A Patient Outcome Prediction based on Random Forest

Since the research and development value of electronic health records (EHRs) which contains a large number of patient treatment data is very high and meaningful, EHRs has gained attention by researchers in recent years. EHRs has some characteristics such as temporality, sparsity, complexity, irregularity, noisiness and so on, which bring many challenges to direct study the medical data. Thus, an effective feature extraction, or phenotyping from patient EHRs is a key step before any further applications. In this paper, MIMIC-III intensive care database is selected for the experiments. To predict the patient's death outcome (namely death due to illness or still alive), we make full use of the visit records of patients and propose a prediction method that combines the medical concept representation model Med2Vec with random forest algorithm. Experimental results indicate that the proposed method is robust to parameter variations and noise. Besides, compared with other prediction methods, the performance metrics of the proposed method are very well. Finally, the effect of the Med2Vec model is superior to that obtained by raw data (i.e., no feature learning applied to EHRs data).