Unbiased sensitivity analysis and pruning techniques in neural networks for surface ozone modelling

Extreme learning machine for soil temperature prediction using only air temperature as input

Soil temperature (TS) is a crucial parameter in many fields, especially agriculture. In developing countries like Algeria, the soil temperatures (TS) and the meteorological data are limited. This study investigates the use of Extreme Learning Machine (ELM) for the accurate prediction of daily ST at three different depths (30 cm, 60 cm, and 100 cm) using a minimal number of climatic inputs. The inputs used in this study include maximum and minimum air temperatures, relative humidity, and day of the year (DOY) as a representative of the temporal component. Five different combinations of inputs were used to develop ELM models and determine the best set of input variables. The ELM models were then compared with traditional methods such as multiple linear regression, artificial neural networks, and adaptive neuro-fuzzy inference system. Based on evaluation metrics such as R, RMSE, and MAPE, the ELM models with air temperatures and DOY as inputs (ELM-M0 and ELM-M3) demonstrated superior performance at all depths when compared to the other techniques. The most accurate predictions were found at a depth of 100 cm using the ELM-M3 model, which employed inputs of minimum and maximum air temperatures and DOY, with R value of 0.98, RMSE of 0.68 °C, and MAPE of 3.4%. The results demonstrate that the inclusion of DOY in the climatic dataset significantly enhances the performance and accuracy of machine learning models for ST prediction. The ELM was found to be a fast, simple, effective, and useful tool for TS prediction.

Predicting medical device failure: a promise to reduce healthcare facilities cost through smart healthcare management

Background

The advancement of biomedical research generates myriad healthcare-relevant data, including medical records and medical device maintenance information. The COVID-19 pandemic significantly affects the global mortality rate, creating an enormous demand for medical devices. As information technology has advanced, the concept of intelligent healthcare has steadily gained prominence. Smart healthcare utilises a new generation of information technologies, such as the Internet of Things (loT), big data, cloud computing, and artificial intelligence, to completely transform the traditional medical system. With the intention of presenting the concept of smart healthcare, a predictive model is proposed to predict medical device failure for intelligent management of healthcare services.

Methods

Present healthcare device management can be improved by proposing a predictive machine learning model that prognosticates the tendency of medical device failures toward smart healthcare. The predictive model is developed based on 8,294 critical medical devices from 44 different types of equipment extracted from 15 healthcare facilities in Malaysia. The model classifies the device into three classes; (i) class 1, where the device is unlikely to fail within the first 3 years of purchase, (ii) class 2, where the device is likely to fail within 3 years from purchase date, and (iii) class 3 where the device is likely to fail more than 3 years after purchase. The goal is to establish a precise maintenance schedule and reduce maintenance and resource costs based on the time to the first failure event. A machine learning and deep learning technique were compared, and the best robust model for smart healthcare was proposed.

Results

This study compares five algorithms in machine learning and three optimizers in deep learning techniques. The best optimized predictive model is based on ensemble classifier and SGDM optimizer, respectively. An ensemble classifier model produces 77.90%, 87.60%, and 75.39% for accuracy, specificity, and precision compared to 70.30%, 83.71%, and 67.15% for deep learning models. The ensemble classifier model improves to 79.50%, 88.36%, and 77.43% for accuracy, specificity, and precision after significant features are identified. The result concludes although machine learning has better accuracy than deep learning, more training time is required, which is 11.49 min instead of 1 min 5 s when deep learning is applied. The model accuracy shall be improved by introducing unstructured data from maintenance notes and is considered the author's future work because dealing with text data is time-consuming. The proposed model has proven to improve the devices' maintenance strategy with a Malaysian Ringgit (MYR) cost reduction of approximately MYR 326,330.88 per year. Therefore, the maintenance cost would drastically decrease if this smart predictive model is included in the healthcare management system.

Water quality modelling using principal component analysis and artificial neural network

The study investigates the latent pollution sources and most significant parameters that cause spatial variation and develops the best input for water quality modelling using principal component analysis (PCA) and artificial neural network (ANN). The dataset, 22 water quality parameters were obtained from Department of Environment Malaysia (DOE). The PCA generated six significant principal component scores (PCs) which explained 65.40 % of the total variance. Parameters for water quality variation are mainlyrelated to mineral components, anthropogenic activities, and natural processes. However, in ANN three input combination models (ANN A, B, and C) were developed to identify the best model that can predict water quality index (WQI) with very high precision. ANN A model appears to have the best prediction capacity with a coefficient of determination (R²) = 0.9999 and root mean square error (RMSE) = 0.0537. These results proved that the PCA and ANN methods can be applied as tools for decision-making and problem-solving for better managing of river quality.

