Multivariate Prediction Research Articles

Application of LSTM neural network in the research of cyclical industries: taking the automotive industry as an illustrative case

Anticipating future developmental trends within an industry stands as one of the paramount responsibilities for industry researchers. Nevertheless, owing to the intricate and subjective nature inherent in industry research, coupled with researchers usually having to rely on their own subjective judgments when making predictions, the forecasting outcomes frequently prove unsatisfactory. In order to address this issue, this study introduces deep learning technology into the industry research field and employs Long & Short-Term Memory (LSTM) networks to forecast a typical cyclical industry - the automotive industry. In the specific experiment, Adam optimization algorithm is employed to enhance model training speed, followed by utilization of the grid search algorithm for optimal parameter selection. Subsequently, the obtained optimal model is utilized for LSTM multivariate sequence prediction in order to forecast the future development trend of the automotive industry over the next six months. The ultimate outcome demonstrates that the LSTM model accurately anticipated the forthcoming U-shaped rebound trend in the automotive industry, albeit with some limitations in predicting precise numerical values and timing. Nevertheless, it aligns closely with the actual developmental trajectory. This investigation represents a constructive application of deep learning technology in industrial research, and its final findings hold significant implications for future industry research endeavors.

The Pre-ECPR Score: Developing and Validating a Multivariable Prediction Model for Favorable Neurological Outcomes in Patients Undergoing Extracorporeal Cardiopulmonary Resuscitation

Extracorporeal cardiopulmonary resuscitation (ECPR) can save patients with refractory cardiac arrest; however, according to recent meta-analyses, only 20% of patients achieve favorable outcomes (Modified Rankin Scale 0-3). We aimed to develop and validate an ECPR prediction model to improve patient selection. Prognostic model development and internal validation study. Single-center study. All 120 normothermic ECPR patients treated at Sahlgrenska University Hospital between January 2010 and October 2021. None. Multivariable logistic regression was used to develop the PRognostic Evaluation of ECPR (Pre-ECPR) score. Model performance was assessed through the area under curve (AUC) and compared with the Extracorporeal Life Support Organization (ELSO) "Example of selection criteria for ECPR" for 1-year survival with favorable outcomes. The positive predictive value (PPV) was calculated. Favorable outcomes occurred in 27.5% of the patients. The Pre-ECPR score, incorporating age, no-flow/initial rhythm (a composite variable), total cardiac arrest time, signs of life, pupil dilation, regional cerebral oxygen saturation, arterial pH, and end-tidal CO2, demonstrated an AUC of 0.87 (95% confidence interval [CI] 0.77-0.93). In internal cross-validation, the AUC of 0.79 (95% CI 0.67-0.88) significantly outperformed the ELSO criteria AUC of 0.63 (95% CI 0.54-0.72, p = 0.012). Pre-ECPR score probabilities >6.4% showed 100% sensitivity and a PPV of 40.5% for favorable outcomes. The Pre-ECPR score combines multiple weighted predictors to provide a single balanced probability of favorable outcomes in ECPR patient selection. In cross-validation, it demonstrated significantly more favorable discriminatory performance than that of the ELSO criteria.

Development and validation of supervised machine learning multivariable prediction models for the diagnosis of Pneumocystis jirovecii pneumonia using nasopharyngeal swab PCR in adults in a low-HIV prevalence setting.

The global burden of the opportunistic fungal disease Pneumocystis jirovecii pneumonia (PJP) remains substantial. Polymerase chain reaction (PCR) on nasopharyngeal swabs (NPS) has high specificity and may be a viable alternative to the gold standard diagnostic of PCR on invasively collected lower respiratory tract specimens, but has low sensitivity. Sensitivity may be improved by incorporating NPS PCR results into machine learning models. Three supervised multivariable diagnostic models (random forest, logistic regression and extreme gradient boosting) were constructed and validated using a 111-person Australian dataset. The predictors were age, gender, immunosuppression type and NPS PCR result. Model performance metrics such as accuracy, sensitivity, specificity and predictive values were compared to select the best-performing model. The logistic regression model performed best, with 80% accuracy, improving sensitivity to 86% and maintaining acceptable specificity of 70%. Using this model, positive and negative NPS PCR results indicated post-test probabilities of 84% (likely PJP) and 26% (unlikely PJP), respectively. The logistic regression model should be externally validated in a wider range of settings. As the predictors are simple, routinely collected patient variables, this model may represent a diagnostic advance suitable for settings where collection of lower respiratory tract specimens is difficult but PCR is available.

