A survey of federated learning for edge computing: Research problems and solutions

This paper introduces a framework for how to appropriately adopt and adjust Machine Learning (ML) techniques used to construct Electrocardiogram (ECG) based biometric authentication schemes. The proposed framework can help investigators and developers on ECG based biometric authentication mechanisms define the boundaries of required datasets and get training data with good quality. To determine the boundaries of datasets, use case analysis is adopted. Based on various application scenarios on ECG based authentication, three distinct use cases (or authentication categories) are developed. With more qualified training data given to corresponding machine learning schemes, the precision on ML-based ECG biometric authentication mechanisms is increased in consequence. ECG time slicing technique with the R-peak anchoring is utilized in this framework to acquire ML training data with good quality. In the proposed framework four new measure metrics are introduced to evaluate the quality of ML training and testing data. In addition, a Matlab toolbox, containing all proposed mechanisms, metrics and sample data with demonstrations using various ML techniques, is developed and made publicly available for further investigation. For developing ML-based ECG biometric authentication, the proposed framework can guide researchers to prepare the proper ML setups and the ML training datasets along with three identified user case scenarios. For researchers adopting ML techniques to design new schemes in other research domains, the proposed framework is still useful for generating ML-based training and testing datasets with good quality and utilizing new measure metrics.

e23086 Background: It is anticipated that implementing machine learning (ML) strategies will impact the clinical decision-making process in the future. This study aimed to compare the accuracy of various machine learning algorithms in predicting suicidal ideation in patients who were receiving treatment for prostate cancer (PC). Further, to predict suicidal ideation and explain the burden of illness (BOI) by utilizing ML algorithms and XAI-based interpretability. Methods: We analyzed the United States 2017 National Inpatient Sample database for patients hospitalized for PC using linear and non-linear ML methods. Using techniques such as forward selection (FS) and backward elimination (BE), Random Forest (RF), decision trees, Multivariate Adaptive Regression Splines, and Gradient Boosting Machine (GBM), we determined subsets and features. We used linear and non-linear MLs-- Lasso, Ridge, RF, and Neural Networks (NN). A 70% and 30% partitioning was performed on the training and test datasets. The performance of the model was tested using discrimination (C-statistics), the Receiver-Operating Characteristics (ROC) curve, and the Hosmer-Lemeshow tests on both the training data and the test data. Several XAI approaches, including permutation significance, global surrogate marker, feature interpretation, interactivity, and local interpretability, were also utilized to examine the BOI (Length of stay (LOS)) among PC cancer cohort with suicide ideation. Results: We identified 680 patients with suicidal ideation among 208,730 PC patients. The most important variable derived through FS and BE were depression, drug abuse, psychiatric disorder, alcohol dependence, race, Medicaid beneficiaries, cardiac arrhythmias, metastasis, weight loss, congestive heart failure, and un-complicated hypertension, further confirmed through the GBM and other methods. There was successful documentation of demographic, socioeconomic, and clinical characteristics through feature selection approaches. Linear models, Lasso, and Ridge demonstrated an excellent area under the ROC for all cohorts ( > 0.9 in train and test datasets). On the other hand, tree-based models, such as RF, performed poorly (AUC 0.72 for train and 0.66 for test). The findings of the XAI methodologies showed that among PC suicide ideation, higher BOI was associated with patients staying in low-income neighborhoods, Whites, Medicaid beneficiaries, and those with weight loss. Conclusions: We showcase the increasing capability of predictive analytics, particularly XAI, in predicting suicidal ideation. The combination of the vast amount of data generated by the digitalization of healthcare systems with the advanced processing power of modern servers enables the development of machine learning-based care pathways, which has the potential to revolutionize clinical decision-making in cancer.

