Leveraging Spatial Patterns in Precipitation Forecasts Using Deep Learning to Support Regional Water Management

Bias correction of ensemble precipitation forecasts in the improvement of summer streamflow prediction skill

Motivation: Certain studies use short-TE MRS to estimate metabolites such as gamma-aminobutyric acid. The findings of such studies hinge on the correlation of this technique to the gold-standard edited MRS technique. Goal(s): We investigate how the false negative rate and false positive rate of studies vary as a function of correlation between short-TE and MEGA-edited metabolite estimates. Approach: Simulated data was used to numerically estimate the false positive and false negative rates, using parameters obtained from the literature. Results: We show that these studies can have largely inflated false positive and false negative rates as compared to their nominally designed values. Impact: The false negative and false positive rate can be markedly inflated when using short-TE MRS to estimate metabolites such as gamma-aminobutyric acid. This calls for increased caution in interpreting results, and the accommodation of these effects when designing statistical analyses.

Ranaviruses have been identified as the etiologic agent in many amphibian die-offs across the globe. Polymerase chain reaction (PCR) is commonly used to detect ranavirus infection in amphibian hosts, but the test results may vary between tissue samples obtained by lethal and non-lethal procedures. Testing liver samples for infection is a common lethal sampling technique to estimate ranavirus prevalence because the pathogen often targets this organ and the liver is easy to identify and collect. However, tail clips or swabs may be more practicable for ranavirus surveillance programs compared with collecting and euthanizing animals, especially for uncommon species. Using PCR results from liver samples for comparison, we defined false-positive test results as occurrences when a non-lethal technique indicated positive but the liver sample was negative. Similarly, we defined false-negative test results as occurrences when a non-lethal technique was negative but the liver sample was positive. Using these decision rules, we estimated false-negative and false-positive rates for tail clips and swabs. Our study was conducted in a controlled facility using American bullfrog Lithobates catesbeianus tadpoles; false-positive and false-negative rates were estimated after different periods of time following exposure to ranavirus. False-negative and false-positive rates were 20 and 6%, respectively, for tail samples, and 22 and 12%, respectively, for swabs. False-negative rates were constant over time, but false-positive rates decreased with post-exposure duration. Our results suggest that non-lethal sampling techniques can be useful for ranavirus surveillance, although the prevalence of infection may be underestimated when compared to results obtained with liver samples.

Sentiment analysis is a subfield of Natural Language Processing (NLP) which tries to process a text to extract opinions or attitudes towards topics or entities. Recently, the use of deep learning methods for sentiment analysis has received noticeable attention from researchers. Generally, different deep learning methods have shown superb performance in sentiment analysis problem. However, deep learning models are different in nature and have different strengths and limitations. For example, convolutional neural networks are useful for extracting local structures from data, while recurrent models are able to learn order dependence in sequential data. In order to combine the advantages of different deep models, in this paper we have proposed a novel approach for aspect-based sentiment analysis which utilizes deep ensemble learning. In the proposed method, we first build four deep learning models, namely CNN, LSTM, BiLSTM and GRU. Then the outputs of these models are combined using stacking ensemble approach where we have used logistic regression as meta-learner. The results of applying the proposed method on the real datasets show that our method has increased the accuracy of aspect-based prediction by 5% to 20% compared to the basic deep learning methods.