Optimization of process parameters in laser cladding multi channel forming using MVBM-NSGA-II method

ABSTRACT Laser cladding multi-channel forming is a complex, nonlinear process characterized by multi-parameter coupling. To determine the technology parameters, an optimization method of multivariate black-box prediction model (MVBM) matching non-dominated sequential genetic algorithm (NSGA-II) was constructed. Multi-channel LC orthogonal experiments were designed. The MVBM was established with process technology parameters as inputs and characteristic parameters reflecting the quality (surface hardness, dilution rate) as responses. The set of optimal solutions for technology parameters by the NSGA-II. The experimental results showed that, under the optimum parameters, the average hardness of the optimized increased by 6.1%, and the dilution rate was reduced by 49.8%. The dilution rate was maintained within the optimal range. The study’s findings can provide theoretical support for optimizing LC multi-objective technology parameters and improving coating quality control.

Advancing water absorption capacity in hard winter wheat using a multivariate genomic prediction approach

AbstractThe water absorption capacity (WAC) of hard wheat (Triticum aestivum L.) flour affects end‐use quality characteristics, including loaf volume, bread yield, and shelf life. However, improving WAC through phenotypic selection is challenging. Phenotyping for WAC is time consuming and, as such, is often limited to evaluation in the latter stages of the breeding process, resulting in the retention of suboptimal lines longer than desired. This study investigates the potential of univariate and multivariate genomic predictions as an alternative to phenotypic selection for improving WAC. A total of 497 hard winter wheat genotypes were evaluated in multi‐environment advanced yield and elite trials over 8 years (2014–2021). Phenotyping for WAC was done via the solvent retention capacity (SRC) using water as a solvent (SRC‐W). Traits that exhibited a significant correlation (r ≥ 0.3) with SRC‐W and were evaluated earlier than SRC‐W were included in the multivariate genomic prediction models. Kernel hardness and diameter were obtained using the single kernel characterization system (SKCS), and break flour yield and total flour yield (T‐Flour) were included. Cross‐validation showed the mean univariate genomic prediction accuracy of SRC to be r = 0.69 ± 0.005, while bivariate and multivariate models showed an improved prediction accuracy of r = 0.82 ± 0.003. Forward validation showed a prediction accuracy up to r = 0.81 for a multivariate model that included SRC‐W + All traits (SRC‐W, Diameter, SKCS hardness and diameter, F‐Flour, and T‐Flour). These results suggest that incorporating correlated traits into genomic prediction models can improve early‐generation prediction accuracy.

Quantitative parameters of HRCT target scan to predict the risk of lung adenocarcinoma based on the detection of lung ground-glass nodules.

To explore the value of high-resolution computed tomography (HRCT) in the differential diagnosis of benign and malignant ground-glass nodules (GGNs), and to provide a theoretical basis for the clinical application of HRCT. A total of 208 patients with GGN who had been clinically confirmed by surgical pathology and clinical confirmation were collected, and HRCT target scanning technology was used to scan and collect general information of patients, and observe the distribution of GGN, GGN size, GGN cross-sectional area, diameter, transverse diameter, solid composition, relationship with bronchi, and relationship with blood vessels and other indicators. Multivariate regression analysis and risk factor prediction are performed. The differences were statistically significant in multivariate regression analysis, such as nodule location, maximum diameter, maximum cross-sectional area, GGN status, nodule boundary and relationship with blood vessels (P < 0.05). The results of ROC curve showed that the AUC value of nodule site and nodule boundary was greater than 0.5, and the nodule boundary AUC value was 0.676, which was more sensitive to predict whether GGN deteriorated to lung adenocarcinoma (LUAD). Nodule site and nodule boundary are effective risk predictors for LUAD in patients with GGN, and nodule boundary is the most valuable independent predictor.

