regression model) based on weighted data points. These weights are assigned according to how close the altered data points are to the original instance.

Two primary SHAP plots (Figures 4 and 5) indicate how each feature affects the model’s decision boundary across the entire dataset. Figure 4 shows the average magnitude of SHAP values (mean absolute SHAP) per feature, highlighting the top predictors. Figure 5 depicts a dot summary where each point corresponds to a specific sample’s SHAP value. Points to the right (positive SHAP) push the classification toward “Depressed,” whereas points to the left (negative SHAP) favor “Non-Depressed.” Colors (red to blue) represent the actual feature value (eg, higher PSD Mean in red and lower PSD Mean in blue).

From these plots, we observe the following. PSD means consistently have the largest average impact. Larger values of the PSD mean (possibly reflecting a stronger low-frequency power in the activity data) correlate with a higher likelihood of depression. Mean activity level and age follow closely, indicating that demographic factors (eg, older individuals) and overall daily activity levels significantly shape the classification. Activity offset and AutoCorr also stand out, aligning with clinical findings that disrupted or rigid circadian rhythms (late offsets and high autocorrelation) are often associated with depressive behaviors.

PSD Mean consistently has the largest average impact. Larger values of the PSD mean (possibly reflecting a stronger low-frequency power in the activity data) correlate with a higher likelihood of depression.

Mean activity level and age follow closely, indicating that demographic factors (eg, older individuals) and overall daily activity levels significantly shape the classification.

We applied LIME to explain a single test instance in Figure 6. LIME creates locally linear “rules” indicating how small perturbations around this instance shift its predicted probability for depression. Features with red bars push the instance toward “Depressed,” while green bars push it toward “Non-Depressed.” For example, a condition like “Gender≤1.0” (female) or older “Age” might increase the model’s belief in depression if other activity-related features also align. LIME reported numeric weights for each condition, revealing how thresholds in AutoCorr, Entropy, or Skewness shape local decision boundaries. These rules coincide well with the SHAP findings (eg, higher autocorrelation frequently signals depressive risk).

While SHAP and LIME are theoretically different (SHAP uses game-theoretic Shapley values, and LIME uses local linear approximations), both converge on similar high-impact features, boosting confidence in the model’s learned representations. Specifically, PSD mean, mean activity level, and daily rhythm indicators (like activity offset, AutoCorr) repeatedly appear among the strongest predictors.

We extended our explainable AI approach to a 3-class setting, distinguishing among “Normal Case,” “Mild Depression,” and “Moderate Depression.” SHAP was used for global insights (per-class and overall feature importance), while LIME was used for local (instance-level) explanations. Figure 7 displays the mean absolute SHAP values aggregated in all 3 classes, highlighting key characteristics that consistently influence the predictions of the model. Notably, AutoCorr and age emerge as dominant factors, aligning with clinical observations that older age groups, coupled with repetitive or rigid activity patterns, may exhibit a higher risk of depression.

For a more in-depth analysis of each class, we generated separate SHAP summary plots. Normal (Figure 8A): Lower AutoCorr typically favors the normal class, suggesting that a more flexible or varied daily rhythm (less repetitive activity) aligns with nondepressed status. Moderate values of age and PSD mean also push predictions away from mild or moderate depression.

Mild depression (Figure 8B): Mild cases show strong contributions from AutoCorr and transition entropy. In many instances, moderate daily rhythms—coupled with subtle changes in the frequency composition of activity (as captured by PSD Mean)—tilt the model toward mild depression. Moderate depression (Figure 8C): This class often manifests higher or more entrenched irregularities. Age, AutoCorr, and an elevated PSD mean appear to push an instance from mild toward moderate severity. Here, demographic and circadian disruptions intertwine more intensively, reflecting a deeper depressive state.

