Missing Value Imputation of Time-Series Air-Quality Data via Deep Neural Networks

Mortality prediction using medical time series on TBI patients

BACKGROUND AND OBJECTIVE

Traumatic Brain Injury (TBI) is one of the leading causes of injury-related mortality in the world, with severe cases reaching mortality rates of 30-40%. It is highly heterogeneous both in causes and consequences making more complex the medical interpretation and prognosis. Gathering clinical, demographic, and laboratory data to perform a prognosis requires time and skill in several clinical specialties. Artificial intelligence (AI) methods can take advantage of existing data by performing helpful predictions and guiding physicians toward a better prognosis and, consequently, better healthcare. The objective of this work was to develop learning models and evaluate their capability of predicting the mortality of TBI. The predictive model would allow the early assessment of the more serious cases and scarce medical resources can be pointed toward the patients who need them most.

METHODS

Long Short Term Memory (LSTM) and Transformer architectures were tested and compared in performance, coupled with data imbalance, missing data, and feature selection strategies. From the Medical Information Mart for Intensive Care III (MIMIC-III) dataset, a cohort of TBI patients was selected and an analysis of the first 48 hours of multiple time series sequential variables was done to predict hospital mortality.

RESULTS

The best performance was obtained with the Transformer architecture, achieving an AUC of 0.907 with the larger group of features and trained with class proportion class weights and binary cross entropy loss.

CONCLUSIONS

Using the time series sequential data, LSTM and Transformers proved to be both viable options for predicting TBI hospital mortality in 48 hours after admission. Overall, using sequential deep learning models with time series data to predict TBI mortality is viable and can be used as a helpful indicator of the well-being of patients.

A Hybrid Missing Data Imputation Method for Batch Process Monitoring Dataset

Batch process monitoring datasets usually contain missing data, which decreases the performance of data-driven modeling for fault identification and optimal control. Many methods have been proposed to impute missing data; however, they do not fulfill the need for data quality, especially in sensor datasets with different types of missing data. We propose a hybrid missing data imputation method for batch process monitoring datasets with multi-type missing data. In this method, the missing data is first classified into five categories based on the continuous missing duration and the number of variables missing simultaneously. Then, different categories of missing data are step-by-step imputed considering their unique characteristics. A combination of three single-dimensional interpolation models is employed to impute transient isolated missing values. An iterative imputation based on a multivariate regression model is designed for imputing long-term missing variables, and a combination model based on single-dimensional interpolation and multivariate regression is proposed for imputing short-term missing variables. The Long Short-Term Memory (LSTM) model is utilized to impute both short-term and long-term missing samples. Finally, a series of experiments for different categories of missing data were conducted based on a real-world batch process monitoring dataset. The results demonstrate that the proposed method achieves higher imputation accuracy than other comparative methods.

A Method of Pruning and Random Replacing of Known Values for Comparing Missing Data Imputation Models for Incomplete Air Quality Time Series

The data obtained from air quality monitoring stations, which are used to carry out studies using data mining techniques, present the problem of missing values. This paper describes a research work on missing data imputation. Among the most common methods, the method that best imputes values to the available data set is analysed. It uses an algorithm that randomly replaces all known values in a dataset once with imputed values and compares them with the actual known values, forming several subsets. Data from seven stations in the Silesian region (Poland) were analyzed for hourly concentrations of four pollutants: nitrogen dioxide (NO2), nitrogen oxides (NOx), particles of 10 μm or less (PM10) and sulphur dioxide (SO2) for five years. Imputations were performed using linear imputation (LI), predictive mean matching (PMM), random forest (RF), k-nearest neighbours (k-NN) and imputation by Kalman smoothing on structural time series (Kalman) methods and performance evaluations were performed. Once the comparison method was validated, it was determine that, in general, Kalman structural smoothing and the linear imputation methods best fitted the imputed values to the data pattern. It was observed that each imputation method behaves in an analogous way for the different stations The variables with the best results are NO2 and SO2. The UMI method is the worst imputer for missing values in the data sets.