Risk Prediction of One-Year Mortality in Patients with Cardiac Arrhythmias Using Random Survival Forest.

Miao F, Cai YP, Zhang YX, Li Y, Zhang YT - Comput Math Methods Med (2015)

Related In: Results  -  Collection

Show All Figures
getmorefigures.php?uid=PMC4562335&req=5

fig4: Ensemble mortality against given continuous variables. Mortality is presented in terms of total death number. Points colored with blue correspond to events, while black ones correspond to censored observations. log∗ indicates log of index.
Mentions: Figure 4 shows the marginal effect of a given continuous variable on the 1-year mortality from RSF analysis. From the figure we can see that the association between the continuous variables like BMI, BR, AST, SCR, and NA and 1-year mortality is nonlinear. Figure 5 displays how the RSF model shows interaction among the 3 most important continuous variables including BUN, AST, and BMI and 1-year predicted mortality. Patients with the highest BUN and AST have the largest mortality and most have low BMI. One-year predicted mortality increased sharply with the elevation of BUN, with about 15% for those with a BUN below 20 mg/dL and nearly 50% for those with a BUN above 58 mg/dL. The mortality rate depends more on AST than BMI for patients with lower BUN, while more on BMI for those with higher BUN. The above mentioned interactions and nonlinear relationships were not prespecified by the analyst but identified by the forest. That is why the RSF has a better discrimination ability than CPH with the error rate (1-c-statistic) of 19% (the details about estimated error rate according to different grown trees can be seen from Figure 6).

Bottom Line: The simplified risk model also achieved a good accuracy of 0.799.Both results outperformed traditional CPH (which achieved a c-statistic of 0.733 for the comprehensive model and 0.718 for the simplified model).As a result, RSF based model which took nonlinearity into account significantly outperformed traditional Cox proportional hazard model and has great potential to be a more effective approach for survival analysis.

View Article: PubMed Central - PubMed

Affiliation: Key Laboratory for Health Informatics of the Chinese Academy of Sciences (HICAS), Shenzhen Institutes of Advanced Technology, Shenzhen 518055, China ; Shenzhen College of Advanced Technology, University of Chinese Academy of Sciences, Beijing 100049, China.

ABSTRACT
Existing models for predicting mortality based on traditional Cox proportional hazard approach (CPH) often have low prediction accuracy. This paper aims to develop a clinical risk model with good accuracy for predicting 1-year mortality in cardiac arrhythmias patients using random survival forest (RSF), a robust approach for survival analysis. 10,488 cardiac arrhythmias patients available in the public MIMIC II clinical database were investigated, with 3,452 deaths occurring within 1-year followups. Forty risk factors including demographics and clinical and laboratory information and antiarrhythmic agents were analyzed as potential predictors of all-cause mortality. RSF was adopted to build a comprehensive survival model and a simplified risk model composed of 14 top risk factors. The built comprehensive model achieved a prediction accuracy of 0.81 measured by c-statistic with 10-fold cross validation. The simplified risk model also achieved a good accuracy of 0.799. Both results outperformed traditional CPH (which achieved a c-statistic of 0.733 for the comprehensive model and 0.718 for the simplified model). Moreover, various factors are observed to have nonlinear impact on cardiac arrhythmias prognosis. As a result, RSF based model which took nonlinearity into account significantly outperformed traditional Cox proportional hazard model and has great potential to be a more effective approach for survival analysis.

No MeSH data available.

Related in: MedlinePlus