Noninvasive Blood Glucose Concentration Measurement Based on Conservation of Energy Metabolism and Machine Learning

Blood glucose (BG) concentration monitoring is essential for controlling complications arising from diabetes, as well as digital management of the disease. At present, finger-prick glucometers are widely used to measure BG concentrations. In consideration of the challenges of invasive BG concentration measurements involving pain, risk of infection, expense, and inconvenience, we propose a noninvasive BG concentration detection method based on the conservation of energy metabolism. In this study, a multisensor integrated detection probe was designed and manufactured by 3D-printing technology to be worn on the wrist. Two machine-learning algorithms were also applied to establish the regression model for predicting BG concentrations. The results showed that the back-propagation neural network model produced better performance than the multivariate polynomial regression model, with a mean absolute relative difference and correlation coefficient of 5.453% and 0.936, respectively. Here, about 98.413% of the predicted values were within zone A of the Clarke error grid. The above results proved the potential of our method and device for noninvasive glucose concentration detection from the human wrist.

Comparative assessment of modeling deep learning networks for modeling ground-level ozone concentrations of pandemic lock-down period

Covid-19 pandemic lock-down has resulted significant differences in air quality levels all over the world. In contrary to decrease seen in primary pollutant species, many of the countries have experienced elevated ground-level ozone levels in this period. Air pollution forecast gains more importance to achieve air quality management and take measures against the risks under such extra-ordinary conditions. Statistical models are indispensable tools for predicting air pollution levels. Considering the complex photochemical reactions involved in tropospheric ozone formation, modeling this pollutant requires efficient non-linear approaches. In this study, deep learning methods were applied to forecast hourly ozone levels during pandemic lock-down for an industrialized region in Turkey. With this aim, different deep learning methods were tested and efficiencies of the models were compared considering the calculated RMSE, MAE, R 2 and loss values.

A hybrid model for PM₂.₅ forecasting based on ensemble empirical mode decomposition and a general regression neural network

Exposure to high concentrations of fine particulate matter (PM₂.₅) can cause serious health problems because PM₂.₅ contains microscopic solid or liquid droplets that are sufficiently small to be ingested deep into human lungs. Thus, daily prediction of PM₂.₅ levels is notably important for regulatory plans that inform the public and restrict social activities in advance when harmful episodes are foreseen. A hybrid EEMD-GRNN (ensemble empirical mode decomposition-general regression neural network) model based on data preprocessing and analysis is firstly proposed in this paper for one-day-ahead prediction of PM₂.₅ concentrations. The EEMD part is utilized to decompose original PM₂.₅ data into several intrinsic mode functions (IMFs), while the GRNN part is used for the prediction of each IMF. The hybrid EEMD-GRNN model is trained using input variables obtained from principal component regression (PCR) model to remove redundancy. These input variables accurately and succinctly reflect the relationships between PM₂.₅ and both air quality and meteorological data. The model is trained with data from January 1 to November 1, 2013 and is validated with data from November 2 to November 21, 2013 in Xi'an Province, China. The experimental results show that the developed hybrid EEMD-GRNN model outperforms a single GRNN model without EEMD, a multiple linear regression (MLR) model, a PCR model, and a traditional autoregressive integrated moving average (ARIMA) model. The hybrid model with fast and accurate results can be used to develop rapid air quality warning systems.

The identification of the relationship between chemical and electrical parameters of honeys using artificial neural networks

A number of significant scientific studies have confirmed the health benefits of honey. Due to the high price of natural honey, it is a common target for adulteration which reduces its medicinal value. Adulteration detection methods require specific laboratory equipment and are very expensive. The development of measurement techniques enables the measurement of electrical characteristics of strained honey. Honey electrical parameters can possibly be used for its quality assessment. The identification of the relationship between chemical and electrical parameters of honeys and analysis to determine if there are frequency-dependent changes, can help in developing of that group of methods. The aim of this research was to determine how the chemical parameters of certain honeys influence the dielectric loss factor and the permittivity of strained honey measured in various frequencies. Another aim was to determine whether the percentage influence of certain chemical parameters of honeys on electrical characteristics significantly depends on frequency value. The research was based on neural network models and sensitivity analysis. The percentage influence of certain chemical parameters on electrical characteristics significantly depends on frequency value.