Across all classes, AutoCorr consistently ranks among the most critical features. However, its specific influence varies. Moderate autocorrelation can lead a case from normal to mild, while higher autocorrelation plus advanced age often indicates moderate depression. These distinctions underscore how the same feature can have class-dependent impacts based on subtle variations in its actual range or interaction with other features.

To validate SHAP’s per-class findings at the instance level, we used LIME for a single test example. Figure 9 focuses on the model’s explanation for the “Moderate Depression” class. LIME produces threshold-based “rules” (eg, “0.74<AutoCorr≤0.78”) that locally shift the probability of class membership. Green bars push the model toward selecting “Moderate Depression.” For instance, an AutoCorr range above 0.74 or age exceeding a certain threshold may substantially raise this class probability. Red bars pull the prediction away from “Moderate Depression,” indicating features or thresholds that conflict with the typical profile for moderate severity (eg, a lower PSD mean or higher Kurtosis).

This local view converges with the SHAP observations: circadian-related variables (eg, AutoCorr and Transition entropy) and demographic factors (eg, age) become pivotal. In mild cases, these metrics may appear only partially disrupted, whereas for moderate severity, they consistently register in ranges linked to stronger depressive indications.

Although XGBoost delivered the highest accuracy, we repeated the XAI analysis for artificial-neural-network (ANN) to test whether the explanatory story was model-dependent. For binary classification, both models ranked PSD mean as the dominant driver of the depressed versus nondepressed decision, confirming the feature’s robustness. Beyond that, the explanations diverged: XGBoost distributed moderate importance across roughly ten additional variables (eg, mean, age, activity offset, and AutoCorr), whereas the ANN concentrated nearly all remaining attribution on just three features (activity onset, kurtosis, and activity offset). This comparison strengthens interpretability in 2 ways: the cross-model agreement on PSD mean validates it as a consistent physiological marker and the differing secondary patterns reveal architecture-specific sensitivities, offering complementary insights that would be missed if explanations were reported for XGBoost alone.

For binary classification, both models ranked PSD mean as the dominant driver of the depressed versus nondepressed decision, confirming the feature’s robustness. Beyond that, the explanations diverged: XGBoost distributed moderate importance across roughly ten additional variables (eg, mean, age, activity offset, and AutoCorr), whereas the ANN concentrated nearly all remaining attribution on just 3 features (activity onset, kurtosis, activity offset). This comparison strengthens interpretability in 2 ways: (1) the cross-model agreement on PSD mean validates it as a consistent physiological marker; and (2) the differing secondary patterns reveal architecture-specific sensitivities, offering complementary insights that would be missed if explanations were reported for XGBoost alone.

For multiclass classification, inspection of the mean-|SHAP| values reveals that XGBoost and the neural network rely on partly overlapping but ultimately different information in multiclass classification, and they do so on strikingly dissimilar attribution scales. XGBoost’s strongest signals come from temporal-structure and demographic variables—AutoCorr carries the largest impact, closely followed by age, with gender and transition entropy also substantial—whereas features such as activity offset contribute almost nothing. The ANN tells a nearly opposite story: its 2 most influential cues are the higher-order statistic kurtosis and activity offset, while the variables that dominate XGBoost shrink to the noise floor. Both models nevertheless converge on PSD mean as a relevant marker, which bolsters confidence in that spectral feature’s physiological validity even though its weight is modest in XGBoost and comparatively prominent in the ANN. It is also worth noting that the ANN’s absolute SHAP magnitudes are roughly 2 orders of magnitude smaller than the tree ensemble’s, reflecting the smoother decision surface learned by the network and the conservative estimates produced by KernelExplainer; consequently, comparisons should be made within each model’s scale rather than across scales. Taken together, the agreement on PSD mean and the divergence elsewhere show that the 2 architectures extract complementary patterns—trees capitalize on heterogeneous temporal-demographic interactions, whereas the network distills the data into distribution-shape metrics—providing a richer, cross-validated picture of how depression severity is encoded.

The combined SHAP–LIME analysis reveals several consistent themes.