Testing for the presence of a condition may give an incorrect result Sensitivity and specificity of a test depend upon the chosen cut-off value ‘No difference’ usually means ‘these data do not suggest a difference’ The power of a test needs to be known to conclude that there is no difference Power calculations should be considered if the null hypothesis is not rejected Calvin Coolidge was a taciturn president of the United States, nicknamed ‘Silent Cal’. When told that he had died, the acerbic writer Dorothy Parker remarked, ‘How could they tell?’ Telling if something has not happened is a perennial problem in science, particularly so in laboratory experiments. Why should this be? In the case of Cal Coolidge, they probably could tell. But in science, random events make it hard to tell, with certainty. There are also different ways of ‘telling’. Mortality is a good categorical measure, an unequivocal endpoint. When diagnostic tests are used to define categories, we meet the concepts of ‘false positive’ and ‘false negative’. Here, we use the expression ‘false’ to indicate that the conclusion we drew from the test would be incorrect. In the example we use below, we have a test for ‘alive’. If we know the true condition, we can classify test results as true or false. Classifying a dead animal as alive would be a false positive. True and false are words usually applied to a diagnosis or similar categorical events but also crop up in other aspects of statistical inference. Categorical analysis and categorical logic (it's either there or it isn’t) are not the same as the frequentist logic (the likelihood is small) that is used in common statistical analysis, and this leads to a great deal of argument and misunderstanding. Our previous comment that statistics has an impoverished vocabulary rings true: in this instance, the concepts of false positive and false negative conclusions are applied in different ways to different procedures, causing confusion. Considering categories, imagine that a pool that has become overheated in the sun: some of the frogs in it have died. Our observations suggest that the frogs on the surface of the pool are still alive, and the ones at the bottom of the pool are dead. Is this correct? A random sample produces the data in Fig. 1. Not all the frogs on the surface are alive. If we used ‘on the surface’ to indicate being alive, we would falsely attribute life to 10 frogs. Since 107 frogs are dead, this is a false positive rate of 10/107 (which we only know if we have an unequivocal means of determining death). The positive predictive power of the test for life being present, if we were to look at the next frog on the surface, would be 80/90. Unfortunately, expressions such as sensitivity, false positive, and so on can be used in two different senses. Here we assume we already have an exact means to determine death, much in the way that some statistics theory will start by defining a population. Thus we consider that the false positive rate of our test for ‘alive’ is the proportion of dead frogs on the surface, expressed as a proportion of the frogs that really ARE dead. An alternative view would be if we were to consider the proportion of frogs on the surface that are later found to be dead, we would then be considering the outcome of the test. The diagnostic false positive rate would be 10/90. This is a diagnostic probability, and this form is less commonly used. Many diagnostic tests don't rely on categories like this: a continuous variable is often used to distinguish categories. Suppose we measure the blood lactate concentration in frogs that are still alive, and relate these concentrations to subsequent survival, we might obtain the results shown in Fig. 2. Plasma lactate values in frogs after overheating, related to survival The groups are different (P < 0.001, Mann–Whitney U test). The median value for survivors is 1.29, and for non-survivors, 1.86 (mmol l−1). Although the median concentrations are different, some frogs with high lactate concentrations can survive. Is there a particular lactate concentration that will distinguish survival from non-survival? The capacity to predict outcome (sensitivity and specificity) alters as we choose different cut-off values. Using a series of cut-off values we can plot sensitivity and (1 – specificity) for each chosen value (Fig. 3). This is known as a Receiver-Operator Characteristic Curve, or ROC curve, because one of its first applications was to find how operators of radar receivers distinguish ‘real’ echoes from background noise. In this case, we are looking for survivors using a lactate concentration that is less than the cut-off value we have chosen. A Receiver Operator Characteristic curve for the capacity of a lactate concentration below a chosen value to predict survival An effective diagnostic test should have a large area below the curve, which would show a combination of good sensitivity and specificity. Since tests vary, so does opinion about the area below the curve that indicates a good test. It should be substantially greater that 50%, which would be no better than chance alone, and the closer to 100% the better. The large overlap of individual values between the two groups in Fig. 2 suggests that lactate concentration is a poor indicator of survival, as there is no value that combines useful specificity and sensitivity. Any cut-off values we could choose would have too many false positive or false negative conclusions. Despite this, there is a statistically significant difference according to the Mann–Whitney test, showing that a significant difference doesn't mean the measure is the basis of a good diagnostic test. In other forms of statistical inference, an expression such as ‘false positive’ is less appropriate and is often not well defined, hidden in the logic of these tests. For example, the t test does not provide black or white conclusions: statistical inference is restricted to judging how large a part chance could play in what we observe. Frequentist statistical tests such as the t tests are measures of uncertainty. The usual laboratory experiment question ‘do these data show an effect?’ has to be worded ‘are these samples statistically different?’ As we know, the question then is re-worded ‘How likely are these results if the samples came from the same population?’ Let's say our test result is that P= 0.01, which would usually be taken to indicate there is an ‘effect’. This is not so: this P value is the probability that we conclude there is an effect, and we still could be wrong. Although small, a chance still exists that we would be wrong, about 1 in 100 times. This is known as the alpha error, the possibility that we might classify our result as showing an effect, when in fact there is none. In the long run, if we were to repeatedly sample two populations that actually were the same, then results like the ones we found would be found 1 in 100 times. This is sometimes called the ‘false positive’ rate. However, it's based on the premise that there is NO difference and is perhaps better called the ‘false conclusion’ rate. Reasoning for ‘no effect’ is even less certain, and also often misconstrued. If the P value is 0.12, then we judge that the results are quite possible, assuming the data are from the same population. If we did the same experiment again and again, then we might obtain data like this in almost 1 out of 8 experiments. Is this unlikely? We usually choose to decide this is quite likely, and thus do not reject the null hypothesis: if we rejected it, the chance of a false conclusion is too great. This ‘chance of a false conclusion’ is often misleadingly called a ‘false negative rate’. The hypothesis the test uses, which is the NULL hypothesis, is resoundingly negative; there are no positive results lurking here. Positive results come if we are persuaded, by an estimate of probability (the P value), that the null hypothesis is not tenable. The context of the question ‘is there an effect?’ is relevant. The consequence of the answer ‘no effect’ may be unimportant. However, in some cases the consequence of concluding that there is no difference, because a difference is not apparent, could be devastating. How often do we hear the claim ‘there is no evidence that X is toxic?’ to assert that X is not toxic? Immediately, the sceptic will ask how toxic, and how much evidence? The statistician may add another qualification: how much variation is present? A reasonable example (not a perfect analogy) is the radar operator: there is a greater certainty about a blip on the screen if the signal is large (accept the answer only if it's very likely), if you can keep looking at the screen (improve the sample size), and if there is not a lot of background noise (Fig. 4). In each of these examples of signal averaging, there is an identical signal with added random noise. The signal is more easily seen if the noise is reduced (vertical arrow). Signal detection is improved by taking more samples to allow more averaging (horizontal arrow). The final feature of signal detection is how certain we wish to be (from ‘just discernible’ to ‘barn door’) that the event is not random (a concept analogous to the P value). In this example, a signal is always present, but we just can't be sure in all the cases. This is equally so when statistics are applied to the null hypothesis. We have to consider the beta error: the possibility that we might classify our result as showing no effect where in fact there could be a difference. To demonstrate, consider two populations that are known to be different, as shown in Fig. 5. A, random samples from two different populations, low and high, where the population means differ by 0.1 mmol l−1. A t test finds the P values shown. B, a valuable further analysis that should be applied before erroneously assuming ‘no difference’ is to inspect the 90% confidence limits of the difference between the means. This indicates the size of the possible differences. Here, a small sample of 20 per group, and a small true difference between the groups, does not reject the null hypothesis, with a P value of 0.05. In this particular case, there is indeed a true difference in the populations from which these samples have been taken, but the power of the test applied has been inadequate. The power of a test is the chance of finding a defined difference: a power of 90% would mean that there was only a 10% chance of failing to find a difference of the defined magnitude. Thus power is (1 −β). The 10% false negative conclusion rate or beta error, is also known as a type II error. Some call this a ‘false negative’ but this could be misleading since there is never a ‘true’ negative here. The frequentist definition is that if the experiment were repeated over and over, the conclusion would be incorrect occasionally, and the rate is the beta error. For the example we have chosen, the relationships between power, sample size and the difference we seek to detect are shown in Fig. 6. Using the samples in Fig. 4, the power of the test (the chance of finding a specified difference between samples, given that it exists) is shown, where n= 20 and n= 40, for two different P values. The populations that provided the example data in Fig. 4 were different, by 0.1 mmol l−1, indicated by the vertical dotted line. If P is accepted at 0.05 and n= 40, the power of the test is only about 40%. The retrospective use of power calculations has been often criticized, but this is usually because the authors were seeking to reject a null hypothesis (they would have liked to find a difference). If a null hypothesis cannot be rejected, this is not sufficient evidence that there is NO difference. Claiming ‘no difference’ can be a serious error. If we wish to convince ourselves, and others, that there is no difference, we must be sure that the test was capable of rejecting a specifically defined difference, if it were present. If a small difference is considered important, and variation in the population is substantial, then a small sample will be inadequate. Put simply, the study is underpowered. Although there are software packages to calculate power, a simple approach for a simple two group comparison is to plot the 90% confidence limits of the difference between the means (Fig. 5B). This will show the likely range of differences that could be detected. The logical error (concluding that no significant difference is equivalent to the same) is frequently made worse when test results are compared, for example: the effect of substance A causes a significant difference, but substance B does not; thus A has more effect than B. As we explain above, several factors, not explicitly stated, could affect this result. It could well be that B has a greater, but more varied, effect. Because most tests are looking for a positive, many scientists fail to consider negatives adequately. In the old song ‘Accentuate the positive’ the accomplished song writer Johnny Mercer also cautioned to ‘eliminate the negative, don't mess with Mr In-Between’.

