Training Subsets Research Articles

Distractor Generation for Lexical Questions Using Learner Corpus Data

Abstract Learner corpora with error annotation can serve as a source of data for automated question generation (QG) for language testing. In case of multiple choice gapfill lexical questions, this process involves two steps. The first step is to extract sentences with lexical corrections from the learner corpus. The second step, which is the focus of this paper, is to generate distractors for the retrieved questions. The presented approach (called DisSelector) is based on supervised learning on specially annotated learner corpus data. For each sentence a list of distractor candidates was retrieved. Then, each candidate was manually labelled as a plausible or implausible distractor. The derived set of examples was additionally filtered by a set of lexical and grammatical rules and then split into training and testing subsets in 4:1 ratio. Several classification models, including classical machine learning algorithms and gradient boosting implementations, were trained on the data. Word and sentence vectors from language models together with corpus word frequencies were used as input features for the classifiers. The highest F1-score (0.72) was attained by a XGBoost model. Various configurations of DisSelector showed improvements over the unsupervised baseline in both automatic and expert evaluation. DisSelector was integrated into an opensource language testing platform LangExBank as a microservice with a REST API.

Deep learning to estimate RECIST in patients with cancer treated in real-world settings.

1564 Background: Creating large oncology clinical-genomic datasets is laborious and time-consuming. Deep learning approaches that extract RECIST outcomes at scale from observational electronic health records (EHR) data could tremendously facilitate precision oncology research. Methods: This retrospective study included patients with solid tumors treated on therapeutic clinical trials from 2004-2022 at Dana-Farber Cancer Institute with radiology reports in EHR and RECIST labels available from the tumor imaging metrics core. Each RECIST label was generated using corresponding radiology reports at a given time point. Patients were sampled into training, validation, and held-out test sets. A deep learning model (RECIST model) was trained to predict two outcomes: overall response and progressive disease at each time point, using reports from that time point and prior time points for each patient. This trained RECIST deep learning model was deployed on real-world radiology reports, and results were compared with true labels abstracted by trained human curators using the PRISSMM framework. Results: This study included 5153 patients with a total of 99,318 radiological reports who had RECIST annotations (median age at protocol enrollment [IQR] 60 years [52-57]; females 61% [3133]; and white 90% [4653]). The most common cancer types included breast (n=1006; 20%) lung (n=573; 11%), and ovarian cancer (n=539; 10%). The training subset included 4121 (79.9%) patients, the validation subset 518 (10.1%), and the held-out test set 514 (9.9%) patients. In the test set, the results showed an AUC of 0.86, and 0.87 with the best F1 scores of 0.72 and 0.63 for predicting overall response, and progressive disease, respectively. The real-world data set included 4482 patients. RECIST model evaluation on real-world radiology reports showed good performance for ascertaining PRISSMM annotations of progression from those reports (AUC 0.83, best F1 0.68), but poor performance for ascertaining PRISSMM annotations of response (AUC 0.63, best F1 0.28). Evaluation metrics are outlined in the Table. Conclusions: This study demonstrated the feasibility of using deep learning approaches to predict RECIST outcomes from radiology reports for patients with solid cancer tumors. The results show that the deep learning model accurately predicted RECIST labels on a held-out test set and real-world radiology reports for the outcome of progressive disease. These findings could accelerate precision oncology research providing a scalable way of ascertaining cancer outcomes from observational EHR data.[Table: see text]

Inferring pseudogene–MiRNA associations based on an ensemble learning framework with similarity kernel fusion

Accumulating evidence shows that pseudogenes can function as microRNAs (miRNAs) sponges and regulate gene expression. Mining potential interactions between pseudogenes and miRNAs will facilitate the clinical diagnosis and treatment of complex diseases. However, identifying their interactions through biological experiments is time-consuming and labor intensive. In this study, an ensemble learning framework with similarity kernel fusion is proposed to predict pseudogene–miRNA associations, named ELPMA. First, four pseudogene similarity profiles and five miRNA similarity profiles are measured based on the biological and topology properties. Subsequently, similarity kernel fusion method is used to integrate the similarity profiles. Then, the feature representation for pseudogenes and miRNAs is obtained by combining the pseudogene–pseudogene similarities, miRNA–miRNA similarities. Lastly, individual learners are performed on each training subset, and the soft voting is used to yield final decision based on the prediction results of individual learners. The k-fold cross validation is implemented to evaluate the prediction performance of ELPMA method. Besides, case studies are conducted on three investigated pseudogenes to validate the predict performance of ELPMA method for predicting pseudogene–miRNA interactions. Therefore, all experiment results show that ELPMA model is a feasible and effective tool to predict interactions between pseudogenes and miRNAs.

