An NI-PSO-LightGBM Model Based on Adversarial Validation: Mortality Prediction in ICU Diabetic Patients

Xiaodong Liang,Minyi Ke

doi:10.1007/978-981-99-2443-1_41

Xiaodong Liang, Minyi Ke

https://doi.org/10.1007/978-981-99-2443-1_41

Copy DOI

Export

Save

Cite

Abstract
Full-Text
Similar Papers

Abstract

Listen

The number of diabetic patients and mortality rates are increasing year by year, imposing a heavy burden on health economies and families. ICU mortality prediction is crucial for patient care and allocating hospital resources. In the paper, a NI-PSO-LightGBM diabetic inpatient mortality prediction model based on adversarial validation is developed. The model was trained and tested using MIT's GOSSIS dataset provided from MIT, which gives various factors involved in hospitalization of diabetic patients. Based on these factors, patients are predicted to survive or not, and the dataset has 91,713 samples with 84 features. Adversarial validation was used to partition the dataset to ensure an even distribution of the training and test set samples and to avoid the model being much less effective on the test set than on the validation set. Feature extraction was performed on the samples using Null Importance to find the optimal subset of features containing the most information. The particle swarm optimization algorithm(PSO) was introduced to adjust the hyperparameters of the LightGBM model to obtain the final NI-PSO-LightGBM model and compare it with the Bayes-optimized LightGBM model (Bayes-LightGBM). The experimental results show that for the Bayes-LightGBM model, the average AUC = 88.66% for 5-fold cross-validation on the validation set, AUC = 88.88% and ACC = 92.84% on the test set. For the NI-PSO-LightGBM model, the validation set Average AUC = 89.52% for 5-fold cross-validation, AUC = 89.53% and ACC = 93.09% on the test set. The difference between the AUC values of the two models in the validation and test sets is extremely small, indicating that the adversarial validation prevents overfitting, and the NI-PSO-LightGBM proposed in the paper performs better than Bayes-LightGBM in both validation and test sets in terms of AUC and ACC, indicating that the NI-PSO-LightGBM model has better predictive ability.

Full Text