Application of machine learning models for predicting depression among older adults with non-communicable diseases in India

Table 1 illustrates both univariate and multivariable logistic regression analyses presenting the unadjusted odds ratio, adjusted odds ratio, 95% confidence interval (CI) and p-value for various factors associated with self-reported depression among individuals aged 45 years and older with NCDs. Each factor is compared to a reference category and the statistical significance of the odds ratios is determined by the p-value.

In terms of the age of participants, the adjusted odds of experiencing depression were elevated in older individuals aged 56–66, 67–77 and 78–88 years when compared to the reference category (aged 45–55 years) with a statistically significant p-value of < 0.001. In the context of gender classification, the adjusted odds of experiencing depression were 1.098 (95% CI: 1.029–1.17) for females, with a p-value of < 0.001 in comparison to males. Regarding the place of residence, there was slightly higher adjusted odds ratio which stood at 1.057 (95% CI: 0.996–1.121) for older adults living in urban areas, with a p-value of < 0.001 suggesting that self-reported depression chances are elevated in urban areas compared to rural areas. In the case of marital status among older adults, the adjusted odds ratio was 1.352 (95% CI: 1.082–1.689) for married older individuals, with a p-value of < 0.001 indicating that self-reported depression probabilities are heightened compared to the reference group. With regards to educational attainment, those who belonged to the “ever attended school” category exhibited slightly higher adjusted odds ratio was 1.081 (95% CI: 1.02–1.147), with a p-value of < 0.001, compared to reference category. The direction of association remained consistent in both univariate and adjusted models. In terms of religious affiliation, the odds of depression were higher among older adults practicing Islam and other faiths when compared to those following Hinduism. The adjusted odds ratio was 1.148 (95% CI: 1.059–1.246) for older adults adhering to Islam and 1.19 (95% CI: 1.101–1.285) for those following other religions, with a p-value of < 0.001. In relation to MPCE quintile, the odds of depression were elevated for older adults in the middle-class household category in contrast to those in the poor-class household category. The adjusted odds ratio was 1.084 (95% CI: 1.011–1.162) for older adults in the “middle” category, all with a p-value of < 0.001. For older adults diagnosed with diabetes, the (lower) adjusted odds ratio stood at 0.906 (95% CI: 0.84–0.978), exhibiting a significant p-value of < 0.001 when compared to peers without diabetes. Similarly, older adults diagnosed with cancer had lower adjusted odds ratio being 0.742 (95% CI: 0.567–0.972), and a p-value of < 0.001 compared to those without cancer. In the case of chronic respiratory diseases, older adults diagnosed with such conditions showed lower adjusted odds ratio was 0.864 (95% CI: 0.782–0.955), with a p-value of < 0.001 against those without chronic respiratory diseases. Older adults facing multiple ADL challenges had three times higher odds of depression than those without such challenges, while the adjusted odds ratio was 1.581 (95% CI: 1.233–2.028), with a significant p-value of < 0.001. Moreover, older adults dealing with mobility issues exhibited higher odds of depression, whether facing single or multiple mobility problems, in comparison to the reference category. The adjusted odds ratio was 1.216 (95% CI: 1.025–1.444) for single mobility and 1.827 (95% CI: 1.579–2.113) for multiple mobility, with a p-value of < 0.001 against older adults without mobility issues. Furthermore, for older adults experiencing challenges in IADL, the odds of depression were notably higher compared to those without such difficulties, with the adjusted odds ratio at 1.214 (95% CI: 1.147–1.284) and a p-value of < 0.001. In the realm of sleep difficulties, older adults had lower adjusted odds ratio i.e., 0.426 (95% CI: 0.402–0.451), with a p-value of < 0.001 from the reference group. When examining self-rated health among older adults, those who perceived their health as “poor” had low adjusted odds ratio was 0.801 (95% CI: 0.759–0.845), with a p-value of < 0.001. Concerning smoking habits among older adults, the odds of depression were elevated in smokers compared to non-smokers, with an adjusted odds ratio of 1.113 (95% CI: 1.048–1.182) and a p-value of < 0.001 relative to the reference category. Similarly, in the context of alcohol consumption among older adults, individuals who consumed alcohol had higher odds of depression than non-consumers. The adjusted odds ratio was 1.147 (95% CI: 1.06–1.24), with a p-value of < 0.001 compared to the reference category. When considering the level of physical activity, inactive older adults had lower adjusted odds ratio i.e., 0.909 (95% CI: 0.857–0.964), with a significant p-value of < 0.001. Exploring the body mass index status among older adults, the lower adjusted odds ratio for underweight older adults was 0.938 (95% CI: 0.877–1.005), all with a p-value of < 0.001.