Using swarm intelligence optimization algorithms to predict the height of fractured water-conducting zone

The accurate calculation of the height of fractured water-conducting zone (FWCZ) is of great significance for mine optimization design, water disaster prevention, and safety production of the coal mines. In this article, a height-prediction model of FWCZ based on extreme learning machine (ELM) is proposed. To address the issues of low prediction accuracy and challenging parameter optimization, we optimized the ELM model using the gray-wolf optimization algorithm (GOA), whale optimization algorithm (WOA), and salp optimization algorithm (SOA). These optimization algorithms mitigate the issues of slow convergence, poor stability, and local optimality associated with traditional neural networks. The mining depth, mining height, overburden strata structure, working face length, and coal seam dip angle are selected as the main controlling factors for the height of FWCZ. A total of 42 fields-measured samples are collected and divided into 2 subsets for training and validating with a ratio of 36/6. The prediction capability of GOA-ELM, WOA-ELM, and SOA-ELM models are evaluated and compared, and the results show that the calculation results of the three models are optimized compared with the ELM model. The prediction capability of GOA and WOA are similar, while the prediction results of SOA-ELM are better than the other two models, and the relative errors of the test sets are all less than 10%. Therefore, the SOA-ELM model is finally applied to predict the height of FWCZ formed after the mining of No.15 coal seam in Xinjian Coal Mine. Finally, we verified the prediction results using measured data from the borehole television detection instrument, which showed good consistency. This provides further evidence of the effectiveness of the swarm intelligence optimization algorithm in predicting the height of FWCZ.

0503 Respiratory Effort During Sleep And Prevalent Diabetes In Obstructive Sleep Apnea

Abstract Introduction Obstructive sleep apnea (OSA) and type 2 diabetes (T2Dia) is a frequent clinical association. Beyond OSA-related hypoxic burden, increased respiratory effort (RE) is one of the main features of OSA and contributes to sympathetic overactivity that in turn might participate to glucose homeostasis dysregulation. The independent contribution of RE to the OSA-T2Dia pathogenesis remains to be demonstrated. The study aims to determine the impact of metrics assessing sleep RE derived from the sleep mandibular jaw movements (MJM) signal on the prevalence of T2Dia in a large cohort of patients addressed for in-laboratory conventional polysomnography (PSG) with suspicion of OSA. Methods An interpretable machine learning model was built to predict T2Dia from clinical data, PSG indices, and MJM-derived parameters (including the time spent with increased RE, expressed in % of TST). The analysis included 1128 subjects, randomly assigned to training (n=853) and validation (n=275) subsets with equal T2Dia prevalence of 11%. Results The risk stratification model based on 19 input features including increased RE showed good performance for predicting prevalent T2Dia (sensitivity=0.81, specificity=0.89). Post-hoc interpretation using the Shapley additive explanation (SHAP) method revealed that increased RE was the most important risk factor of T2Dia after other clinical factors for T2Dia (i.e., sex, BMI and average SpO2), ahead of standard PSG metrics (including the apnea-hypopnea index and oxygen desaturation index). Conclusion These findings suggested that the proportion of sleep time spent with increased RE automatically derived from MJM is a potential new reliable metric to predict prevalent T2Dia in patients with OSA. Support (if any)

Practical Knowledge Level Characteristics and Innovative Practice Achievement Transformation Mechanism of Vocational College Students

Cultivating talents with innovation ability is both the goal of higher vocational colleges and their responsibility given by the era. Focusing on improving the practical knowledge level of students means that students should be exposed to nature, society and problems through practical activities, thus providing the maximum space for them to think, explore, discover and innovate. As for existing studies on cultivating innovative ability of students and improving their practical knowledge level, they mainly have focused on ability evaluation and analysis of influencing factors, and few of them has involved analysis of their practical knowledge level characteristics. Therefore, this paper studied the characteristics and transformation mechanism of their innovative practice achievements. A multi-level comprehensive evaluation system was constructed, which aimed to evaluate the practical knowledge level of those students. The input dimension of quantized data of indexes was reduced in accordance with the principal component analysis (PCA). Then the quantized data of the low-dimensional indexes were mined by hierarchical clustering, and different training subsets were generated to participate in the evaluation model training. Based on the idea of clustering-prediction-integration, Gated Recurrent Unit (GRU) model was used for nonlinear integration of the output results of two-way Long Short-Term Memory (LSTM) model, which further fit the nonlinear characteristics in the quantized data of all indexes. By introducing Bootstrap-Data Envelopment Analysis (DEA) method, the transformation rate of innovative practice achievements of the students was collected and calculated in order to obtain a small calculation error. Experimental results verified the effectiveness of the proposed calculation method of the model.

