Predicting Depression in Canadians with or at Risk of Diabetes: A Cross-Sectional Machine Learning Analysis

A. Data Collection and Preparation

We obtained a comprehensive dataset (N = 8,602) from the Canadian Primary Care Sentinel Surveillance Network (CPCSSN), a national repository of de-identified electronic medical records spanning 2003 to 2015 [10]. This cross-sectional dataset comprises 34 features capturing essen-tial demographic attributes (e.g., age, sex), anthropometric measurements (e.g., body mass index), biomarkers (e.g., hemoglobin A1c), comorbidity diagnoses, and prescription records. To ensure clinical relevance, the cohort was restricted to patients with either a confirmed diagnosis of diabetes mellitus (type 1 or type 2) [14] or an elevated hemoglobin A1c (≥ 5.7%) indicative of prediabetes [18], yielding a final analytic sample of 6,219 individuals. Depression labels were derived from clinician-documented di-agnoses, reflecting routine clinical practice.

We evaluated data quality and completeness, noting that fewer than 3% of observations had missing values. Given the minimal scope of missingness, multiple imputation by chained equations (MICE) was applied to the training portion of the data [19]. All continuous variables were normalized to mean zero and unit variance to stabilize subsequent model fitting. An 80:20 train-test split was used, and the training subset exhibited class imbalance (approximately 20% with depression). To manage this, we first removed Tomek links, where a pair (x_i, x_j) from opposite classes satisfies each being the other’s nearest neighbor in feature space. Removing the majority class instance in each Tomek pair eliminates borderline points that can confuse the model’s decision boundary. Next, we employed SMOTE, which synthesizes minority samples using linear interpolations: , where x^(k) is a randomly chosen minority class neighbor of x, and λ∈ [0, 1]. This two-step balance strategy addresses the challenge in clinical data where differences between depressed and non-depressed patients can be obscured by the larger control group [20].

New variables were derived to reflect chronic disease burden, which is clinically salient for mental health outcomes. For each documented comorbidity (e.g., osteoarthritis, hypertension), we computed a duration metric by subtracting the date of diagnosis from the most recent clinical encounter, thus representing the approximate years lived with that condition. Undiagnosed cases were set to zero. Medication records were parsed into 45 binary indicators, excluding categories with fewer than ten occurrences to avoid sparse features. These feature engineering steps were motivated by evidence that longer disease duration, multiple comorbidities, and certain prescriptions may compound the risk of depression in patients with metabolic disorders.

B. Model Development

Recognizing the inherent constraints of primary care datasets, such as modest sample sizes and heterogeneous data structures, our approach prioritizes models that strike a balance between interpretability and robust performance. While recent advances in deep learning and generative modeling have expanded the analytical toolkit, the practical limitations of routine EHR data often favor well-established, transparent methods. Accordingly, we conducted a systematic comparison of six supervised learning algorithms: LR, NB, RF, AB, XGB, and an ANN. Each algorithm was chosen based on its unique trade-off between interpretability, computational complexity, and capacity for capturing non-linear relationships. Formally, given N patients , with x_i ∈ℝ^d representing a vector of demographics, comorbidity durations, and medication indicators, and y_i ∈ {0, 1} indicating a depression diagnosis, these methods learn a mapping f : ℝ^d {0, 1} that generalizes to new individuals.

LR provides a straightforward linear boundary whose coefficients can serve as intuitive risk indicators, while NB uses feature-independence assumptions to yield a generative view of P (y| x). RF, AB, and XGB build ensembles of decision trees, enabling them to adaptively capture interactions among predictors such as age, BMI, and disease history. The ANN leverages a feedforward, multilayer structure to approximate non-linear functions, which can be helpful in capturing more complex relationships. However, as illustrated in Figure 1, larger neural networks may become over-sensitized to noise and outliers when real-world data are limited or contain label inaccuracies. Our intent was to cover a range of baseline architectures suitable for typical EHR scenarios, where advanced deep or generative models often require bigger, more uniform datasets to avoid overfitting or opaque decision boundaries. All source code for data preprocessing, model training, and interpretability analyses is publicly available¹, facilitating replication and further refinement by both clinicians and data scientists.

• Download figure
• Open in new tab

Fig. 1:

A synthetic dataset illustrating typical complexities of clinical data: correlated features, overlapping subpopulations, outliers, and label noise. Each subplot shows a learned decision boundary f : ℝ ² →{0, 1} for a different model (Logistic Regression, Naive Bayes, Random Forest, AdaBoost, XGBoost, a smaller ANN, and a larger ANN). Simpler or ensemble methods yield stable partitions under moderate sample sizes, whereas the larger ANN overfits to minor clusters and mislabeled points. Axis labels are omitted to highlight differences in boundary shapes.