The results underscore a multifaceted relationship between demographic, health, functional and behavioural aspects linked to depression among older Indian adults. The findings highlight the importance of tailored mental health strategies that consider both clinical and socio-behavioural influences on depression in later life.

Table 2 presents the performance metrics of eight ML models evaluated on a 70/30 train-test split, aimed at predicting depression among older adults with NCDs. The performance metrics include sensitivity, specificity, precision, accuracy with 95% confidence intervals (CI), area under the receiver operating characteristic curve (AUROC) with 95% CI, F1 score, negative predictive value (NPV), Cohen’s Kappa, PR-AUC value, Matthews Correlation Coefficient (MCC) and balanced accuracy.

The random forest model demonstrated the highest predictive performance across almost all metrics. It achieved the highest accuracy of 95.6% (95% CI: 0.954–0.958), AUROC of 0.996 (95% CI: 0.995–0.997), F1 score of 0.954, precision of 0.994 and Cohen’s Kappa of 0.912, indicating excellent agreement beyond chance. Moreover, the PR-AUC value of 0.994 suggests superior performance in precision-recall space, important in datasets with class imbalance. The balanced accuracy was also highest i.e, 0.956, showing the model’s effectiveness in correctly classifying both positive and negative cases. These findings support the random forest as the most robust classifier in this study. Moreover, the decision tree model also performed well, with a balanced accuracy of 0.915, AUROC of 0.915 (95% CI: 0.912–0.918) and accuracy of 91.5% (95% CI: 0.912–0.918). It maintained high precision (0.994) and F1 score (0.908), making it a strong alternative model with slightly reduced complexity compared to random forest.

Among the moderate performers, KNN achieved a balanced accuracy of 0.793, AUROC of 0.877 and F1 score of 0.760, indicating its competence in handling nonlinear patterns. It also maintained a high precision of 0.902 but had lower sensitivity (0.656), suggesting that it may miss some true positive cases. The neural network model had mixed results, with moderate AUROC (0.663) and balanced accuracy (0.619). However, its Cohen’s Kappa (0.239) and MCC (0.239) were higher than other low-performing models, indicating better generalizability than naïve bayes and linear classifiers in this context. On the other hand, logistic regression, support vector machine (SVM), naïve bayes and ridge classifier exhibited limited performance. These models consistently showed low accuracy (~ 0.57–0.60), AUROC values around 0.61–0.63 and PR-AUC values under 0.65. Their low Cohen’s Kappa (< 0.20) and MCC values (< 0.21) indicate weak agreement and predictive power. Although these models had moderate sensitivity (~ 0.73–0.84), their specificity values were substantially lower (ranging from 0.319 to 0.462), which could lead to an unacceptable rate of false positives in public health interventions.

Figure 1 shows the receiver operating characteristic (ROC) curves for all eight predictive models used in this study to classify depression among older adults with NCDs. The area under the ROC curve (AUROC) is a robust metric for assessing model performance, representing the model’s ability to distinguish between classes across various thresholds.

Among the models, the random forest model achieved a perfect AUROC score of 1.00, indicating outstanding classification capability and an exceptional ability to distinguish between individuals with and without depressive symptoms. Similarly, the decision tree model followed closely with an AUROC of 0.92, suggesting high discriminative power with relatively fewer misclassifications. The KNN model also performed strongly, with an AUROC of 0.88, indicating a favourable trade-off between sensitivity and specificity. In contrast, models such as neural network (AUROC = 0.66), logistic regression (AUROC = 0.63), SVM (AUROC = 0.63), ridge classifier (AUROC = 0.63) and naïve Bayes (AUROC = 0.62) demonstrated moderate to low discriminatory performance.

The diagonal dashed line in the plot represents the baseline (AUROC = 0.5), reflecting random guessing. All models performed above this baseline, yet only ensemble and tree-based methods significantly outperformed traditional or simpler classifiers.

Figure 2 presents the precision-recall (PR) curves for all eight machine learning models used to predict depression in older adults with NCDs. The precision-recall area under the curve (PR-AUC) metric provides a robust assessment of model’s performance, particularly useful in imbalanced datasets where the proportion of positive (depressed) cases is low.