A Nano-QSTR model to predict nano-cytotoxicity: an approach using human lung cells data

BackgroundThe widespread use of new engineered nanomaterials (ENMs) in industries such as cosmetics, electronics, and diagnostic nanodevices, has been revolutionizing our society. However, emerging studies suggest that ENMs present potentially toxic effects on the human lung. In this regard, we developed a machine learning (ML) nano-quantitative-structure-toxicity relationship (QSTR) model to predict the potential human lung nano-cytotoxicity induced by exposure to ENMs based on metal oxide nanoparticles.ResultsTree-based learning algorithms (e.g., decision tree (DT), random forest (RF), and extra-trees (ET)) were able to predict ENMs’ cytotoxic risk in an efficient, robust, and interpretable way. The best-ranked ET nano-QSTR model showed excellent statistical performance with R2 and Q2-based metrics of 0.95, 0.80, and 0.79 for training, internal validation, and external validation subsets, respectively. Several nano-descriptors linked to the core-type and surface coating reactivity properties were identified as the most relevant characteristics to predict human lung nano-cytotoxicity.ConclusionsThe proposed model suggests that a decrease in the ENMs diameter could significantly increase their potential ability to access lung subcellular compartments (e.g., mitochondria and nuclei), promoting strong nano-cytotoxicity and epithelial barrier dysfunction. Additionally, the presence of polyethylene glycol (PEG) as a surface coating could prevent the potential release of cytotoxic metal ions, promoting lung cytoprotection. Overall, the current work could pave the way for efficient decision-making, prediction, and mitigation of the potential occupational and environmental ENMs risks.

Comparing Machine Learning Techniques for Predictions of Motorway Segment Crash Risk Level

Motorways are typically the safest road environment in terms of injury crashes per million vehicle kilometres; however, given the high severity of crashes occurring therein, there is still space for road safety improvements. The objective of this study is to compare the classification performance of five machine learning techniques for predictions of crash risk levels of motorway segments. To that end, data on crash risk levels, driving behaviour metrics, and road geometry characteristics of 668 motorway segments were exploited. The utilized dataset was divided into training and test subsets, with a proportion of 75% and 25%, respectively. The training subset was used to train the models, whereas the test subset was used for the evaluation of their performance. The response variable of the models was the crash risk level of the considered motorway segments, while the predictors were various road design characteristics and naturalistic driving behaviour metrics. The techniques considered were Logistic Regression, Decision Tree, Random Forest, Support Vector Machine, and K-Nearest Neighbours. Among the five techniques, the Random Forest model achieved the best classification performance (overall accuracy: 89.3%, macro-averaged precision: 89.0%, macro-averaged recall: 88.4%, macro-averaged F1 score: 88.6%). Moreover, the Shapley additive explanations were calculated in order to assist with the interpretation of the model’s outcomes. The findings of this study are particularly useful as the Random Forest model could be used as a highly promising proactive road safety tool for identifying potentially hazardous motorway segments.

Spatial Prediction of Monthly Precipitation in Sulaimani Governorate using Artificial Neural Network Models

ANN modeling is used here to predict missing monthly precipitation data in one station of the eight weather stations network in Sulaimani Governorate. Eight models were developed, one for each station as for prediction. The accuracy of prediction obtain is excellent with correlation coefficients between the predicted and the measured values of monthly precipitation ranged from (90% to 97.2%). The eight ANN models are found after many trials for each station and those with the highest correlation coefficient were selected. All the ANN models are found to have a hyperbolic tangent and identity activation functions for the hidden and output layers respectively, with learning rate of (0.4) and momentum term of (0.9), but with different data set sub-division into training, testing and holdout data sub-sets, and different number of hidden nodes in the hidden layer. It is found that it is not necessary that the nearest station to the station under prediction has the highest effect; this may be attributed to the high differences in elevation between the stations. It can also found that the variance is not necessary has effect on the correlation coefficient obtained.

Development of a Prediction Model for Healthy Life Years Without Activity Limitation: National Cross-sectional Study.

