Using data from the publicly available National Health and Nutrition Examination Survey (NHANES) dataset, the aim of the study was to develop a machine learning model to predict Patient Health Questionnaire (PHQ-9) scores, a clinical tool to screen depression, using urinary incontinence (UI) features. We hypothesized that PHQ-9 score can be predicted using the demographic and subjective UI features, especially those reflecting the impact on social dynamics and daily activities. This is the first study to leverage advanced machine learning algorithms to identify an important condition associated with UI.

The dataset was procured from nationally representative NHANES responses between 2008 to 2018. Inclusion criteria were any individual older than 20 years with demonstrated incontinence and complete responses to the urology questionnaire (KIQ) and depression screener (DLQ). The feature included both UI and non-UI features. Non-UI features included demographics (DEMO), medical conditions including hypertension and diabetes (MCQ, BPQ, and DIQ), healthcare utilization and access cases (HUQ), disability questionnaire (DIQ), and smoking status (SMQ). Numerical responses to the depression screener (DPQ) were summed to determine the overall PHQ-9 score. A total of 47 UI and non-UI features were included in the machine learning model. Selected urology variables were logically recoded, and an incontinence severity index (ISI) was calculated. Two feature sets were evaluated with the machine learning model: UI and non-UI features versus UI-only features. 

We utilized a boosted decision tree-based architecture known as a Light Gradient Boosting Machine (LightGBM). The algorithm was chosen due to its high interpretability and efficiency. Categorical and ordinal (likert-scale) variables were one-hot encoded and continuous variables were transformed using the Yeo-Johnson method. Similarly, PHQ-9 scores were transformed with a Box-Cox method. After omitting the outcome variable, K-Nearest Neighbor imputation was utilized for partially missing data pertaining to 10 non-UI features representing 8% of the total data. This was done to maximize the available data for UI and PHQ-9 scores and enhance model performance and generalizability. For the UI-only model, no imputation was necessary. Feature reduction was conducted using collinearity with a threshold of R2 > 0.9. For hyperparameter optimization and model performance evaluation, nested k-fold cross-validation was utilized with ten inner and five outer folds in tandem with a randomized grid search. Scoring was evaluated with mean absolute error (MAE). MAE denotes the mean difference between the predicted continuous PHQ-9 score and the known value from the dataset. Individual feature impacts were interpreted using the Shapley Explanations Framework which determines the numerical effect of each feature on the model PHQ-9 prediction. Python (Version 3.8.1, with packages LightGBM, Scikit-learn, and NumPy) was used for model fitting, prediction, and evaluation.

From a study cohort of 5,717 patients with incontinence, 74% were female and the mean age (std) was 56.0 (16.5) years. The mean (std) PHQ-9 score of the cohort was 10.7 (6.6). Bothersome and effect on daily activities reflected the subjective aspects of UI while nocturia frequency and ISI are objective characteristics. Four variables were most impactful among the ten most influential features to predict PHQ-9 score: bothersome (KIQ052), effect on daily activities (KIQ050), nocturia (KIQ480), and ISI (Figure 1). Higher severity in each aspect of UI produced an increased PHQ-9 score (Figure 2). However, there tended to be a severity threshold in which the effect on PHQ-9 changed from positive to negative. 

The model containing UI-only features resulted in an MAE (std) of 3.71 (0.11) for outer loop performance on the held-out test set. When UI features were incorporated with variables known to be associated with depression, the resulting MAE (std) decreased to 3.24 (0.09). Other features including decreased age, diagnosis of diabetes, and lower-income were most influential in predicting a higher PHQ-9 score. Furthermore, the same aspects of UI shown such as bothersome, effect on daily activities, nocturia, and ISI demonstrated a similar high impact in prediction in the UI-only model. Interestingly, the effect on daily activities and nocturia were among the ten most influential features and even exceeded overall health, marriage, and smoking status. The effect of UI on daily activities produced an impact in the top five of all demographic and UI features.

PHQ-9 scores are most predicted by subjective UI features than objective features. This supports the hypothesis that the individual experience of UI tends to have a more significant impact on depressive symptoms in comparison to frequency or volume-based characteristics. Subjective UI features, as well as nocturia and ISI, appear to be better predictors than other non-UI features known to be associated with depression in patients with UI. 

Depression amongst UI patients is correlated to worse quality of living and other medical comorbidities. The findings demonstrate that we may be able to predict and provide early intervention to those at-risk for depressive symptoms in primary care and/or urologic clinical care settings. Additionally, high-performing machine learning models provide highly accurate predictions that allow us to potentially prioritize patient care and value in UI cohort.

There remain limitations to this study. NHANES involves cross-sectional responses, hence, no conclusions can be drawn regarding causation. Additionally, PHQ-9 scores when classified according to no depression, mild, moderate, and severe depression are inherently imbalanced. For patients with lower PHQ-9 scores, the model may perform better than those with higher scores. In the future, class imbalance may be addressed with random oversampling to better predict the PHQ-9 scores in patients with moderate-severe to severe depression.

Machine learning models incorporating UI features can accurately predict PHQ-9 scores, a marker of depression and worse quality of life. Subjective UI features may be most influential in the relationship with depression severity, hence, social dynamics, patient well-being, and coping strategies should be emphasized during UI diagnosis and treatment.

Continence 2S2 (2022) 100480
DOI: 10.1016/j.cont.2022.100480