What makes a good prediction? Feature importance and beginning to open the black box of machine learning in genetics

Genetic data have become increasingly complex within the past decade, leading researchers to pursue increasingly complex questions, such as those involving epistatic interactions and protein prediction. Traditional methods are ill-suited to answer these questions, but machine learning (ML) techniques offer an alternative solution. ML algorithms are commonly used in genetics to predict or classify subjects, but some methods evaluate which features (variables) are responsible for creating a good prediction; this is called feature importance. This is critical in genetics, as researchers are often interested in which features (e.g., SNP genotype or environmental exposure) are responsible for a good prediction. This allows for the deeper analysis beyond simple prediction, including the determination of risk factors associated with a given phenotype. Feature importance further permits the researcher to peer inside the black box of many ML algorithms to see how they work and which features are critical in informing a good prediction. This review focuses on ML methods that provide feature importance metrics for the analysis of genetic data. Five major categories of ML algorithms: k nearest neighbors, artificial neural networks, deep learning, support vector machines, and random forests are described. The review ends with a discussion of how to choose the best machine for a data set. This review will be particularly useful for genetic researchers looking to use ML methods to answer questions beyond basic prediction and classification.

Using Textual and Economic Features to Predict the RMB Exchange Rate

This research proposes an integrated framework for the use of textual and economic features to predict the exchange rate of the TWD (Taiwan dollar) against the RMB (Chinese Renminbi). The exchange rate is affected by the current economic situation and expectations for the future economic climate. Exchange rate forecasting studies focus mainly on overall economic indices and the actual exchange rate, but overlook the influence of news. This research considers both textual and economic factors and builds three basic prediction models, i.e. multiple linear regression (MLR), support vector regression (SVR), and Gaussian process regression (GPR) for the prediction of the RMB exchange rate. In addition to the three basic prediction models, this research uses ensemble learning and feature selection techniques to improve prediction performance. Our experiments demonstrate that textual features also play an important role in predicting the RMB exchange rate. The SVR model is shown to outperform the other models and the MLR model is shown to perform worst. The ensemble of three basic models performs better than its individual counterparts. Finally, the models which use feature selection techniques demonstrate improved results in general, and different feature selection techniques are shown to be more suitable for different prediction models. JEL classification numbers: D80, F31, F47. Keywords: Exchange rate prediction, Text mining, Ensemble learning, Time series forecasting.

Dominance of γ-γ electron-positron pair creation in a plasma driven by high-intensity lasers

Creation of electrons and positrons from light alone is a basic prediction of quantum electrodynamics, but yet to be observed. Our simulations show that the required conditions are achievable using a high-intensity two-beam laser facility and an advanced target design. Dual laser irradiation of a structured target produces high-density γ rays that then create > 108 positrons at intensities of 2 × 1022 Wcm−2. The unique feature of this setup is that the pair creation is primarily driven by the linear Breit-Wheeler process (γγ → e+e−), which dominates over the nonlinear Breit-Wheeler and Bethe-Heitler processes. The favorable scaling with laser intensity of the linear process prompts reconsideration of its neglect in simulation studies and also permits positron jet formation at experimentally feasible intensities. Simulations show that the positrons, confined by a quasistatic plasma magnetic field, may be accelerated by the lasers to energies >200 MeV.

P140 Predicting Non-Responsiveness to Iron Therapy in Anaemic Children with Inflammatory Bowel Disease

