Red-flagging multimorbidity clusters for Alzheimer's disease risk using explainable machine learning: Evidence from a national emergency department sample

Figures

• 

  Figure 1.. Prevalence of chronic conditions among patients with and without AD. This bar chart illustrates the percentage of selected chronic comorbidities among ED visits, stratified by AD status. The x-axis displays 18 chronic conditions, while the y-axis represents the proportion of comorbid diagnoses (%) within each group. AD patients show higher prevalence of urinary tract infection (UTI), depressive disorder, hypothyroidism, and anemia. In contrast, non-AD patients exhibit higher rates of hyperlipidemia, controlled type 2 diabetes, and back pain. These contrasting comorbidity profiles informed the selection of predictor variables for subsequent regression and machine learning analyses.

• 

  Figure 2.. Odds ratios of chronic conditions associated with AD. This horizontal bar plot displays adjusted odds ratios (ORs) with 95% confidence intervals for 17 chronic conditions based on multivariable logistic regression analysis. An OR greater than 1 indicates increased odds of an AD diagnosis associated with the respective condition, while an OR less than 1 indicates reduced odds. The x-axis is presented on a logarithmic scale to enhance interpretability. Urinary tract infection (UTI; OR = 2.74), depression (OR = 1.93), hypothyroidism (OR = 1.68), and anemia (OR = 1.57) showed the strongest positive associations with AD. In contrast, conditions such as controlled type 2 diabetes and back pain were inversely associated with AD.

• 

  Figure 3.. Receiver operating characteristic (ROC) curves for predicting AD. This figure compares the classification performance of four supervised learning models in predicting AD diagnosis using ROC curves. Each curve shows the trade-off between sensitivity (true positive rate) and 1–specificity (false positive rate) across various thresholds. XGBoost and Random Forest models yielded the highest area under the curve (AUC = 0.783), followed closely by Logistic Regression (AUC = 0.782) and Decision Tree (AUC = 0.778). The diagonal dashed line represents the baseline performance of a random classifier (AUC = 0.5).

• 

  Figure 4.. SHAP summary plot for the XGBoost model predicting AD. This summary plot displays SHAP values quantifying the contribution of each feature to the AD prediction model. Each point represents an individual case; color denotes the magnitude of the feature value (red = high, blue = low), and horizontal position reflects the impact on model output (log-odds of AD). Age was the most influential predictor, followed by urinary tract infection (UTI), depression, hypothyroidism, and anemia. Higher-order interaction terms (e.g., UTI + Depression + Anemia) were also among the top contributors, underscoring the additive risk conveyed by multimorbidity clusters. This plot highlights both the global importance and directional influence of features in the model's decision-making process.