Short-term air pollution prediction using graph convolutional neural networks

Pollution is a major concern in the present day, causing multiple illnesses and deaths, specifically in developing countries in Asia and Africa. While it has drawn worldwide attention as governments try to issue laws to meet certain criteria for air pollution levels, pollution concentration forecasting has become a major challenge. Particularly, short term forecasting will help to gain information regarding concentrations of harmful pollutants for the upcoming hours and enable better decision-making with regards to controlling air pollution. In this paper, we investigate spatio-temporal graph-based models to determine the best methods for spatial and temporal analysis of data. The models have the additional capacity to perform multi-variate predictions of correlated data, i.e., predicting multiple pollutant concentrations simultaneously, thus requiring lower computational efforts. A real-world pollution dataset measured over Delhi, India, is used to comparing the proposed models with baselines, which shows the Spatio-Temporal Graph Convolution Neural Network (STGCN) models to be performing better than others. For a better understanding of model architectures with the most effective strategies for spatial and temporal data analysis, three models, namely STGCN-A, STGCN-B, STGCN-C have been developed. The models have been compared with 6 other baselines over multiple forecasting horizons of 1 h, 24 h, and 48 h timesteps using various metrics such as mean absolute error (MAE), root mean square error (RMSE), mean absolute percent error (MAPE). On the PM2.5 dataset of Delhi, STGCN-B achieves a performance of 10.53 MAE, 6.92 RMSE and 25.25 MAPE for a 1 h forecast, while STGCN-C achieves 20.18 MAE, 14.73 RMSE and 55.45 MAPE for a 24 h forecast. In general, both structures achieve similar results, with STGCN-C being better in many cases. They are further analysed through observation-prediction graphs and Taylor diagrams, which give an insight into our findings. The models are additionally validated on a benchmark real-world dataset from California, USA for better understanding of the spatio-temporal relations and model performances on a more stable dataset, where STGCN-C performs best for PM2.5 with 4.30 RMSE, 1.98 MAE, 25.96 MAPE for 1 h predictions for univariate data and 3.63 RMSE, 1.88 MAE and 25.91 MAPE in multivariate forecasting. The developed spatio-temporal graph-based models hold promising applications in urban air quality management, aiding policymakers in implementing targeted interventions to mitigate pollution-related health risks. Furthermore, these models can support public health agencies by providing timely and accurate forecasts of pollutant concentrations, enabling proactive measures to safeguard community well-being. Our study showcases the efficacy of spatio-temporal graph-based models in accurately forecasting air pollutant concentrations, with particular emphasis on short-term predictions. By leveraging multi-variate capabilities, our proposed models demonstrate superior performance compared to baseline approaches across various forecasting horizons.

Multivariate time series approaches to extract predictive asthma biomarkers from prospectively patient-collected diary data: a systematic review

ObjectivesLongitudinal data are common in asthma studies, to assess asthma progression in patients and identify predictors of future outcomes, including asthma exacerbations and asthma control. Different methods can quantify temporal...

Forecasting international tourist arrivals in South Korea: a deep learning approach

Purpose This study aims to develop a robust long short-term memory (LSTM)-based forecasting model for daily international tourist arrivals at Incheon International Airport (ICN), incorporating multiple predictors including exchange rates, West Texas Intermediate (WTI) oil prices, Korea composite stock price index data and new COVID-19 cases. By leveraging deep learning techniques and diverse data sets, the research seeks to enhance the accuracy and reliability of tourism demand predictions, contributing significantly to both theoretical implications and practical applications in the field of hospitality and tourism. Design/methodology/approach This study introduces an innovative approach to forecasting international tourist arrivals by leveraging LSTM networks. This advanced methodology addresses complex managerial issues in tourism management by providing more accurate forecasts. The methodology comprises four key steps: collecting data sets; preprocessing the data; training the LSTM network; and forecasting future international tourist arrivals. The rest of this study is structured as follows: the subsequent sections detail the proposed LSTM model, present the empirical results and discuss the findings, conclusions and the theoretical and practical implications of the study in the field of hospitality and tourism. Findings This research pioneers the simultaneous use of big data encompassing five factors – international tourist arrivals, exchange rates, WTI oil prices, KOSPI data and new COVID-19 cases – for daily forecasting. The study reveals that integrating exchange rates, oil prices, stock market data and COVID-19 cases significantly enhances LSTM network forecasting precision. It addresses the narrow scope of existing research on predicting international tourist arrivals at ICN with these factors. Moreover, the study demonstrates LSTM networks’ capability to effectively handle multivariable time series prediction problems, providing a robust basis for their application in hospitality and tourism management. Originality/value This research pioneers the integration of international tourist arrivals, exchange rates, WTI oil prices, KOSPI data and new COVID-19 cases for forecasting daily international tourist arrivals. It bridges the gap in existing literature by proposing a comprehensive approach that considers multiple predictors simultaneously. Furthermore, it demonstrates the effectiveness of LSTM networks in handling multivariable time series forecasting problems, offering practical insights for enhancing tourism demand predictions. By addressing these critical factors and leveraging advanced deep learning techniques, this study contributes significantly to the advancement of forecasting methodologies in the tourism industry, aiding decision-makers in effective planning and resource allocation.

