A restricted Boltzmann machine based method for efficient processing of large biomedical datasets

Abstract

In the biomedical and biomedicine fields, big data processing and analysis has been widely recognized as a fundamental but challenging task. As missing or lost values in datasets are inevitable for various reasons in biomedical studies and clinical practice, the imputation of missing data is critical for providing appropriate datasets in biomedical studies. Incorrect imputation may affect the accuracy of data analysis and results prediction. Previously, a number of algorithms and tools were used to impute missing data. But most of them focused on datasets with low interaction of variables, or a small number of samples or variables. These problems limit further application of the existing methods in more complicated biomedical studies. Also, risk prediction models are increasingly used in diagnosis and prognosis of diseases, clinical interventions, and so forth. The quality of the missing data imputation affects modeling performance.

In this study, I developed a Restricted Boltzmann Machine (RBM)-based methodology that can use biomedical datasets to impute missing values and predict disease risk. The new RBM algorithm can process continuous and categorical data simultaneously. The RBM approach is more effective than six existing imputation algorithms for imputing missing values in six disease-related datasets. In particular, this method takes much less time to impute datasets containing a large number of samples or variables than five other algorithms. I also built a Deep Belief Network based on the modified RBM algorithm to predict the risk of human disease. Using the same datasets, this model makes better predictions.

The RBM-based method was applied to analyze pregnancy and live birth prediction for embryos after in-vitro fertilization. The results show the efficiency of the algorithm in missing data imputation and risk prediction, and show its potential as an alternative method of embryo selection for transfer in-vitro fertilization. Thus, this study provides guidance for biomedical programs containing complex data structures with a large diversity of data types, patients and other variables.