Machine learning (ML) approaches to predicting quantum mechanical (QM) properties have made great strides toward achieving the computational chemist's holy grail of structure-based property prediction. In contrast to direct ML methods, which encode a molecule with only structural information, in this work, we show that QM descriptors improve ML predictions of dimer interaction energy, both in terms of accuracy and data efficiency, by incorporating electronic information into the descriptor. We present the electron deformation density interaction energy machine learning (EDDIE-ML) model, which predicts the interaction energy as a function of Hartree-Fock electron deformation density. We compare its performance with leading direct ML schemes and modern DFT methods for the prediction of interaction energies for dimers of varying charge type, size, and intermolecular separation. Under a low-data regime, EDDIE-ML outperforms other direct ML schemes and is the only model readily transferrable to larger, more complex systems including base pair trimers and porous cages. The underlying physical connection between the density and interaction energy enables EDDIE-ML to reach an accuracy comparable to modern DFT functionals in fewer training data points compared to other ML methods.

Impact craters, resulting from the collision of meteorites, asteroids, or comets with planetary surfaces, manifest as circular-elliptical depressions with diverse sizes and shapes influenced by various factors. These morphological features play a crucial role in planetary exploration, offering insights into the geological composition and structure of celestial bodies. Beyond their scientific importance, craters may also hold valuable natural resources, such as frozen water in the Moon's permanently shadowed craters. Furthermore, understanding craters’ spatial distribution is pivotal for terrain-relative navigation and for selecting future landing sites. Manual crater mapping through visual inspection is an impractical and laborious process, often unattainable for large-scale investigations. Moreover, manual crater mapping is susceptible to human errors and biases, leading to potential disagreements of up to 40%. In order to tackle these issues, semi-automatic crater detection algorithms (CDA) have been developed to mitigate human biases, and to enable large-scale and real-time crater detection and mapping. The majority of CDAs’ are based on machine learning (ML) and data-driven methods. ML-based CDAs’ are trained in a supervised manner using specific datasets that were manually labelled. Because of that, existing ML-based CDAs’ are constrained to specific data types according to the type of their training data. This makes current ML-based CDAs’ unstable and un-practical, since applying an ML scheme to a different type of data requires acquiring and labelling a new training set, and subsequently use it to train a new ML scheme, or fine-tune an already existing one. In this study, we describe a universal approach [1] for crater identification based on Segment Anything Model (SAM), a foundational computer vision and image segmentation model developed by META [2]. SAM was trained with over 1 billion masks, and is capable to segment various data types (e.g., photos, DEM, spectra, gravity) from different celestial bodies (e.g., Moon, Mars) and measurement setups. The segmentation output undergoes further classification into crater and no-crater based on geometric indices assessing circular and elliptical attributes of the investigated mask. The proposed framework is proven effective across different datasets from various planetary bodies and measurement configurations. The outcomes of this study underlines the potential of foundational segmentation models in planetary science. Foundational models tuned for planetary data can provide universal classifiers contributing towards an automatic scheme for identifying, detecting and mapping various morphological and geological targets in different celestial bodies.  

In recent power grid systems, data-driven approach has been taken to grid condition evaluation and classification after successful adoption of big data techniques in internet applications. However, the raw training data from single monitoring system, e.g. dissolved gas analysis (DGA), are rarely sufficient for training in the form of valid instances and the data quality can rarely meet the requirement of precise data analytics since raw data set usually contains samples with noisy data. This paper proposes a machine learning scheme (PCA_IR) to improve the accuracy of fault diagnose, which combines dimension-increment procedure based on association analysis, dimension-reduction procedure based on principal component analysis and back propagation neural network (BPNN). First, the dimension of training data is increased by adding selected data which originates from different source such as production management system (PMS) to the original data obtained by DGA. The added data would also inevitably result in more noise. Thus, we then take advantage of the PCA method to reduce the noise in the training data as well as retaining significant information for classification. Finally, the new training data yielded after PCA procedure is inputted into BPNN for classification. We test the PCA_IR scheme on fault diagnosis of power transformers in power grid system. The experimental results show that the classifiers based on our scheme achieve higher accuracy than traditional ones. Therefore, the scheme PCA_IR would be successfully deployed for fault diagnosis in power grid system.