Study on the changing patterns of production performance of laying hens and their relationships with environmental factors in a large-scale henhouse

The production performance of laying hens is influenced by various environmental factors within the henhouse. The intricate interactions among these factors make the impact process highly complicated. The exact relationships between production performance and environmental variables are still not well understood. In this study, we measured the production performance of laying hens and various environmental variables across different parts of the henhouse, evaluated the weight of each environmental variable, and constructed a laying rate prediction model. Results displayed that body weight, laying rate, egg weight and eggshell thickness of hens decrease gradually from WCA to FA (P < 0.05). Serum levels of FSH and LH, as well as antibody level of H5 Re-13, gradually decrease from WCA to FA (P < 0.05). Moreover, the values for temperature (T), temperature-humidity index (THI), air velocity (AV), carbon dioxide (CO2), and particulate matter (PM2.5) gradually increase from WCA to FA (P < 0.05). Conversely, the relative humidity (RH) value gradually decreases from FA to WCA (P < 0.05). Additionally, the weights of the environmental variables, determined using a combination of the grey relational analysis (GRA) and analytic hierarchy process (AHP), were as follows in descending order: RH, THI, T, light intensity (LI), AV, PM2.5, NH3, and CO2. When the number of decision trees in the laying rate prediction model was set to 2,500, the results displayed a high level of agreement between the model's predictions and the observed outcomes. The model's performance evaluation yielded an R2 value of 0.89995 for the test set, suggesting strong predictive effects. In conclusion, the current study revealed significant differences in both the production performance of laying hens and the environmental variables across different parts of the henhouse. Furthermore, the study demonstrated that different environmental factors have distinct impacts on laying rate, with humidity and temperature identified as the primary factors. Finally, a multi-variable prediction model was constructed, exhibiting high accuracy in predicting laying rate.

Prediction of transforaminal epidural injection success in sciatica (POTEISS): a protocol for the development of a multivariable prediction model for outcome after transforaminal epidural steroid injection in patients with lumbar radicular pain due to disc herniation or stenosis

BackgroundTransforaminal epidural injections (TEI) can alleviate symptoms and help to maintain physical functioning and quality of life in patients with lumbar radicular pain. We aim to develop a prediction model for patient outcome after TEI in patients suffering from unilateral lumbar radicular pain due to lumbar disc herniation (LDH) or single-level spinal stenosis (LSS). The secondary aim is to estimate short-term patient outcome differences between LDH and LSS patients, the association between psychological variables and patient outcome, the rate of additional injections, surgery and complications, and to explore the short-term cost-effectiveness of TEI.MethodsThis study is designed as a multi-centre, observational, prospective cohort study in two large regional hospitals in the Netherlands. Patients diagnosed with unilateral lumbar radicular pain secondary to LDH or LSS and congruent with MRI findings, who are referred for TEI along usual care pathways, are eligible for study participation. A total of 388 patients with LDH or LSS will be included. A pre-defined set of demographic, clinical and radiological variables will be used as the predictors in the model. The primary outcome measure is the Numerical Rating Scale (NRS) for leg pain. Secondary outcome measures include back pain, physical functioning, perceived recovery, pain coping strategies, anxiety and depression and use of analgesics and physical therapy. Patients will be evaluated at baseline, 2 weeks and 6 weeks after treatment. NRS leg pain and Likert perceived recovery data will be used as the dependent variables in a generalized linear mixed model for prediction of TEI outcome, with internal validation of performance (explained variation) by bootstrap resampling. Cost-effectiveness for a period of 6 weeks prior to and after treatment will be performed with decision-analytic modelling.DiscussionPatients with severe lumbar radicular pain often request additional treatment when conservative care is insufficient. TEI can offer relief of symptoms. Currently, it is not possible to predict responsiveness to this treatment for individual patients. This study is designed to explore predictors that can differentiate between patients that will and will not have a positive outcome after TEI. This information may support treatment strategies for this patient group.Trial registrationThis study is registered at ClinicalTrials.gov database under registry number NCT04540068 on September 1, 2020.