COVID-19 is highly infectious and causes acute respiratory disease. Machine learning (ML) and deep learning (DL) models are vital in detecting disease from computerized chest tomography (CT) scans. The DL models outperformed the ML models. For COVID-19 detection from CT scan images, DL models are used as end-to-end models. Thus, the performance of the model is evaluated for the quality of the extracted feature and classification accuracy. There are four contributions included in this work. First, this research is motivated by studying the quality of the extracted feature from the DL by feeding these extracted to an ML model. In other words, we proposed comparing the end-to-end DL model performance against the approach of using DL for feature extraction and ML for the classification of COVID-19 CT scan images. Second, we proposed studying the effect of fusing extracted features from image descriptors, e.g., Scale-Invariant Feature Transform (SIFT), with extracted features from DL models. Third, we proposed a new Convolutional Neural Network (CNN) to be trained from scratch and then compared to the deep transfer learning on the same classification problem. Finally, we studied the performance gap between classic ML models against ensemble learning models. The proposed framework is evaluated using a CT dataset, where the obtained results are evaluated using five different metrics The obtained results revealed that using the proposed CNN model is better than using the well-known DL model for the purpose of feature extraction. Moreover, using a DL model for feature extraction and an ML model for the classification task achieved better results in comparison to using an end-to-end DL model for detecting COVID-19 CT scan images. Of note, the accuracy rate of the former method improved by using ensemble learning models instead of the classic ML models. The proposed method achieved the best accuracy rate of 99.39%.

