Machine Learning Approach Predicts Probability of Time to Stage-Specific Conversion of Alzheimer's Disease

Abstract

Background: The progression of Alzheimer's disease (AD) varies in different patients at different stages, which makes predicting the time of disease conversions challenging.

Objective: We established an algorithm by leveraging machine learning techniques to predict the probability of the conversion time to next stage for different subjects during a given period.

Methods: Firstly, we used Kaplan-Meier (KM) estimation to get the transition curves of different AD stages, and calculated Log-rank statistics to test whether the progression rate between different stages was identical. This quantitatively confirmed the progression rates known in the literature. Then, we developed an approach based on deep learning model, DeepSurv, to predict the probabilities of time-to-conversion. Finally, to help interpret the deep learning model in our approach, we identified important variables contributing the most to the DeepSurv prediction, whose significance were validated with the analysis of variance (ANOVA).

Results: Our machine learning approach predicted the time to conversion with a high accuracy. For each of the different stages, the concordance index (CI) of our approach was at least 86%, and the integrated Brier score (IBS) was less than 0.1. To facilitate interpretability of the prediction results, our approach identified the top 10 variables for each disease conversion scenario, which were clinicopathologically meaningful, and most of them were also statistically significant.

Conclusion: Our study has the potential to provide individualized prediction for future time course of AD conversions years before their actual occurrence, thus facilitating personalized prevention and intervention strategies to slow down the progression of AD.

Keywords: Analysis of variance; dementia; machine learning; survival analysis.

Fig. 1.

Schematic illustration of our workflow for deep learning-based stage-stratified prediction of probabilities of time to AD conversion.

Fig. 2.

Statistics of subjects for different transitions.

Fig. 3.

The comparison of KM estimations for different scenarios. The dotted line is a trend curve fitted by a quadratic polynomial. The horizontal axis denotes the time in days, and the vertical axis denotes the probability of staying in the current stage.

Fig. 4.

Visualization of the predictions for 5 randomly chosen patients in different conversion scenarios: (a) $\{0,0.5\}⇝\left\{1\right\}$; (b) $\left\{1\right\}⇝\left\{2\right\}$; (c) $\left\{2\right\}⇝\left\{3\right\}$; (d) $\left\{1\right\}⇝\{2,3\}$; (e) $\left\{1\right\}⇝\left\{3\right\}$; (f) $\{0,0.5,1\}⇝\left\{2\right\}$. The horizontal axis denotes the time in days, with the origin being the onset of the current stage; the vertical axis denotes the probability of staying in the current stage. The vertical dashed line denotes the actual time when the transition event occurred.