Artificial neural network modeling of the water quality index for Kinta River (Malaysia) using water quality variables as predictors

This article describes design and application of feed-forward, fully-connected, three-layer perceptron neural network model for computing the water quality index (WQI)(1) for Kinta River (Malaysia). The modeling efforts showed that the optimal network architecture was 23-34-1 and that the best WQI predictions were associated with the quick propagation (QP) training algorithm; a learning rate of 0.06; and a QP coefficient of 1.75. The WQI predictions of this model had significant, positive, very high correlation (r=0.977, p<0.01) with the measured WQI values, implying that the model predictions explain around 95.4% of the variation in the measured WQI values. The approach presented in this article offers useful and powerful alternative to WQI computation and prediction, especially in the case of WQI calculation methods which involve lengthy computations and use of various sub-index formulae for each value, or range of values, of the constituent water quality variables.

Analysis and fit of surface CO2 concentrations at a rural site

PURPOSE

The impact of CO(2) urban plume in a rural area was investigated by concentrations recorded near surface.

METHODS

CO(2) dry concentrations at three levels near surface were recorded for about 8 months at a rural site. Daily cycles were obtained and directional analysis was made with percentiles. Several functions were used to fit background and plume concentrations and the goodness of fit was evaluated with different statistics, which were also compared.

RESULTS

Daily cycle showed a difference of around 2 ppm during the night between the lowest (1.8 m) and the highest (8.3 m) levels. Weighting functions of the directional analysis revealed the influence of the Valladolid urban plume. Two regions were established, with local factors prevailing below 3 m s(-1) and transport dominating above 6 m s(-1). The best fit was achieved with a quadratic function for the background and a cubic function for the plume due to the lack of symmetry observed. Gamma and Weibull distributions were also successfully used. Some statistics, such as the root mean square error (RMSE), stood out when evaluating the goodness of fit, whilst others were discarded due to their extremely low values and the lack of sensitivity against the functions used. Finally, a comprehensive metric merging several statistics was also tested with slight differences against RMSE.

Evaluation of a multiple regression model for the forecasting of the concentrations of NOx and PM10 in Athens and Helsinki

Forecasting models based on stepwise multiple linear regression (MLR) have been developed for Athens and Helsinki. The predictor variables were the hourly concentrations of pollutants (NO, NO(2), NO(x), CO, O(3), PM(2.5) and PM(10)) and the meteorological variables (ambient temperature, wind speed/direction, and relative humidity) and in case of Helsinki also Monin-Obukhov length and mixing height of the present day. The variables to be forecasted are the maximum hourly concentrations of PM(10) and NO(x), and the daily average PM(10) concentrations of the next day. The meteorological pre-processing model MPP-FMI was used for computing the Monin-Obukhov length and the mixing height. The limitations of such statistical models include the persistence of both the meteorological and air quality situation; the model cannot account for rapid changes (on a temporal scale of hours or less than a day) that are commonly associated, e.g., with meteorological fronts, or episodes of a long-range transport origin. We have selected the input data for the model from one urban background and one urban traffic station both in Athens and Helsinki, in 2005. We have used various statistical evaluation parameters to analyze the performance of the models, and inter-compared the performance of the predictions for both cities. Forecasts from the MLR model were also compared to those from an Artificial Neural Network model (ANN) to investigate, if there are substantial gains that might justify the additional computational effort. The best predictor variables for both cities were the concentrations of NO(x) and PM(10) during the evening hours as well as wind speed, and the Monin-Obukhov length. In Athens, the index of agreement (IA) for NO(x) ranged from 0.77 to 0.84 and from 0.69 to 0.72, in the warm and cold periods of the year. In Helsinki, the corresponding values of IA ranged from 0.32 to 0.82 and from 0.67 to 0.86 for the warm and cold periods. In case of Helsinki the model accuracy was expectedly better on the average, when Monin-Obukhov length and mixing height were included as predictor variables. The models provide better forecasts of the daily average concentration, compared with the maximum hourly concentration for PM(10). The results derived by the ANN model where only slightly better than the ones derived by the MLR methodology. The results therefore suggest that the MLR methodology is a useful and fairly accurate tool for regulatory purposes.