Multi-Energy Load Prediction Method for Integrated Energy System Based on Fennec Fox Optimization Algorithm and Hybrid Kernel Extreme Learning Machine.

To meet the challenges of energy sustainability, the integrated energy system (IES) has become a key component in promoting the development of innovative energy systems. Accurate and reliable multivariate load prediction is a prerequisite for IES optimal scheduling and steady running, but the uncertainty of load fluctuation and many influencing factors increase the difficulty of forecasting. Therefore, this article puts forward a multi-energy load prediction approach of the IES, which combines the fennec fox optimization algorithm (FFA) and hybrid kernel extreme learning machine. Firstly, the comprehensive weight method is used to combine the entropy weight method and Pearson correlation coefficient, fully considering the information content and correlation, selecting the key factors affecting the prediction, and ensuring that the input features can effectively modify the prediction results. Secondly, the coupling relationship between the multi-energy load is learned and predicted using the hybrid kernel extreme learning machine. At the same time, the FFA is used for parameter optimization, which reduces the randomness of parameter setting. Finally, the approach is utilized for the measured data at Arizona State University to verify its effectiveness in multi-energy load forecasting. The results indicate that the mean absolute error (MAE) of the proposed method is 0.0959, 0.3103 and 0.0443, respectively. The root mean square error (RMSE) is 0.1378, 0.3848 and 0.0578, respectively. The weighted mean absolute percentage error (WMAPE) is only 1.915%. Compared to other models, this model has a higher accuracy, with the maximum reductions on MAE, RMSE and WMAPE of 0.3833, 0.491 and 2.8138%, respectively.

Are we crossing a minimum of the Gleissberg centennial cycle? Multivariate machine learning-based prediction of the sunspot number using different proxies of solar activity and spectral analysis

We propose a new method for predicting the solar cycle in terms of the sunspot number (SN) based on multivariate machine learning algorithms, various proxies of solar activity, and the spectral analysis of all considered time series via the fast Fourier transform (through the latter we identify periodicities with which to lag these series and thus generate new attributes –predictors– for incorporation in the prediction model). This combination of three different techniques in a single method is expected to enhance the accuracy and reliability of the solar activity prediction models developed to date. Thus, predictive results for SN are presented for Solar Cycles 25 (the current one) and 26 (using the 13-month smoothed SN, version 2) up until January 2038, yielding maximum values of 134.2 (in June 2024) and 115.4 (in May 2034), respectively, with a root mean squared error (RMSE) of 9.8. These results imply, on the one hand, a maximum of Cycle 25 below the average and, on the other hand, a lower peak than the preceding ones for Cycle 26, suggesting that Solar Cycles 24, 25, and 26 are part of a minimum of the centennial Gleissberg cycle, as occurred with Cycles 12, 13, and 14 in the final years of the 19th century and the early 20th century.

Association between adherence to life’s simple 7 metrics and risk of obstructive sleep apnea among adults in the United States