The random forest model outperformed all other models, achieving a PR-AUC of 0.99, reflecting its exceptional ability to maintain both high precision and recall across thresholds. This indicates the model can accurately identify individuals with depressive symptoms while minimizing false positives. Moreover, the decision tree model also showed strong performance with a PR-AUC of 0.91, closely matching random forest. This suggests its ability to capture relevant patterns in the dataset despite its comparatively simpler structure. The KNN model yielded a PR-AUC of 0.88, demonstrating solid performance and its effectiveness in handling class imbalance through neighbourhood-based voting. On the other hand, models such as neural network (PR-AUC = 0.65), logistic regression (PR-AUC = 0.61), ridge classifier (PR-AUC = 0.61), support vector machine (PR-AUC = 0.60) and naïve bayes (PR-AUC = 0.60) showed lower discriminatory power in distinguishing positive cases from negative ones, with relatively flat PR curves.

Figure 3 presents a SHAP summary plot, evaluated on a 70/30 train-test split, highlighting the relative importance and directional impact of each feature on the model’s prediction of depression among older adults with NCDs. The x-axis represents the SHAP value (impact on the model’s output), while each dot represents an individual prediction. The colour gradient indicates the original feature value (red = high, blue = low).

Notably, poor sleep emerged as the most influential predictor of depressive symptoms, with higher values (in red) significantly increasing the model’s predicted risk. Other prominent contributors include BMI, age, self-rated health and MPCE quintile, all of which exhibit substantial SHAP values and consistent directional influence across the sample. These top features had broad SHAP distributions, signifying a wide range of influence across individuals. Interestingly, marital status, IADL limitations and education also show meaningful impact, although with slightly lower SHAP values, indicating their moderate predictive utility. Features such as cancer, CVD, ADL status and chronic respiratory diseases demonstrated minimal SHAP values, suggesting limited contribution to model output.

Overall, the SHAP summary plot provides a comprehensive visualization of both the strength and direction of influence of each variable, offering interpretable insights into the model’s decision-making process.

Figure 4 represents the binary-shaded McNemar’s test pairwise comparison matrix of eight machine learning models for predicting depression among older adults using a 70/30 train-test split. Dark red cells indicate statistically significant differences (p ≤ 0.05) between model pairs, while light grey cells denote non-significant differences (p > 0.05). The diagonal is not relevant for interpretation, as it represents each model compared with itself.

The McNemar’s test pairwise comparison showed a heterogeneous pattern of statistically significant differences among the eight models. Logistic regression differed significantly (p ≤ 0.05) from decision tree, random forest, KNN and naïve bayes, while comparisons with ridge classifier, SVM and neural network were non-significant (p > 0.05). Ridge classifier showed significance against decision tree, random forest, KNN and naïve bayes, but not against logistic regression, SVM and neural network. Decision tree was significantly different from all other models (logistic regression, ridge classifier, random forest, SVM, KNN, naïve bayes and neural network). Random forest likewise differed significantly from all other models. SVM showed significant differences versus decision tree, random forest, KNN and naïve bayes, but not versus logistic regression, ridge classifier and neural network. KNN differed significantly from all other models. Naïve bayes also differed significantly from all other models. Finally, neural network showed significance against decision tree, random forest, KNN and naïve bayes, with non-significant differences versus logistic regression, ridge classifier and SVM. Overall, significant pairwise differences were denser in the central/right regions of the matrix, suggesting clustering of comparatively stronger-performing models in those zones.

In summary, the McNemar’s test confirms the robustness and superiority of the random forest model, with statistically significant differences in classification outcomes compared to other models, reinforcing its suitability for deployment in predictive modelling of depression among older adults.

Table 3 represents the comparative performance of eight ML models i.e., decision tree, random forest, logistic regression,  SVM,  KNN, naive bayes, neural network and ridge classifier for predicting depression among older adults with NCDs in India, based on stratified 10-fold cross-validation using all predictors.

