Neural networks: further insights into error function, generalized weights and others

Artificial neural network for flood susceptibility mapping in Bangladesh

The objective of the research is to investigate flood susceptibility in the Sylhet division of Bangladesh. Eight influential factors (i.e., elevation, slope, aspect, curvature, TWI, SPI, roughness, and LULC) were applied as inputs to the model. In this work, 1280 samples were taken at different locations based on flood and non-flood characteristics; of these, 75% of the inventory dataset was used for training and 25% for testing. An artificial neural network was applied to develop a flood susceptibility model, and the results were plotted on a map using ArcGIS. According to the finding, 40.98% (i.e., 499433.50 hectors) of the study area is found within the very high-susceptibility zone, and 37.43% (i.e., 456168.76 hectors) are in the highly susceptible zone. Only 6.52% and 15% of the area were found in low and medium flood susceptibility zones, respectively. The results of model validation show that the overall prediction rate is around 89% and the overall model success rate is around 98%. The study's findings assist policymakers and concerned authorities in making flood risk management decisions in order to mitigate the negative impacts.

The effect of using fresh farmyard manure (animal manure) on the severity of Fusarium verticilioides in soil, root, stem, and kernels as well as lodging and borer incidence of maize plants

Fusarium verticillioides, an important maize pathogen, produce fumonisins, causes stalk rot and consequentially reduce crop growth and yield. Therefore, herein we aimed to evaluate the potential use of two farmyard soil organic manures, i.e., fresh (5-6 days old) and stored (5-6 months old) organic manure, to manage F. verticillioides infections as well as borer incidence and lodging in maize plants. After 30, 60, and 90 days of sowing, samples of soil, roots, and stems were collected to isolate F. verticillioides. Moreover, we estimated ear and kernel rot induced by F. verticillioides at the final harvest. Fresh organic manure treatment increased infection rates of F. verticillioides in soil, roots, stem and kernels compared to the control treatment. In contrast, stored organic manure plots treatments decrease F. verticillioides frequency. At 90 days after sowing, stored organic manure suppressed the survival of F. verticillioides, which reduced the F. verticillioides incidence percent. These results were similar to the effect of herbicides-and insecticide-treated plots demonstrated, which show a significant decrease in F. verticillioides incidence rates. Mycological analysis on symptomless kernels revealed a higher % of pathogen infection in opened husks variety (Balady) than closed husks variety (SC10). Compared with stored organic manure, the stem borer incidence and lodging percentage were the highest in fresh organic manure plots. Finally, these results demonstrated that storing organic manure within five to six months as farmyard manure led to high-temperature centigrade within organic manure, thereby destroying spores of F. verticillioides, whereas fresh organic manure did not.

Landslide Susceptibility Mapping of Landslides with Artificial Neural Networks: Multi-Approach Analysis of aBackpropagation Algorithm Applying the Neuralnet Package in Cuenca, Ecuador

Natural hazards generate disasters and huge losses in several aspects, with landslides being one of the natural risks that have caused great impacts worldwide. The aim of this research was to explore a method based on machine learning to evaluate susceptibility to rotational landslides in an area near Cuenca city, Ecuador, which has a high incidence of these phenomena, mainly due to its environmental conditions, and in which, however, such studies are scarce. The implemented method consisted of an artificial neural network multilayer perceptron (ANN MLP), generated with the neuralnet R package, with which, by means of different backpropagation algorithms (RPROP+, RPROP−, SLR, SAG, and Backprop), five landslide susceptibility maps (LSMs) were generated for the study area. A landslide inventory updated to 2019 and 10 conditioning factors, mainly topographical, geological, land cover, and hydrological, were considered. The results obtained, which were validated through the AUC-ROC value and statistical parameters of precision, recall, accuracy, and F-Score, showed a good degree of adjustment and an acceptable predictive capacity. The resulting maps showed that the area has mostly sectors of moderate, high, and very high susceptibility, whose landslide occurrence percentages vary between approximately 63% and 80%. In this research, different variants of the backpropagation algorithm were implemented to verify which one gave the best results. With the implementation of additional methodologies and correct zoning, future analyses could be developed, contributing to adequate territorial planning and better disaster risk management in the area.

Application of GIS and Machine Learning to Predict Flood Areas in Nigeria

Floods are one of the most devastating forces in nature. Several approaches for identifying flood-prone locations have been developed to reduce the overall harmful impacts on humans and the environment. However, due to the increased frequency of flooding and related disasters, coupled with the continuous changes in natural and social-economic conditions, it has become vital to predict areas with the highest probability of flooding to ensure effective measures to mitigate impending disasters. This study predicted the flood susceptible areas in Nigeria based on historical flood records from 1985~2020 and various conditioning factors. To evaluate the link between flood incidence and the fifteen (15) explanatory variables, which include climatic, topographic, land use and proximity information, the artificial neural network (ANN) and logistic regression (LR) models were trained and tested to develop a flood susceptibility map. The receiver operating characteristic curve (ROC) and area under the curve (AUC) were used to evaluate both model accuracies. The results show that both techniques can model and predict flood-prone areas. However, the ANN model produced a higher performance and prediction rate than the LR model, 76.4% and 62.5%, respectively. In addition, both models highlighted that those areas with the highest susceptibility to flood are the low-lying regions in the southern extremities and around water areas. From the study, we can establish that machine learning techniques can effectively map and predict flood-prone areas and serve as a tool for developing flood mitigation policies and plans.