Background Anaemia negatively impacts physical fitness and quality-of-life, and therapy is aimed at normalization of haemoglobin (Hb) levels. In the majority of patients with inflammatory bowel disease (IBD) anaemia has a mixed aetiology resulting from chronic inflammation and iron deficiency. A considerable proportion of patients will fail to respond to iron therapy, but it has been difficult to identify non-responders at baseline with currently used iron indicators. Hepcidin is a peptide hormone involved in iron homeostasis. Upregulation leads to decreased iron availability, whereas downregulation facilitates iron absorption and release from macrophages to allow for erythropoiesis. We evaluated commonly used iron indicators (ferritin and transferrin saturation [TSAT]) and emerging biomarkers (soluble transferrin receptor [sTfR] and hepcidin levels) at baseline to predict non-responsiveness to iron therapy in anaemic children with IBD. Methods We performed a prospective multi-centre cohort study among patients with IBD and anaemia (defined as Hb > 2 standard deviations (SD) below the reference mean according to WHO cut-offs). We assessed iron indicators, sTfR, and hepcidin at baseline and again one month after the initiation of oral or intravenous iron therapy. Therapy was given according to international guidelines. Primary outcome was based on the change of Hb z-score (one month after treatment vs baseline) divided by baseline SD, where non-responsiveness was defined as a standardised change score of less than 1. Hepcidin was expressed as z-score to allow correction for age and gender. Baseline data of ferritin and TSAT were used to construct a basic logistic regression model. Hepcidin and/or sTfR were then added to the basic prediction model (models [M] 1–3). Optimal sensitivity and specificity were identified using the Youden’s J Index. Results Of 40 anaemic IBD patients (mean age 12.8 years; mean Hb z-score -3.1 SD), sixteen (40%) were non-responsive to iron therapy one month after initiation. The basic prediction model yielded an area under the curve (AUC) of 0.69. Figure 1 shows that adding sTfR, hepcidin or both to the model increased the AUC to 0.78 (M1), 0.82 (M2), and 0.90 (M3), respectively. For model 3, sensitivity and specificity at the optimal cut-off value were 94% and 71%, respectively. Figure 1: Receiver operating characteristic (ROC) curves representing the accuracy of iron therapy non-responsiveness. AUC, area under the curve; CI, confidence interval Conclusion Based on prediction quality of the models, triaging with a strategy that involves baseline ferritin, TSAT, sTfR, and hepcidin is preferred to assess non-responsiveness to iron therapy in anaemic children with IBD.

The Role of Maternal Weight in the Hierarchy of Macrosomia Predictors; Overall Effect of Analysis of Three Prediction Indicators

So far it has not been established which maternal features play the most important role in newborn macrosomia. The aim of this study is to provide assessment of a hierarchy of twenty six (26) maternal characteristics in macrosomia prediction. A Polish prospective cohort of women with singleton pregnancy (N = 912) which was recruited in the years 2015−2016 has been studied. Two analyses were performed: for probability of macrosomia >4000 g (n = 97) (vs. 755 newborns 2500−4000 g); and for birthweight >90th percentile (n = 99) (vs. 741 newborns 10−90th percentile). A multiple logistic regression was used (with 95% confidence intervals (CI)). A hierarchy of significance of potential predictors was established after summing up of three prediction indicators (NRI, IDI and AUC) calculated for the basic prediction model (maternal age + parity) extended with one (test) predictor. ‘Net reclassification improvement’ (NRI) focuses on the reclassification table describing the number of women in whom an upward or downward shift in the disease probability value occurred after a new factor had been added, including the results for healthy and ill women. ‘Integrated discrimination improvement’ (IDI) shows the difference between the value of mean change in predicted probability between the group of ill and healthy women when a new factor is added to the model. The area under curve (AUC) is a commonly used indicator. Results. The macrosomia risk was the highest for prior macrosomia (AOR = 7.53, 95%CI: 3.15−18.00), p < 0.001). A few maternal characteristics were associated with more than three times higher macrosomia odds ratios, e.g., maternal obesity and gestational age ≥38 weeks. A different hierarchy was shown by the prediction study. Compared to the basic prediction model (AUC = 0.564 (0.501−0.627), p = 0.04), AUC increased most when pre-pregnancy weight (kg) was added to the base model (AUC = 0.706 (0.649−0.764), p < 0.001). The values of IDI and NRI were also the highest for the model with maternal weight (IDI = 0.061 (0.039−0.083), p < 0.001), and NRI = 0.538 (0.33−0.746), p < 0.001). Adding another factor to the base model was connected with significantly weaker prediction, e.g., for gestational age ≥ 38 weeks (AUC = 0.602 (0.543−0.662), p = 0.001), IDI = 0.009 (0.004; 0.013), p < 0.001), and NRI = 0.155 (0.073; 0.237), p < 0.001). After summing up the effects of NRI, IDI and AUC, the probability of macrosomia was most strongly improved (in order) by: pre-pregnancy weight, body mass index (BMI), excessive gestational weight gain (GWG) and BMI ≥ 25 kg/m2. Maternal height, prior macrosomia, fetal sex-son, and gestational diabetes mellitus (GDM) occupied an intermediate place in the hierarchy. The main conclusions: newer prediction indicators showed that (among 26 features) excessive pre-pregnancy weight/BMI and excessive GWG played a much more important role in macrosomia prediction than other maternal characteristics. These indicators more strongly highlighted the differences between predictors than the results of commonly used odds ratios.