Among the models evaluated, the random forest model consistently outperformed all others across nearly every evaluation metric. It achieved the highest accuracy (89.5% with 95% CI: 0.893–0.897), AUROC (0.955), F1 score (0.898), Cohen’s Kappa (0.791), MCC (0.792) and PR-AUC (0.945). Additionally, it demonstrated strong sensitivity (0.919) and specificity (0.872), indicating high capability in correctly identifying both depressed and non-depressed individuals. The model’s balanced accuracy (0.895) reinforces its superior and well-calibrated performance. Similarly, the decision tree model also showed strong performance, particularly in sensitivity (0.838), precision (0.854) and F1 score (0.845). Its AUROC (0.847) and accuracy (84.7%) were moderately high, indicating it is a reliable, interpretable alternative to more complex models. Although not as powerful as random forest, decision tree maintained moderate agreement (Cohen’s Kappa = 0.694) and classification power (MCC = 0.694), showing its utility as a lightweight yet effective model. KNN demonstrated good performance, especially in specificity (0.929), precision (0.905) and AUROC (0.885). With a balanced accuracy of 0.803 and F1 score of 0.774, KNN proved to be a strong contender, particularly for minimizing false positives. Its MCC of 0.626 and Cohen’s kappa value of 0.605 indicate substantial agreement between predictions and actual outcomes. SVM performed reasonably well, achieving accuracy of 71.9%, AUROC of 0.798 and F1 score of 0.734. While its sensitivity (0.775) and specificity (0.663) were lower than random forest, SVM maintained acceptable predictive consistency (Cohen’s kappa = 0.438; MCC = 0.441). Neural network model performed moderately well, with a balanced accuracy of 0.742, AUROC of 0.825 and F1 score of 0.739. Although it lagged behind random forest and KNN, it outperformed traditional models in most metrics. Naive bayes and ridge classifier showed relatively poor performance with lowest specificity and lowest accuracy. Logistic regression demonstrated limited predictive power, with accuracy (59.3%), AUROC (0.632), F1 score (0.629) and PR-AUC (0.614) all relatively low. It also had the lowest Cohen’s kappa (0.186) among all non-naive models, confirming its inadequacy for complex nonlinear relationships in this context. These results support the adoption of ensemble and nonlinear models for public health prediction tasks involving complex behavioural and clinical data.

Figure 5 illustrates the receiver operating characteristic (ROC) curves and corresponding area under the ROC curve (AUROC) values for eight ML models developed to predict depression among older adults in India. The evaluation was performed using a stratified 10-fold cross-validation approach to ensure balanced representation of depressed and non-depressed cases across folds.

Among these, the random forest model achieved the highest AUROC of 0.96, demonstrating excellent discriminatory ability. This was followed by KNN (AUROC = 0.88), neural network (AUROC = 0.83), SVM (AUROC = 0.80) and decision tree (AUROC = 0.85), all of which exhibited strong performance above the acceptable threshold of 0.80. In contrast, logistic regression (AUROC = 0.63), ridge classifier (AUROC = 0.63) and naïve bayes (AUROC = 0.62) showed considerably lower performance, only marginally better than the random guessing baseline (AUROC = 0.50).

All models outperformed this threshold, though only a subset (random forest, KNN, decision tree, neural network) demonstrated clinically and statistically meaningful predictive value. These findings highlight the value of advanced ML approaches in handling complex health data.

Figure 6 presents the precision-recall curves for eight ML models used to predict depression among older adults in India. The models were evaluated using stratified 10-fold cross-validation and their respective PR-AUC (Area Under the Precision-Recall Curve) values are shown in the figure legend. PR curves are particularly useful for evaluating classification performance on imbalanced datasets, where the number of non-depressed individuals may substantially outweigh those with depression.

Random forest model demonstrated the highest PR-AUC of 0.94, indicating superior performance in maintaining high precision across a wide range of recall values. This suggests that the model is highly effective at correctly identifying true cases of depression while minimizing false positives, even in a class-imbalanced setting. KNN achieved the second-best performance with a PR-AUC of 0.88, followed by the neural network (PR-AUC = 0.82) and decision tree (PR-AUC = 0.80). SVM model showed moderate performance, with a PR-AUC of 0.79, indicating a reasonable balance between precision and recall, though not as robust as the top-performing models. In contrast, logistic regression, naive bayes and ridge classifier all exhibited lower PR-AUC values of 0.61, suggesting limited effectiveness in handling the imbalanced nature of the depression prediction task. These models struggled to maintain high precision at higher levels of recall, implying that they are more likely to misclassify non-depressed individuals as depressed, or vice versa.

Moreover, random forest’s superior PR-AUC confirms its robustness, especially in real-world deployment scenarios where identifying depressed individuals with minimal false alarms is critical.