The traditional method for gaining knowledge on the state of a motor is to take test data in time increments and plot the progression of the machine parameters, looking for trending data. This leads to a reactive reliability mode, where one performs maintenance or takes corrective action once data have been collected that indicates cause for concern. Standard test methods are generally employed and over time one may become a knowledgeable expert on the condition of the motor and when to repair it prior to failure.Emulation and Uncertainty Quantification (UQ) can be used as a machine learning powered approach to improve upon traditional predictive maintenance practices, allowing for corrective action to be taken prior to the real-time data itself indicating concern. Machine Learning strategies also take the potentially unreliable human guess work out of the intuition-based approach to predicting failure based on prior experience and knowledge.Emulators (aka predictive models) are statistical models trained using advanced analytics and machine learning algorithms to learn the input-output relationships of an underlying data set, often called the training data. Once trained, the key strength of the emulator is the ability to rapidly make predictions of the output of a system for input combinations not contained in the training data, eliminating the need for further direct data collection to perform any desired analyses.Data collected from a motor can be used to train an emulator capable of predicting the future performance of that motor. If these predictions indicate a need for corrective action, the predictive speed of the emulator can be used to check the outcomes of different "what-if" scenarios to determine the best course of action. UQ tools can be used along with the statistical prediction process to place error bounds on the various outcomes.This paper will present the above concepts in more detail. This will include the steps of emulator training, beginning with a design of experiments to select appropriate training data to be collected from the motor, validation of the emulator’s predictive accuracy, and its use for predictive maintenance and UQ, including sensitivity analyses of inputs on the output(s) of interest, uncertainty propagation, and optimization.

Clinical drug–drug interactions (DDIs) have been a major cause for not only medical error but also adverse drug events (ADEs). The published literature on DDI clinical toxicity continues to grow significantly, and high-performance DDI information retrieval (IR) text mining methods are in high demand. The effectiveness of IR and its machine learning (ML) algorithm depends on the availability of a large amount of training and validation data that have been manually reviewed and annotated. In this study, we investigated how active learning (AL) might improve ML performance in clinical safety DDI IR analysis. We recognized that a direct application of AL would not address several primary challenges in DDI IR from the literature. For instance, the vast majority of abstracts in PubMed will be negative, existing positive and negative labeled samples do not represent the general sample distributions, and potentially biased samples may arise during uncertainty sampling in an AL algorithm. Therefore, we developed several novel sampling and ML schemes to improve AL performance in DDI IR analysis. In particular, random negative sampling was added as a part of AL since it has no expanse in the manual data label. We also used two ML algorithms in an AL process to differentiate random negative samples from manually labeled negative samples, and updated both the training and validation samples during the AL process to avoid or reduce biased sampling. Two supervised ML algorithms, support vector machine (SVM) and logistic regression (LR), were used to investigate the consistency of our proposed AL algorithm. Because the ultimate goal of clinical safety DDI IR is to retrieve all DDI toxicity–relevant abstracts, a recall rate of 0.99 was set in developing the AL methods. When we used our newly proposed AL method with SVM, the precision in differentiating the positive samples from manually labeled negative samples improved from 0.45 in the first round to 0.83 in the second round, and the precision in differentiating the positive samples from random negative samples improved from 0.70 to 0.82 in the first and second rounds, respectively. When our proposed AL method was used with LR, the improvements in precision followed a similar trend. However, the other AL algorithms tested did not show improved precision largely because of biased samples caused by the uncertainty sampling or differences between training and validation data sets.