EMI Shielding Properties of Polymer Blend with the Inclusion of Graphene Nano Platelets

Polymer nano composites based on poly vinyl chloride matrix were fabricated using polyaniline (PANI) and graphene nano platelets (GNP) as electrically conductive nano filler for the application of electromagnetic interference (EMI) shielding. DC conductivity was first evaluated by using cyclic voltammetry and found an increasing trend of electrical conductivity as PANI and GNP was added in PVC matrix that confirms the formation of electrically conductive network structure. Dielectric properties like dielectric constant, dielectric loss and AC conductivity were evaluated in frequency range of 100 Hz to 3 MHz that gives basic prediction for EMI shielding effectiveness. Vector Network Analyzer (VNA) was used to assess EMI shielding properties using coaxial cable method in 11GHz to 20GHz range and it was found that a maximum of 29 dB shielding was archived with the incorporation of 15 wt% of PANI in PVC. This value increased to 56 dB as 5 wt% GNP added in PVC/PANI 15 wt% blend. Interaction of matrix with filler, nature of filler and dispersion of filler in matrix are the key parameters for achieving high shielding effectiveness.

A Distributionally Robust Power Dispatch Model for Active Distribution Network

The uncertainty of distribution network operation is increasing with the integration of large-scale renewable distributed generation (DG) units. To reduce the conservativeness of traditional robust optimization (RO) solutions, a data-driven robust optimal approach, which incorporates the superiority of both stochastic and robust approaches, is employed to solve the dispatch model in this paper. Firstly, a deterministic optimal dispatching model is established with the minimum total operation cost of distribution network; secondly, a two-stage distributed robust dispatching model is constructed based on the historical data of renewable-generators output available. The first stage of the model aims at finding optimal values under the basic prediction scenario. In the second stage, the uncertain probability distribution confidence sets with norm-1 and norm-∞ constraints are integrated to find the optimal solution under the worst probability distribution. The model is solved by column-and-constraint generation (CCG) algorithm. Numerical simulation on the IEEE 33-bus system has been performed. Comparisons with the traditional stochastic and robust approaches demonstrate the effectiveness of the proposal.

Passive Underwater Target Tracking: Conditionally Minimax Nonlinear Filtering with Bearing-Doppler Observations

The paper presents an application of the Conditionally-Minimax Nonlinear Filtering (CMNF) algorithm to the online estimation of underwater vehicle movement given a combination of sonar and Doppler discrete-time noisy sensor observations. The proposed filter postulates recurrent “prediction–correction” form with some predefined basic prediction and correction terms, and then they are optimally fused. The CMNF estimates have the following advantageous features. First, the obtained estimates are unbiased. Second, the theoretical covariance matrix of CMNF errors meets the real values. Third, the CMNF algorithm gives a possibility to choose the preliminary observation transform, basic prediction, and correction functions in any specific case of the observation system to improve the estimate accuracy significantly. All the features of conditionally-minimax estimates are demonstrated by the regression example of random position estimate given the noisy bearing observations. The contribution of the paper is the numerical study of the CMNF algorithm applied to the underwater target tracking given bearing-only and bearing-Doppler observations.

Basic prediction methodology for covid-19: estimation and sensitivity considerations

SummaryThe purpose of the present paper is to present simple estimation and prediction methods for basic quantities in an emerging epidemic like the ongoing covid-10 pandemic. The simple methods have the advantage that relations between basic quantities become more transparent, thus shedding light to which quantities have biggest impact on predictions, with the additional conclusion that uncertainties in these quantities carry over to high uncertainty also in predictions.A simple non-parametric prediction method for future cumulative case fatalities, as well as future cumulative incidence of infections (assuming a given infection fatality risk f), is presented. The method uses cumulative reported case fatalities up to present time as input data. It is also described how the introduction of preventive measures of a given magnitude ρ will affect the two incidence predictions, using basic theory of epidemic models. This methodology is then reversed, thus enabling estimation of the preventive magnitude ρ, and of the resulting effective reproduction number RE. However, the effects of preventive measures only start affecting case fatalities some 3-4 weeks later, so estimates are only available after this time has elapsed. The methodology is applicable in the early stage of an outbreak, before, say, 10% of the community have been infected.Beside giving simple estimation and prediction tools for an ongoing epidemic, another important conclusion lies in the observation that the two quantities f (infection fatality risk) and ρ (the magnitude of preventive measures) have very big impact on predictions. Further, both of these quantities currently have very high uncertainty: current estimates of f lie in the range 0.2% up to 2% ([9], [7]), and the overall effect of several combined preventive measures is clearly very uncertain.The two main findings from the paper are hence that, a) any prediction containing f, and/or some preventive measures, contain a large amount of uncertainty (which is usually not acknowledged well enough), and b) obtaining more accurate estimates of in particular f, should be highly prioritized. Seroprevalence testing of random samples in a community where the epidemic has ended are urgently needed.