The Effect of Testing Reliability on Visual Field Sensitivity in Normal Eyes: The Singapore Chinese Eye Study

"Test and Standard Precautions"-Is It Enough to Protect Us From False-Negative Severe Acute Respiratory Syndrome Coronavirus 2 Test Results?

Deep learning algorithms can be used to design systems that report data on patients and deliver warnings to medical applications or electronic health information if there are changes in the patient's health. These systems could be created using deep Learning. This helps verify that patients get the proper effective care at the proper time for each specific patient. A healthcare decision support system was presented using the Internet of Things and deep learning methods. In the proposed system, we examined the capability of integrating deep learning technology into automatic diagnosis and IoT capabilities for faster message exchange over the Internet. We have selected the suitable Neural Network structure (number of best-hidden layers and activation function classes) to construct the e-health system. In addition, the e-health system relied on data from doctors to understand the Neural Network. In the validation method, the total evaluation of the proposed healthcare system for diagnostics provides dependability under various patient conditions. Based on evaluation and simulation findings, a dual hidden layer of feed-forward NN and its neurons store the tanh function more effectively than other NN. To overcome challenges, this study will integrate artificial intelligence with IoT. This study aims to determine the NN's optimal layer counts and activation function variations.

[Departement_IRSTEA]Eaux [TR1_IRSTEA]ARCEAU

This paper compares the ability of deep learning and entropic learning methods to predict the probability of the Niño-3.4 index being above 0.4° (El Niño), below −0.4° (La Niña), or within both of these thresholds (neutral) at lead times of 3 up to 24 months. In particular, the performance, interpretability, and training cost of entropic learning methods, represented by the entropy-optimal scalable probabilistic approximation (eSPA) algorithm, are compared with deep learning methods, represented by a long short-term memory (LSTM) classifier, trained on the same dataset. Using only monthly data derived from a surface-forced ocean model over the period 1958–2018, the problem manifests as a canonical small-data challenge. Relative to the LSTM model, eSPA exhibits equivalent or better performance in terms of area under the receiver operating characteristic curve (AUC) for all lead times with a smaller computational cost by two and three orders of magnitude. Comparisons of AUC with other state-of-the-art deep learning models in the literature show that eSPA appears to also be more accurate than these models across all three classes. Composite images are generated for each of the cluster centroids from each trained eSPA model at each lead time. At shorter lead times, the composite images for the most significant clusters correspond to patterns representing mature or emerging/declining El Niño or La Niña states, while at longer lead times, they correspond to precursor states consisting of extratropical anomalies. Finally, modifications to the baseline dataset are explored, showing that improvements can be made in the parsimony of the trained eSPA model without sacrificing predictive power.