Container security has received much research attention recently. Previous work has proposed to apply various machine learning techniques to detect security attacks in containerized applications. On one hand, supervised machine learning schemes require sufficient labelled training data to achieve good attack detection accuracy. On the other hand, unsupervised machine learning methods are more practical by avoiding training data labelling requirements, but they often suffer from high false alarm rates. In this paper, we present SHIL, a self-supervised hybrid learning solution, which combines unsupervised and supervised learning methods to achieve high accuracy without requiring any manual data labelling. We have implemented a prototype of SHIL and conducted experiments over 41 real world security attacks in 28 commonly used server applications. Our experimental results show that SHIL can reduce false alarms by 39-91% compared to existing supervised or unsupervised machine learning schemes while achieving a higher or similar detection rate.

Container security has received much research attention recently. Previous work has proposed to apply various machine learning techniques to detect security attacks in containerized applications. On one hand, supervised machine learning schemes require sufficient labeled training data to achieve good attack detection accuracy. On the other hand, unsupervised machine learning methods are more practical by avoiding training data labeling requirements, but they often suffer from high false alarm rates. In this article, we present a generic self-supervised hybrid learning (SHIL) framework for achieving efficient online security attack detection in containerized systems. SHIL can effectively combine both unsupervised and supervised learning algorithms but does not require any manual data labeling. We have implemented a prototype of SHIL and conducted experiments over 46 real-world security attacks in 29 commonly used server applications. Our experimental results show that SHIL can reduce false alarms by 33%–93% compared to existing supervised, unsupervised, or semi-supervised machine learning schemes while achieving a higher or similar detection rate.

A Machine-Learning Classification of Walking in the Community Using Inertial Sensors and Smart Insole

A machine learning strategy for optimal path planning of space robotic manipulator in on-orbit servicing

Machine learning may assist with the identification of potentially inappropriate penicillin allergy labels. Strategies to improve the performance of existing models for this task include the use of additional training data, synthetic data and transfer learning. The aims of this study were to investigate the use of additional training data and novel machine learning strategies, namely synthetic data and transfer learning, to improve the performance of penicillin adverse drug reaction (ADR) machine learning classification. Machine learning natural language processing was applied to free-text penicillin ADR data extracted from a public health system electronic health record (EHR). The models were developed by training on various labelled data sets. ADR entries were split into training and testing data sets and used to develop and test a variety of machine learning models. The effect of training on additional data and synthetic data versus the use of transfer learning was analysed. Following the application of these techniques, the area under the receiver operator curve of best-performing models for the classification of penicillin allergy (vs intolerance) and high-risk allergy (vs low-risk allergy) improved to 0.984 (using the artificial neural network model) and 0.995 (with the transfer learning approach) respectively. Machine learning models demonstrate high levels of accuracy in the classification and risk stratification of penicillin ADR labels using the reaction documented in the EHR. The model can be further optimised by incorporating additional training data and using transfer learning. Practical applications include automating case detection for penicillin allergy delabelling programmes.

Spider neurotoxins are commonly used as pharmacological tools and are a popular source of novel compounds with therapeutic and agrochemical potential. Since venom peptides are inherently toxic, the host spider must employ strategies to avoid adverse effects prior to venom use. It is partly for this reason that most spider toxins encode a protective proregion that upon enzymatic cleavage is excised from the mature peptide. In order to identify the mature toxin sequence directly from toxin transcripts, without resorting to protein sequencing, the propeptide cleavage site in the toxin precursor must be predicted bioinformatically. We evaluated different machine learning strategies (support vector machines, hidden Markov model and decision tree) and developed an algorithm (SpiderP) for prediction of propeptide cleavage sites in spider toxins. Our strategy uses a support vector machine (SVM) framework that combines both local and global sequence information. Our method is superior or comparable to current tools for prediction of propeptide sequences in spider toxins. Evaluation of the SVM method on an independent test set of known toxin sequences yielded 96% sensitivity and 100% specificity. Furthermore, we sequenced five novel peptides (not used to train the final predictor) from the venom of the Australian tarantula Selenotypus plumipes to test the accuracy of the predictor and found 80% sensitivity and 99.6% 8-mer specificity. Finally, we used the predictor together with homology information to predict and characterize seven groups of novel toxins from the deeply sequenced venom gland transcriptome of S. plumipes, which revealed structural complexity and innovations in the evolution of the toxins. The precursor prediction tool (SpiderP) is freely available on ArachnoServer (http://www.arachnoserver.org/spiderP.html), a web portal to a comprehensive relational database of spider toxins. All training data, test data, and scripts used are available from the SpiderP website.