BackgroundWe aimed to explore the impact of adherence to Life’s Simple 7 (LS7) metrics on risk of obstructive sleep apnea (OSA), and the impact of inflammation on the association, in adults in the United States.MethodsData from 13,825 community-dwelling adults aged ≥ 20 years recruited in the National Health and Nutrition Examination Surveys (NHANES) 2005–2008, 2015–2018 was analyzed. The LS7 score was calculated based on the AHA definition of LS7 metrics. The diagnosis of OSA was based on self-reported symptoms of sleep disturbance using a standard questionnaire. The Multivariable Apnea Prediction (MAP) Index score was also calculated to assess the risk of OSA. Log-binominal regression and negative binomial regression were performed to estimate the associations between LS7 and OSA and MAP index, with odds ratios (ORs) and prevalence ratios (PRs) and their 95% confidence intervals (CIs) calculated. Mediation analysis was performed to estimate the mediating effects of inflammatory indicators on the associations.ResultsA total of 4473 participants (32.4%) had OSA, and the mean MAP index was 0.39. In fully adjusted log-binominal regression models, with total score < 6 as the reference, the ORs (95% CIs) for risk of OSA were 0.90 (0.73, 1.10), 0.76 (0.65, 0.89), 0.78 (0.64, 0.95), and 0.45 (0.38, 0.54) for total score = 6, total score = 7, total score = 8, and total score > 8, respectively (P for trend < 0.001). When LS7 score was analyzed as a continuous variable, each 1-point increase in LS7 score was associated with a 15% decrease in OSA risk (P < 0.001). In negative binominal regression models, the adjusted PRs (95% CIs) for the MAP index were 0.93 (0.90, 0.97), 0.87 (0.84, 0.91), 0.80 (0.77, 0.84), and 0.55 (0.53, 0.57) for total score = 6, total score = 7, total score = 8, and total score > 8, respectively (P for trend < 0.001). For each 1-point increase in LS7 score, the risk of OSA decreased by 13% (P < 0.001). Consistent results were observed in subgroup analysis. Mediation analysis indicated that inflammatory factors, including blood cell count, neutrophil count, and C-reactive protein, positively mediated the association of LS7 with OSA, with a mediation proportion of 0.022 (P = 0.04), 0.02 (P = 0.04), and 0.02 (P = 0.02), respectively.ConclusionsIn a nationally representative sample of US adults, adherence to LS7 metrics was independently associated with reduced OSA risk. Inflammation plays a mediating role in the association between LS7 and OSA.

Natriuretic Peptides to Classify Risk of Atrial Fibrillation Detection After Stroke: Analysis of the BIOSIGNAL and PRECISE Cohort Studies.

Prolonged cardiac monitoring (PCM) increases atrial fibrillation (AF) detection after ischemic stroke, but access is limited, and it is burdensome for patients. Our objective was to assess whether midregional proatrial natriuretic peptide (MR-proANP) and N-terminal pro-B-type natriuretic peptide (NT-proBNP) could classify people who are unlikely to have AF after ischemic stroke and allow better targeting of PCM. We analyzed people from the Biomarker Signature of Stroke Aetiology (BIOSIGNAL) study with ischemic stroke, no known AF, and ≥3 days cardiac monitoring. External validation was performed in the Preventing Recurrent Cardioembolic Stroke: Right Approach, Right Patient (PRECISE) study of 28 days of cardiac monitoring in people with ischemic stroke or transient ischemic attack and no known AF. The main outcome is no AF detection. We assessed the discriminatory value of MR-proANP and NT-proBNP combined with clinical variables to identify people with no AF. A decision curve analysis was performed with combined data to determine the net reduction in people who would undergo PCM using the models based on a 15% threshold probability for AF detection. We included 621 people from the BIOSIGNAL study. The clinical multivariable prediction model included age, NIH Stroke Scale score, lipid-lowering therapy, creatinine, and smoking status. The area under the receiver-operating characteristic curve (AUROC) for clinical variables was 0.68 (95% CI 0.62-0.74), which improved with the addition of log10MR-proANP (0.72, 0.66-0.78; p = 0.001) or log10NT-proBNP (0.71, 0.65-0.77; p = 0.009). Performance was similar for the models with log10MR-proANP vs log10NT-proBNP (p = 0.28). In 239 people from the PRECISE study, the AUROC for clinical variables was 0.68 (0.59-0.76), which improved with the addition of log10NT-proBNP (0.73, 0.65-0.82; p < 0.001) or log10MR-proANP (0.79, 0.72-0.86; p < 0.001). Performance was better for the model with log10MR-proANP vs log10NT-proBNP (p = 0.03). The models could reduce the number of people who would undergo PCM by 30% (clinical and log10MR-proANP), 27% (clinical and log10NT-proBNP), or 20% (clinical only). MR-proANP and NT-proBNP help classify people who are unlikely to have AF after ischemic stroke. Measuring MR-proANP or NT-proBNP could reduce the number of people who need PCM by 30%, without reducing the amount of AF detected. NCT02274727; clinicaltrials.gov/study/NCT02274727.

