Big data ordination towards intensive care event count cases using fast computing GLLVMS.

Area of Science:

• Data Mining
• Machine Learning
• Biostatistics

Background:

• Dimension reduction is crucial in data mining and machine learning for multicollinearity removal and improved model interpretation.
• High-dimensional data analysis presents computational challenges, necessitating efficient dimension reduction techniques.

Purpose of the Study:

• To apply high-dimensional ordination using Generalized Linear Latent Variable Models (GLLVMs) to event count data from various intensive care units.
• To evaluate the performance and computational time of GLLVMs with different approximation methods and probability distributions.

Main Methods:

• Utilized Generalized Linear Latent Variable Models (GLLVMs) for high-dimensional ordination of event counts from intensive care units.
• Employed variational approximation and Laplace approximation to measure GLLVM performance and computational time.
• Compared performance across Negative Binomial, Poisson, Gaussian, ZIP, and Tweedie distributions.

Main Results:

• GLLVMs demonstrated superior performance, achieving 98% variance, outperforming other methods.
• The Negative Binomial distribution combined with Variational Approximation yielded the best accuracy, indicated by AIC, AICc, and BIC values.
• The best model identified was GLLVM-VA Negative Binomial (AIC 7144.07) and GLLVM-LA Negative Binomial (AIC 6955.922).

Conclusions:

• GLLVMs provide an effective approach for high-dimensional data analysis in intensive care settings.
• The Negative Binomial distribution with Variational Approximation offers the highest accuracy and best model performance.
• GLLVMs enhance both the interpretability and computational efficiency of complex healthcare datasets.