Timely identification of patients likely to miss seasonal influenza vaccination (SIV) could help health care practitioners tailor services and gain efficiency. To develop and validate a predictive model of SIV uptake among at-risk adults. This prognostic study constructed a prediction model for vaccine uptake by adults at increased risk of influenza-associated complications. Drawing from the Clinical Practice Research Datalink database's records of primary care data of 324 284 adults routinely collected at general practices across England from January 2011 to December 2016, logistic regression models were trained on data from patients registered from January 2012 to December 2013 and validated with out-of-sample data from patients registered from January 2015 to December 2016. Data were extracted from the database December 2018 and analyzed between September 2019 and December 2019. Covariates included sex, age, race/ethnicity, smoking status, socioeconomic status, previous pneumococcal vaccination, prior season SIV uptake, and clinical risk conditions. The main outcome was patient-level SIV uptake. Model performance was measured via misclassification rate, Brier score, sensitivity, specificity, and area under the curve. The training data sets consisted of 324 284 (aged 18 to 64 years) and 186 426 (aged 65 years or older) patients. The mean (SD) age in the training data among patients aged 18 to 64 years was 45 (13) years; 161 487 (49.8%) were women, and 102 133 (31.5%) were categorized as white. Among patients aged 65 years or older, the mean (SD) age was 77 (8) years; 96 169 (51.6%) were women, and 64 996 (34.9%) were categorized as white. The validation data sets consisted of 35 210 patients aged 18 to 64 years and 25 497 aged 65 years or older. The mean (SD) age in the validation data set among patients aged 18 to 64 years was 42 (14) years; 17 296 (49.1%) were women, and 13 346 (37.9%) were categorized as white. Among patients aged 65 years or older, the mean (SD) age was 73 (8) years; 13 135 (51.5%) were women, and 9641 (37.8) were categorized as white. Among patients aged 18 to 64 years, SIV uptake was 35.9% (95% CI, 35.7%-36.0%) and 32.6% (95% CI, 32.1%-33.1%) for the training and validation data sets, respectively. Among patients aged 65 years or older, SIV uptake was 83.1% (95% CI, 82.9%-83.2%) and 76.1% (95% CI, 75.5%-76.6%) for the training and validation data sets, respectively. Prior season SIV uptake and pneumococcal vaccination status were the best predictors of SIV uptake. Predicted SIV uptake probabilities for patients aged 18 to 64 years were reliable, but biased toward underpredicting, whereas, among patients aged 65 years or older, they were variable and biased toward overpredicting. Briefly, in out-of-sample validation among patients aged 18 to 64 years, misclassification rates were 0.163 to 0.164, Brier scores were 0.124 to 0.125, area under the receiver operating characteristic curve values ranged from 0.876 to 0.877, sensitivity ranged from 0.705 to 0.720, and specificity ranged from 0.896 to 0.902. In patients aged 65 years or older, misclassification rates were 0.120 to 0.125, Brier scores were 0.0953 to 0.0959, area under the receiver operating characteristic curve was 0.877, sensitivity ranged from 0.919 to 0.936, and specificity ranged from 0.680 to 0.753. This study suggests that data obtained from primary care records could accurately predict SIV uptake among at-risk adults. Further research is needed to assess the feasibility and efficacy of implementing this model in clinical settings.

Joint think locally and globally: Communication-efficient federated learning with feature-aligned filter selection