A. Model Benchmarking and Interpretability

Among the six models evaluated, XGBoost emerged as the top performer, achieving an AUC of 0.64 and a weighted average F1 score of 0.72 on the held-out test set. Other methods demonstrated AUCs ranging from 0.54 to 0.64. Although these performance metrics may appear modest relative to more complex deep learning architectures, they are in line with similar studies using routine EHR data for depression screening, where challenges such as imprecise labeling and the absence of detailed psychosocial measures are common [17]. The high precision in identifying non-depressed cases, alongside lower recall for the depressed minority, reflects the intrinsic class imbalance (20.6% depression rate) and the fact that some depressive conditions may remain undocumented or embedded within unstructured clinical notes. Despite these constraints, the consistent detection of several core risk factors, such as BMI, hemoglobin A1c, osteoarthritis, and younger age, suggests that even traditional machine learning algorithms can highlight clinically meaningful patterns in cross-sectional EHR data. This is important since purely deep or generative models may not always yield decisive gains when exposed to the sparse, heterogeneous records found in primary care [21]. Moreover, these reproducible but modest findings provide a transparent benchmark for researchers seeking to refine depression detection methods, underscoring two key points:

B. Feature Importance and Clinical Implications of XGBoost

Figure 2 highlights two essential aspects of the XGBoost model’s behavior on the test set. The upper panel confusion matrix (1,244 test instances) shows 795 true negatives (non-depressed) and 96 true positives (depressed), with relatively few false positives (198) corresponding to a precision of 0.33. However, a recall of 0.38 indicates that 150 depressed patients were missed (false negatives). From a clinical stand-point, this balance may be acceptable as an initial risk stratification tool, given the well-documented challenges of accurately capturing mental health conditions in EHRs.

• Download figure
• Open in new tab

Fig. 2:

Performance metrics for six classifiers (left), XGBoost confusion matrix (top right), and SHAP summary plot (bottom). The matrix shows correctly (TN, TP) and incorrectly (FN, FP) classified patients. In the SHAP plot, each dot’s horizontal position indicates how much that feature pushes the model toward (positive) or away from (negative) a depression prediction, with purple denoting lower feature values and orange denoting higher values.

The lower panel is a SHAP summary plot, which explains feature-level contributions to the model’s predicted log-odds of depression. Specifically, for each instance x_i and feature j, the SHAP value ϕ_j(x_i) quantifies how much x_ij shifts the model’s output relative to a baseline: logi , where ϕ₀ is the average prediction on the training set. Positive SHAP values push the prediction toward ‘depressed’, negative values do the opposite, and the color scale (purple to orange) reflects each patient’s actual feature value. ‘Sex’ (female) and ‘Age at Exam’ (younger) emerged as influential, suggesting that female sex and earlier onset of metabolic conditions may elevate psychosocial burdens [24, 25]. Osteoarthritis (‘OA’) and BMI also significantly increased depression risk, likely reflecting pain, reduced mobility, and social constraints. Systolic blood pressure (‘sBP’) and antihypertensives (e.g., ‘HTRX HYDROCHLOROTHIAZIDE’) further emphasize the link between cardiovascular strain and mood disturbances. Although hemoglobin A1c did not dominate on its own, it contributed synergistically with factors like high BMI, reinforcing that depression risk in diabetic and prediabetic populations arises from multifaceted metabolic profiles rather than glycemic control alone.

Despite these clinically coherent findings, the 155 false negatives emphasize a need for improved data capture, through standardized mental health tools or more frequent longitudinal assessments, to enhance recall. Nonetheless, the SHAP interpretation demonstrates how a well-calibrated machine learning model can provide nuanced, patient-level insights into depression risk, thereby informing targeted interventions in real-world primary care settings.

C. Limitations

Although this study leveraged real-world CPCSSN data to demonstrate the feasibility of depression risk modeling in a diverse adult population [10], several limitations remain. The cross-sectional design precludes capturing temporal changes or identifying whether evolving glycemic or comorbidity profiles trigger depressive symptoms. Reliance on clinician-documented diagnoses introduces potential label bias, as subclinical or undiagnosed cases may be overlooked; integrating standardized tools (e.g., PHQ-9) could improve label reliability. Finally, sparse demographic data restricted sub-group analysis, limiting insights into racial or socioeconomic disparities. Despite these constraints, the findings provide a valuable benchmark for future efforts to refine and expand depression detection in patients with or at risk of diabetes.