Figure 7 presents the SHAP summary plot derived from eight ML models trained to predict depression among older adults in India, using a stratified 10-fold cross-validation approach. SHAP values provide a consistent and model-agnostic interpretation of feature importance by quantifying each variable’s contribution to the prediction outcome for individual samples. Each dot represents a SHAP value for a feature and an individual sample. Red indicates higher feature values and blue indicates lower values. Features are ordered by their average importance (mean absolute SHAP value) and pick top predictors of depression.

Poor sleep emerged as the most influential predictor of depression. The SHAP values are predominantly positive for high values of poor sleep (shown in red), suggesting that individuals who report poor sleep significantly contribute to the model’s prediction of depression risk. BMI, self-rated health and age also showed substantial impact on depression prediction. For example, higher BMI (red) appears to slightly reduce depression risk in some cases (negative SHAP values). Poor self-rated health (red) contributes positively to the predicted risk of depression. Older age (red) generally increases depression risk, although some blue dots indicate protective effects in younger individuals. Marital status, IADL limitations, MPCE quintile and education were moderately important. Smoking and alcohol consumption (red) generally increased the model’s prediction for depression. Being physically inactive (red) was positively associated with depression risk. Functional limitations like difficulty in mobility, ADL status and chronic respiratory diseases also played relevant roles, although with more modest SHAP impacts. Less influential variables (towards the bottom of the plot), such as social activity status, diabetes, CVD and cancer had relatively smaller and more centred SHAP distributions, indicating minimal and inconsistent influence on model outputs. This SHAP summary plot enhances interpretability by revealing not only which variables were most predictive, but also how their values influenced predictions.

Figure 8 represents the binary-shaded McNemar’s test pairwise comparison matrix for the eight ML models used to predict depression among older adults in India. The evaluation is based on stratified 10-fold cross-validation. McNemar’s test is a non-parametric statistical test used to determine whether the performance differences between two classifiers on the same dataset are statistically significant, based on their misclassification patterns. Dark red cells indicate statistically significant differences (p ≤ 0.05) between model pairs, while light grey cells denote non-significant differences (p > 0.05).

Logistic regression differed significantly (p ≤ 0.05) from decision tree, random forest, KNN, naïve bayes and neural network, while comparisons with ridge classifier and SVM were non-significant (p > 0.05). Ridge classifier was significantly different from decision tree, random forest, KNN, naïve bayes and neural network, but not from logistic regression and SVM. Decision tree showed significant differences versus all other models. Random forest likewise differed significantly from all others. SVM was significantly different from decision tree, random forest, KNN, naïve bayes and neural network, with non-significant differences versus logistic regression and ridge classifier. KNN differed significantly from all models except neural network. Naïve bayes showed significant differences versus every other model. Neural network differed significantly from logistic regression, ridge classifier, decision tree, random forest, SVM and naïve bayes, but not from KNN. Collectively, the only non-significant pairs were (logistic regression vs. ridge classifier), (logistic regression vs. SVM), (ridge classifier vs. SVM) and (KNN vs. neural network), all remaining off-diagonal comparisons were significant.

Feature selection

Figure 9 illustrates the relative importance of features in identifying depression among older adults in India, as determined by the information gain through random forest model trained on a 70/30 train-test split. Feature importance here is computed based on the mean decrease in gini impurity, a metric that evaluates how well each variable contributes to improving the purity of nodes in the decision trees of the ensemble model.

Among all the features, BMI and age emerged as the most dominant features, with substantially higher importance scores than all other variables. This highlights the strong role of physical health and ageing in shaping depression risk among older adults. MPCE quintile and poor sleep were the next most influential factors, suggesting that both economic resources and sleep quality meaningfully contribute to depressive outcomes. Additional features including religion, place of residence, smoking, education and IADL limitations also contributed to model performance reflecting the interplay of social, behavioral and functional health dimensions. Variables such as gender, diabetes and physical activity were included among the top features but carried relatively lower importance compared to BMI, age and economic/ health related measures.  Other variables like ADL status, chronic respiratory diseases, cancer and CVD had relatively low importance scores, indicating their limited predictive value for depression within this modelling framework.

Table 4 represents the classification performance of eight ML models trained using the top 12 features identified through random forest-based information gain. This reduced-feature modelling approach aimed to improve interpretability while maintaining robust predictive performance.