GRAPHICAL REPRESENTATION OF CAUSE-EFFECT RELATIONSHIPS AMONG CHEMICAL PROCESS VARIABLES USING A NEURAL NETWORK APPROACH

The extraction of information from tabular data is not a natural task for human beings, which is worse when dealing with high dimensional systems. On the other hand, graphical representations make the understanding easier by exploring the human capacity of processing visual information. Such representations can be used for many purposes, e.g., complex systems structuring which contributes to a better understanding of it. This paper constructs a cause-effect map relating the influence of each input process variable on the steam generated by a boiler. The real case study is based on the operations of a chemical recovery boiler of a Kraft pulp mill in Brazil. The map is obtained by two steps, namely the identification of a neural predictive model for the steam and a study of sensitivity analysis. The numerical results are then depicted in a graphical format using a cause-effect map. This representation highlights the relative importance of the predictor variables to the steam generation. The results, in agreement with the literature, show the higher contribution of the heat released during the fuel burning, and the lower influence of both the fuel temperature and the operating variables associated with the primary level of injection of the combustion air.

Spatial succession modeling of biological communities: a multi-model approach

Strong spatial correlation may exist in the spatial succession of biological communities, and the spatial succession can be mathematically described. It was confirmed by our study on spatial succession of both plant and arthropod communities along a linear transect of natural grassland. Both auto-correlation and cross-correlation analyses revealed that the succession of plant and arthropod communities exhibited a significant spatial correlation, and the spatial correlation for plant community succession was stronger than arthropod community succession. Theoretically it should be reasonable to infer a site's community composition from the last site in the linear transect. An artificial neural network for state space modeling (ANNSSM) was developed in present study. An algorithm (i.e., Importance Detection Method (IDM)) for determining the relative importance of input variables was proposed. The relative importance for plant families Gramineae, Compositae and Leguminosae, and arthropod orders Homoptera, Diptera and Orthoptera, were detected and analyzed using IDM. ANNSSM performed better than multivariate linear regression and ordinary differential equation, while ordinary differential equation exhibited the worst performance in the simulation and prediction of spatial succession of biological communities. A state transition probability model (STPM) was proposed to simulate the state transition process of biological communities. STPM performed better than multinomial logistic regression in the state transition modeling. We suggested a novel multi-model framework, i.e., the joint use of ANNSSM and STPM, to predict the spatial succession of biological communities. In this framework, ANNSSM and STPM can be separately used to simulate the continuous and discrete dynamics.

Predicting sorption of pharmaceuticals and personal care products onto soil and digested sludge using artificial neural networks

A comprehensive analytical investigation of the sorption behaviour of a large selection of over-the-counter, prescribed pharmaceuticals and illicit drugs to agricultural soils and freeze-dried digested sludges is presented. Batch sorption experiments were carried out to identify which compounds could potentially concentrate in soils as a result of biosolid enrichment. Analysis of aqueous samples was carried out directly using liquid chromatography-tandem mass spectrometry (LC-MS/MS). For solids analysis, combined pressurised liquid extraction and solid phase extraction methods were used prior to LC-MS/MS. Solid-water distribution coefficients (K(d)) were calculated based on slopes of sorption isotherms over a defined concentration range. Molecular descriptors such as log P, pK(a), molar refractivity, aromatic ratio, hydrophilic factor and topological surface area were collected for all solutes and, along with generated K(d) data, were incorporated as a training set within a developed artificial neural network to predict K(d) for all solutes within both sample types. Therefore, this work represents a novel approach using combined and cross-validated analytical and computational techniques to confidently study sorption modes within the environment. The logarithm plots of predicted versus experimentally determined K(d) are presented which showed excellent correlation (R(2) > 0.88), highlighting that artificial neural networks could be used as a predictive tool for this application. To evaluate the developed model, it was used to predict K(d) for meclofenamic acid, mefenamic acid, ibuprofen and furosemide and subsequently compared to experimentally determined values in soil. Ratios of experimental/predicted K(d) values were found to be 1.00, 1.00, 1.75 and 1.65, respectively.

Prediction of ozone concentrations in Oporto city with statistical approaches

The performance of three statistical methods: time-series, multiple linear regression and feedforward artificial neural networks models were compared to predict the daily mean ozone concentrations. The study here reported was based on data from one urban site with traffic influences and one rural background site. The studies were performed for the year 2002 and the respective four trimesters separately. In the multiple linear regression and feedforward artificial neural network models, the concentrations of ozone, the concentrations of its precursors (nitrogen oxides) and some meteorological variables for one and two days before the prediction day were used as predictors. For the application of these models in the validation step, the inputs of ozone concentration for one and two days before were replaced by the ozone concentrations predicted by the models. The results showed that time-series modelling was not profitable. In the development step, similar performances were obtained with multiple linear regression and feedforward artificial neural network. Better performance indexes were achieved with feedforward artificial neural network models in validation step. Concluding, feedforward artificial neural network models were more efficient to predict ozone concentrations.