Developing a multi-variate prediction model for COVID-19 from crowd-sourced respiratory voice data

Aim: COVID-19 has affected more than 223 countries worldwide and in the post-COVID era, there is a pressing need for non-invasive, low-cost, and highly scalable solutions to detect COVID-19. This study focuses on the analysis of voice features and machine learning models in the automatic detection of COVID-19. Methods: We develop a deep learning model to identify COVID-19 from voice recording data. The novelty of this work is in the development of deep learning models for COVID-19 identification from only voice recordings. We use the Cambridge COVID-19 Sound database which contains 893 speech samples, crowd-sourced from 4,352 participants via a COVID-19 Sounds app. Voice features including Mel-spectrograms and Mel-frequency cepstral coefficients (MFCC) and convolutional neural network (CNN) Encoder features are extracted. Based on the voice data, we develop deep learning classification models to detect COVID-19 cases. These models include long short-term memory (LSTM), CNN and Hidden-Unit BERT (HuBERT). Results: We compare their predictive power to baseline machine learning models. HuBERT achieves the highest accuracy of 86% and the highest AUC of 0.93. Conclusions: The results achieved with the proposed models suggest promising results in COVID-19 diagnosis from voice recordings when compared to the results obtained from the state-of-the-art.

Machine Learning-Based Algorithm for the Early Prediction of Postoperative Hypocalcemia Risk After Thyroidectomy.

We used machine learning to develop and validate a multivariable algorithm allowing the accurate and early prediction of postoperative hypocalcemia risk. Postoperative hypocalcemia is frequent after total thyroidectomy. An early and accurate individualized prediction of the risk of hypocalcemia could guide the selective prescription of calcium supplementation only to patients most likely to present with hypocalcemia after total thyroidectomy. This retrospective study enrolled all patients undergoing total thyroidectomy in a single referral center between November 2019 and March 2022 (derivation cohort) and April 2022 and September 2022 (validation cohort). The primary study outcome was postoperative hypocalcemia (serum calcium under 80mg/L). Exposures were multiple clinical and biological variables prospectively collected and analyzed with various machine learning methods to develop and validate a multivariable prediction algorithm. Among 610/118 participants in the derivation/validation cohorts, 100 (16.4%)/26 (22%) presented postoperative hypocalcemia. The most accurate prediction algorithm was obtained with random forest and combined intraoperative parathyroid hormone measurements with 3 clinical variables (age, sex, and body mass index) to calculate a postoperative hypocalcemia risk for each patient. After multiple cross-validation, the area under the receiver operative characteristic curve was 0.902 (0.829-0.970) in the derivation cohort, and 0.928 (95% CI: 0.86; 0.97) in the validation cohort. Postoperative hypocalcemia risk values of 7% (low threshold) and 20% (high threshold) had, respectively, a sensitivity of 92%, a negative likelihood ratio of 0.11, a specificity of 90%, and a positive of 7.6 for the prediction of postoperative hypocalcemia. Using machine learning, we developed and validated a simple multivariable model that allowed the accurate prediction of postoperative hypocalcemia. The resulting algorithm could be used at the point of care to guide clinical management after total thyroidectomy.