Analytics with artificial intelligence to advance the treatment of acute respiratory distress syndrome

Artificial intelligence (AI) has found its way into clinical studies in the era of big data. Acute respiratory distress syndrome (ARDS) or acute lung injury (ALI) is a clinical syndrome that encompasses a heterogeneous population. Management of such heterogeneous patient population is a big challenge for clinicians. With accumulating ALI datasets being publicly available, more knowledge could be discovered with sophisticated analytics. We reviewed literatures with big data analytics to understand the role of AI for improving the caring of patients with ALI/ARDS. Many studies have utilized the electronic medical records (EMR) data for the identification and prognostication of ARDS patients. As increasing number of ARDS clinical trials data is open to public, secondary analysis on these combined datasets provide a powerful way of finding solution to clinical questions with a new perspective. AI techniques such as Classification and Regression Tree (CART) and artificial neural networks (ANN) have also been successfully used in the investigation of ARDS problems. Individualized treatment of ARDS could be implemented with a support from AI as we are now able to classify ARDS into many subphenotypes by unsupervised machine learning algorithms. Interestingly, these subphenotypes show different responses to a certain intervention. However, current analytics involving ARDS have not fully incorporated information from omics such as transcriptome, proteomics, daily activities and environmental conditions. AI technology is assisting us to interpret complex data of ARDS patients and enable us to further improve the management of ARDS patients in future with individual treatment plans.

Identification of risk factors for mortality associated with COVID-19

OBJECTIVES

Coronavirus Disease 2019 (COVID-19) has become a pandemic outbreak. Risk stratification at hospital admission is of vital importance for medical decision making and resource allocation. There is no sophisticated tool for this purpose. This study aimed to develop neural network models with predictors selected by genetic algorithms (GA).

METHODS

This study was conducted in Wuhan Third Hospital from January 2020 to March 2020. Predictors were collected on day 1 of hospital admission. The primary outcome was the vital status at hospital discharge. Predictors were selected by using GA, and neural network models were built with the cross-validation method. The final neural network models were compared with conventional logistic regression models.

RESULTS

A total of 246 patients with COVID-19 were included for analysis. The mortality rate was 17.1% (42/246). Non-survivors were significantly older (median (IQR): 69 (57, 77) vs. 55 (41, 63) years; p < 0.001), had higher high-sensitive troponin I (0.03 (0, 0.06) vs. 0 (0, 0.01) ng/L; p < 0.001), C-reactive protein (85.75 (57.39, 164.65) vs. 23.49 (10.1, 53.59) mg/L; p < 0.001), D-dimer (0.99 (0.44, 2.96) vs. 0.52 (0.26, 0.96) mg/L; p < 0.001), and α-hydroxybutyrate dehydrogenase (306.5 (268.75, 377.25) vs. 194.5 (160.75, 247.5); p < 0.001) and a lower level of lymphocyte count (0.74 (0.41, 0.96) vs. 0.98 (0.77, 1.26) × 109/L; p < 0.001) than survivors. The GA identified a 9-variable (NNet1) and a 32-variable model (NNet2). The NNet1 model was parsimonious with a cost on accuracy; the NNet2 model had the maximum accuracy. NNet1 (AUC: 0.806; 95% CI [0.693-0.919]) and NNet2 (AUC: 0.922; 95% CI [0.859-0.985]) outperformed the linear regression models.

CONCLUSIONS

Our study included a cohort of COVID-19 patients. Several risk factors were identified considering both clinical and statistical significance. We further developed two neural network models, with the variables selected by using GA. The model performs much better than the conventional generalized linear models.

Predicting in-hospital mortality of patients with acute kidney injury in the ICU using random forest model

OBJECTIVES

We aimed to construct a mortality prediction model using the random forest (RF) algorithm for acute kidney injury (AKI) patients in the intensive care unit (ICU), and compared its performance with that of two other machine learning models and the customized simplified acute physiology score (SAPS) II model.

METHODS

We used medical information mart for intensive care (MIMIC) III database for model development and performance comparison. The RF model uses the same predictor variable set as that of the SAPS II model. We also developed three other models and compared the RF model with the other three models in prediction performance. Three other models include support vector machine (SVM) model, artificial neural network (ANN) model and customized SAPS II model. In model comparison, the prediction performance of each model was assessed by the Brier score, the area under the receiver operating characteristic curve (AUROC), accuracy and F₁ score.

RESULTS

The final cohort consisted of 19044 patients with AKI in the ICU. The observed in-hospital mortality of AKI patients is 13.6% in the final cohort. The results of model performance comparison show that the Brier score of the RF model is 0.085 (95%CI: 0.084-0.086) and AUROC of the RF model is 0.866 (95%CI: 0.862-0.870). The accuracy of the RF model is 0.728 (95%CI: 0.715-0.741). The F₁ score of the RF model is 0.459 (95%CI: 0.449-0.470). The calibration plots show that the RF model slightly overestimates mortality in patients with low risk of death and underestimates mortality in patients with high risk of death.

CONCLUSION

There is great potential for the RF model in mortality prediction for AKI patients in ICU. The RF model may be helpful to aid ICU clinicians to make timely clinical intervention decisions for AKI patients, which is critical to help reduce the in-hospital mortality of AKI patients. A prospective study is necessary to evaluate the clinical utility of the RF model.