Research On Mortality Prediction Based On High-Dimensional Imbalanced ICU Data

The Intensive Care Units(ICUs)gather the most important resources of the hospital,aiming to provide the comprehensive and reliable treatment for the critical patients.Making mortality predictions for ICU patients can help doctors design diagnostic plans and rationally allocate medical resources,thereby reducing the mortality rate of ICU patients and decreasing the patient's medical expenses.However,the accuracy of the existing mortality prediction model is still not high,and the prediction models have some shortcomings as well.Due to the particularity of ICU,ICU data has the properties of high dimensionality,imbalance and data missing.The traditional machine learning prediction models are mostly designed from the perspective of models combination and are less designed for data characteristics.In addition,there are fewer reports of parameter optimization in the design of the model,which is an important factor affecting the performance of the model.This paper designed series of analysis methods and tools to solve the above issues,focusing on the preprocessing of high-dimensional unbalanced data and the parameter optimization.This paper modified Cost-sensitive Principal Component Analysis(called MCSPCA)to improve the performance of Cost-sensitive Principal Component Analysis(CSPCA)method.In the adjustment of the cost coefficient,different coefficients are used,and the final choice is to use positive coefficients C_i~+=1,negative coefficients C_i~-=N_+/N_-.Through such adjustment of the cost coefficients between the positive and negative samples,we can calculate sample matrix after dimensionality reduction,thus the high-dimensional and unbalanced problems of the samples can be improved in the feature extraction phase;For parameter optimization problem,we designed the Chaos Particle Swarm Optimization(CPSO)algorithm,trying to improve the performance of algorithm from two aspects,namely,particle chaotic sequence initialization and premature convergence judgment processing mechanism,and these two stages are also improved,to increase the dispersion of population particles in the solution space and improve the ability of the population breaking away from premature convergence.In order to obtain the best prediction model,this paper investigated different preprocessing methods,parameter optimization methods and different classifiers,.and finally evaluated their AUC performances in a real benchmark dataset(Physionet Challenge 2012 Dataset).The test results showed that this model improved the performancesofthetraditionalmachinelearningmethods.Theproposed MCSPCA+CPSO+SVM model achieved the best AUC performance value of 0.7718 and the minimum elapsed running time 814s,thereby improving the performance of the ICU prediction mortality model.