Systematic review of methods used in prediction models with recurrent event data

BackgroundPatients who suffer from chronic conditions or diseases are susceptible to experiencing repeated events of the same type (e.g. seizures), termed ‘recurrent events’. Prediction models can be used to predict the risk of recurrence so that intervention or management can be tailored accordingly, but statistical methodology can vary. The objective of this systematic review was to identify and describe statistical approaches that have been applied for the development and validation of multivariable prediction models with recurrent event data. A secondary objective was to informally assess the characteristics and quality of analysis approaches used in the development and validation of prediction models of recurrent event data.MethodsSearches were run in MEDLINE using a search strategy in 2019 which included index terms and phrases related to recurrent events and prediction models. For studies to be included in the review they must have developed or validated a multivariable clinical prediction model for recurrent event outcome data, specifically modelling the recurrent events and the timing between them.The statistical analysis methods used to analyse the recurrent event data in the clinical prediction model were extracted to answer the primary aim of the systematic review. In addition, items such as the event rate as well as any discrimination and calibration statistics that were used to assess the model performance were extracted for the secondary aim of the review.ResultsA total of 855 publications were identified using the developed search strategy and 301 of these are included in our systematic review. The Andersen-Gill method was identified as the most commonly applied method in the analysis of recurrent events, which was used in 152 (50.5%) studies. This was closely followed by frailty models which were used in 116 (38.5%) included studies. Of the 301 included studies, only 75 (24.9%) internally validated their model(s) and three (1.0%) validated their model(s) in an external dataset.ConclusionsThis review identified a variety of methods which are used in practice when developing or validating prediction models for recurrent events. The variability of the approaches identified is cause for concern as it indicates possible immaturity in the field and highlights the need for more methodological research to bring greater consistency in approach of recurrent event analysis. Further work is required to ensure publications report all required information and use robust statistical methods for model development and validation.PROSPERO registrationCRD42019116031.

Performance evaluation of multivariate deep-time convolution neural architectures for short-term electricity forecasting: Findings and failures

Deep learning (DL) models hold great promise in enhancing the decision-making abilities of electricity market participants and system operators in the short term, as they excel at capturing the intricate interdependencies and uncertainties in electricity consumption. Nevertheless, currently, the adoption of DL models for electricity prediction is not keeping up with the advancements in DL research. Within this framework, the aims of this study are to examine the probabilistic, multivariate, and multihorizon time series prediction of deep time convolution neural (DTCN) models. The effectiveness of DTCN in forecasting electricity consumption over daily, weekly and monthly time horizons is emperically assessed. We analyze both probabilistic and deterministic predictions. The evaluation data comprises real residential electricity consumption, average temperature and humidity with hourly granularity, collected over one year from a residential building. For evaluation, we considered accuracy and uncertainties, sensitivity to hyperparameters, the impact of exogeneous variables, as well as computational efficiency. Results show that the probabilistic models offer greater advantages when it comes to forecasting on multihorizons, especially for diverse datasets with non-uniform time series models. Unlike traditional models, the DTCN models demonstrated superior performance in capturing complex patterns and reducing forecast errors in short-term scenarios. However, their performance can be influenced by the sensitivity of hyperparameters. These findings can act as a benchmark for quantitatively assessing the performance of novel multivariate DL models in predicting electricity demand.

Development and validation of the eMCI-CHD tool: A multivariable prediction model for the risk of mild cognitive impairment in patients with coronary heart disease.

This study aimed to develop and validate an eMCI-CHD tool based on clinical data to predict mild cognitive impairment (MCI) risk in patients with coronary heart disease (CHD). This cross-sectional study prospectively collected data from 400 patients with coronary heart disease (aged 55-90 years, 62% men) from July 2022 to September 2023 and randomized (7:3 ratio) them into training and validation sets. After determining the modeling variables through least absolute shrinkage and selection operator regression analysis, four ML classifiers were developed: logistic regression, extreme gradient boosting (XGBoost), support vector machine, and random forest. The performance of the models was evaluated using area under the ROC curve, accuracy, sensitivity, specificity, and F1 score. Decision curve analysis was used to assess the clinical performance of the established models. The SHapley Additive exPlanations (SHAP) method was applied to determine the significance of the features, the predictive model was visualized with a nomogram, and an online web-based calculator for predicting CHD-MCI risk scores was developed. Of 400 CHD patients (average age 70.86±8.74 years), 220 (55%) had MCI. The XGBoost model demonstrated superior performance (AUC: 0.86, accuracy: 78.57%, sensitivity: 0.74, specificity: 0.84, F1: 0.79) and underwent validation. An online tool (https://mr.cscps.com.cn/mci/index.html) with seven predictive variables (APOE gene typing, age, education, TyG index, NT-proBNP, C-reactive protein, and occupation) assessed MCI risk in CHD patients. This study highlights the potential for predicting MCI risk among CHD patients using an ML model-driven nomogram and risk scoring tool based on clinical data.