Machine learning algorithms and traditional statistical models for detection of dementia: a population-based study

Early detection of dementia can facilitate timely therapeutic interventions to mitigate the disease progression. We evaluate the performance of machine learning algorithms and traditional logistic regression (LR) model in diagnosing dementia amongst a rural Chinese older population.

In this population-based study, predictors for dementia in LR model were screened using the least absolute shrinkage and selection operator regression. We implemented random forest (RF) and eXtreme Gradient Boosting (XGBoost) models. We used isotonic regression for probability recalibration, the area under the receiver operating characteristic curve (AUC) and calibration plots to assess the model performance, and the decision curve analysis (DCA) to assess clinical usefulness of the models.

Of the 5200 participants (mean age = 71.0 years, 57.1% females), 302 (5.8%) were diagnosed with dementia. The AUC of LR, RF and XGBoost models for detecting dementia were 0.88 [95% confidence interval (CI): 0.85-0.91], 0.88 (0.84-0.90) and 0.95 (0.94-0.98), respectively, in the validation set (n = 2080). Calibration plots and Brier scores showed that in the validation set, the three models had good overall alignment between the actual and predicted risks. The DCA showed better net benefit of the three models than the intervention for all or no intervention.

We developed and evaluated the performance of LR, RF and XGBoost models for dementia detection in a rural older population in China. All the three models showed excellent discrimination and calibration. The XGBoost model exhibited the best performance and clinical utility in dementia detection. These models are based on information readily available in routine clinical practice, which could be easily implemented amongst rural older populations for dementia detection.