Recent advances in microarray technology have opened new ways for functional annotation of previously uncharacterised genes on a genomic scale. This has been demonstrated by unsupervised clustering of co-expressed genes and, more importantly, by supervised learning algorithms. Using prior knowledge, these algorithms can assign functional annotations based on more complex expression signatures found in existing functional classes. Previously, support vector machines (SVMs) and other machine-learning methods have been applied to a limited number of functional classes for this purpose. Here we present, for the first time, the comprehensive application of supervised neural networks (SNNs) for functional annotation. Our study is novel in that we report systematic results for ~100 classes in the Munich Information Center for Protein Sequences (MIPS) functional catalog. We found that only ~10% of these are learnable (based on the rate of false negatives). A closer analysis reveals that false positives (and negatives) in a machine-learning context are not necessarily "false" in a biological sense. We show that the high degree of interconnections among functional classes confounds the signatures that ought to be learned for a unique class. We term this the "Borges effect" and introduce two new numerical indices for its quantification. Our analysis indicates that classification systems with a lower Borges effect are better suitable for machine learning. Furthermore, we introduce a learning procedure for combining false positives with the original class. We show that in a few iterations this process converges to a gene set that is learnable with considerably low rates of false positives and negatives and contains genes that are biologically related to the original class, allowing for a coarse reconstruction of the interactions between associated biological pathways. We exemplify this methodology using the well-studied tricarboxylic acid cycle.

The revised ICH E14 Question and Answer (R3) document issued in December 2015 enables pharmaceutical companies to use concentration-QTc (C-QTc) modeling as the primary analysis for assessing QTc prolongation risk of new drugs. A new approach by including the time effect into the current C-QTc model is introduced. Through a simulation study, we evaluated performances of different C-QTc modeling with different dependent variables, covariates, and covariance structures. This simulation study shows that C-QTc models with ΔQTc being dependent variable without time effect inflate false negative rate and that fitting C-QTc models with different dependent variables, covariates, and covariance structures impacts the control of false negative and false positive rates. Appropriate C-QTc modeling strategies with good control of false negative rate and false positive rate are recommended.

Due to the complexity of medical images, traditional medical image classification methods have been unable to meet actual application needs. In recent years, the rapid development of deep learning theory has provided a technical approach for solving medical image classification tasks. However, deep learning has the following problems in medical image classification. First, it is impossible to construct a deep learning model hierarchy for medical image properties; second, the network initialization weights of deep learning models are not well optimized. Therefore, this paper starts from the perspective of network optimization and improves the nonlinear modeling ability of the network through optimization methods. A new network weight initialization method is proposed, which alleviates the problem that existing deep learning model initialization is limited by the type of the nonlinear unit adopted and increases the potential of the neural network to handle different visual tasks. Moreover, through an in‐depth study of the multicolumn convolutional neural network framework, this paper finds that the number of features and the convolution kernel size at different levels of the convolutional neural network are different. In contrast, the proposed method can construct different convolutional neural network models that adapt better to the characteristics of the medical images of interest and thus can better train the resulting heterogeneous multicolumn convolutional neural networks. Finally, using the adaptive sliding window fusion mechanism proposed in this paper, both methods jointly complete the classification task of medical images. Based on the above ideas, this paper proposes a medical classification algorithm based on a weight initialization/sliding window fusion for multilevel convolutional neural networks. The methods proposed in this study were applied to breast mass, brain tumor tissue, and medical image database classification experiments. The results show that the proposed method not only achieves a higher average accuracy than that of traditional machine learning and other deep learning methods but also is more stable and more robust.

Food and Agriculture Organization/World Health Organization (FAO/WHO) recommended that IgE cross-reactivity between a transgenic protein and allergen be considered when there is >or= 35% identity over a sliding "window" of 80 amino acids. Our objective was to evaluate the false positive and negative rates observed using the FAO/WHO versus conventional FASTA analyses. Data used as queries against allergen databases and analyzed to assess false positive rates included: 1,102 hypothetical corn ORFs; 907 randomly selected proteins; 89 randomly selected corn proteins; and 97 corn seed proteins. To evaluate false negative rates of both methods: Bet v 1a along with several crossreacting fruit/vegetable allergens and a bean alpha-amylase inhibitor were used as queries. Both methods were also evaluated for their ability to detect a putative nonallergenic test protein containing a sequence derived from Ara h 1. FASTA versions 3.3t0 and 3.4t25 were utilized. Data indicate a conventional FASTA analysis produced fewer false positives and equivalent false negative rates. Conventional FASTA versus sliding window derived E scores were generally more significant. Results suggest a conventional FASTA search provides more relevant identity to the query protein and better reflects the functional similarities between proteins. It is recommended that the conventional FASTA analysis be conducted to compare identities of proteins to allergens.