Among all the models, random forest remained the most effective model across nearly all metrics. It achieved the highest accuracy (93.8%), AUROC (0.994), PR-AUC (0.993), F1 Score (0.934), Cohen’s Kappa (0.876) and MCC (0.881), highlighting its superior ability to distinguish between depressed and non-depressed individuals using a limited set of informative features. Similarly, decision tree also performed well with an accuracy of 91.1%, AUROC of 0.911 and F1 score of 0.904. While slightly lower than random forest, its performance was robust, offering interpretability with good sensitivity (0.828) and extremely high specificity (0.995). KNN model demonstrated strong balanced accuracy (0.787) and AUROC (0.875), indicating a good balance between sensitivity (0.644) and specificity (0.930). It outperformed both SVM and logistic regression in terms of discrimination ability. SVM model achieved moderate results with a sensitivity of 0.847 and F1 score of 0.670. However, its specificity was low (0.318), suggesting a tendency to overpredict depression cases, leading to a moderate overall accuracy (58.3%). Neural network yielded a moderate AUROC of 0.629 and PR-AUC of 0.610, with balanced sensitivity (0.812) and F1 score (0.663), but struggled with specificity (0.362), reflecting challenges in identifying true negatives. Logistic regression, ridge classifier and naive bayes performed poorly compared to tree-based and KNN models. These models reported low AUROC values (around 0.61–0.63), precision (~ 0.56) and F1 scores (~ 0.64), indicating limited ability to handle complex, nonlinear relationships even with the top features selected. Cohen’s kappa and MCC scores were highest for random forest (0.876 and 0.881, respectively), reflecting strong agreement between predicted and true labels.

These results indicate that using a reduced set of 12 top features still preserves substantial predictive power, especially for ensemble-based methods like random forest. These findings underscore the value of machine learning feature selection in optimizing both model efficiency and interpretability.

Figure 10 presents the ROC curves of eight ML models evaluated using the top 12 most important features identified by random forest’s information gain method. The models were tested on a 70/30 train-test split. The ROC curve illustrates the trade-off between sensitivity (true positive rate) and 1-specificity (false positive rate) across different classification thresholds, with the area under the ROC curve (AUROC) mentioning as a summary performance metric.

Among all the models, random forest demonstrated exceptional discriminative power, achieving the highest AUROC of 0.99, indicating near-perfect classification performance. Its ROC curve closely touches the top-left corner, reflecting high sensitivity and specificity regardless of the threshold. Meanwhile, decision tree also performed strongly, with an AUROC of 0.91, suggesting effective separation between the depressed and non-depressed classes even when using a limited feature set. KNN attained a competitive AUROC of 0.88, closely following the performance of the decision tree model. This highlights KNN’s ability to generalize well using the top 12 predictive features. Neural network achieved a moderate AUROC of 0.63, marginally outperforming linear classifiers but still indicating limited ability to separate the two classes based on the reduced feature set. Logistic regression, SVM, naive bayes and ridge classifier each reported low AUROC values of 0.61–0.63, with ROC curves falling close to the diagonal reference line. This reflects performance near random chance and indicates these models struggled to learn discriminative patterns from the selected features. This analysis reaffirms that careful feature selection using embedded methods like information gain can lead to highly parsimonious yet powerful model suitable for real-world implementation in population-level mental health screening.

Figure 11 illustrates the precision-recall (PR) curves for eight ML models trained using the top 12 features identified via random forest’s information gain approach, evaluated on a 70/30 train-test split. The precision-recall area under the curve (PR-AUC) is a critical performance metric, particularly in imbalanced classification problems such as predicting depression, where the depressed individuals warrant heightened attention.

Among all the models, random forest achieved the highest PR-AUC of 0.99, indicating exceptional precision across a wide range of recall values. This underscores the model’s robustness in correctly identifying positive cases of depression while maintaining minimal false positives. Meanwhile, decision tree also demonstrated strong performance with a PR-AUC of 0.91, suggesting a favourable balance between sensitivity and precision. Its curve remained consistently above most other models, indicating reliable positive class identification. KNN followed with a PR-AUC of 0.87, showing competitive performance and an ability to generalize effectively using the reduced feature set. Neural network achieved a modest PR-AUC of 0.61, indicating intermediate performance with less stable precision across recall levels. Logistic regression, naive bayes, ridge classifier and SVM performed poorly, each yielding PR-AUC values between 0.58 and 0.59, suggesting difficulty in distinguishing true positives from false positives, especially in imbalanced settings. These findings highlight the importance of selecting both appropriate features and model architectures in sensitive health-related prediction tasks.