In some countries, including Japan-the leading country in terms of longevity, life expectancy has been increasing; meanwhile, healthy life years have not kept pace, necessitating an effective health policy to narrow the gap. The aim of this study is to develop a prediction model for healthy life years without activity limitations and deploy the model in a health policy to prolong healthy life years. The Comprehensive Survey of Living Conditions, a cross-sectional national survey of Japan, was conducted by the Japanese Ministry of Health, Labour and Welfare in 2013, 2016, and 2019. The data from 1,537,773 responders were used for modelling using machine learning. All participants were randomly split into training (n=1,383,995, 90%,) and test (n=153,778, 10%) subsets. Extreme gradient boosting classifier was implemented. Activity limitations were set as the target. Age, sex, and 40 types of diseases or injuries were included as features. Healthy life years without activity limitations were calculated by incorporating the predicted prevalence rate of activity limitations in a life table. For the wide utility of the model in individuals, we developed an application tool for the model. In the groups without (n=1,329,901) and with (n=207,872) activity limitations, the median age was 47 (IQR 30-64) and 69 (IQR 54-80) years, respectively (P<.001); female sex comprised 51.3% (n=681,794) in the group without activity limitations and 56.9% (n=118,339) in the group with activity limitations (P<.001). A total of 42 features were included in the feature set. Age had the highest impact on model accuracy, followed by depression or other mental diseases; back pain; bone fracture; other neurological disorders, pain, or paralysis; stroke, cerebral hemorrhage, or infarction; arthritis; Parkinson disease; dementia; and other injuries or burns. The model exhibited high performance with an area under the receiver operating characteristic curve of 0.846 (95% CI 0.842-0.849) with exact calibration for the average probability and fraction of positives. The prediction results were consistent with the observed values of healthy life years for both sexes in each year (range of difference between predictive and observed values: -0.89 to 0.16 in male and 0.61 to 1.23 in female respondents). We applied the prediction model to a regional health policy to prolong healthy life years by adjusting the representative predictors to a target prevalence rate. Additionally, we presented the health condition without activity limitations index, followed by the application development for individual health promotion. The prediction model will enable national or regional governments to establish an effective health promotion policy for risk prevention at the population and individual levels to prolong healthy life years. Further investigation is needed to validate the model's adaptability to various ethnicities and, in particular, to countries where the population exhibits a short life span.

Development of a machine learning framework for radiation biomarker discovery and absorbed dose prediction.

Molecular radiation biomarkers are an emerging tool in radiation research with applications for cancer radiotherapy, radiation risk assessment, and even human space travel. However, biomarker screening in genome-wide expression datasets using conventional tools is time-consuming and underlies analyst (human) bias. Machine Learning (ML) methods can improve the sensitivity and specificity of biomarker identification, increase analytical speed, and avoid multicollinearity and human bias. To develop a resource-efficient ML framework for radiation biomarker discovery using gene expression data from irradiated normal tissues. Further, to identify biomarker panels predicting radiation dose with tissue specificity. A strategic search in the Gene Expression Omnibus database identified a transcriptomic dataset (GSE44762) for normal tissues radiation responses (murine kidney cortex and medulla) suited for biomarker discovery using an ML approach. The dataset was pre-processed in R and separated into train and test data subsets. High computational cost of Genetic Algorithm/k-Nearest Neighbor (GA/KNN) mandated optimization and 13 ML models were tested using the caret package in R. Biomarker performance was evaluated and visualized via Principal Component Analysis (PCA) and dose regression. The novelty of ML-identified biomarker panels was evaluated by literature search. Caret-based feature selection and ML methods vastly improved processing time over the GA approach. The KNN method yielded overall best performance values on train and test data and was implemented into the framework. The top-ranking genes were Cdkn1a, Gria3, Mdm2 and Plk2 in cortex, and Brf2, Ccng1, Cdkn1a, Ddit4l, and Gria3 in medulla. These candidates successfully categorized dose groups and tissues in PCA. Regression analysis showed that correlation between predicted and true dose was high with R2 of 0.97 and 0.99 for cortex and medulla, respectively. The caret framework is a powerful tool for radiation biomarker discovery optimizing performance with resource-efficiency for broad implementation in the field. The KNN-based approach identified Brf2, Ddit4l, and Gria3 mRNA as novel candidates that have been uncharacterized as radiation biomarkers to date. The biomarker panel showed good performance in dose and tissue separation and dose regression. Further training with larger cohorts is warranted to improve accuracy, especially for lower doses.

An HMM-DNN-Based System for the Detection and Classification of Low-Frequency Acoustic Signals from Baleen Whales, Earthquakes, and Air Guns off Chile

Marine passive acoustic monitoring can be used to study biological, geophysical, and anthropogenic phenomena in the ocean. The wide range of characteristics from geophysical, biological, and anthropogenic sounds sources makes the simultaneous automatic detection and classification of these sounds a significant challenge. Here, we propose a single Hidden Markov Model-based system with a Deep Neural Network (HMM-DNN) for the detection and classification of low-frequency biological (baleen whales), geophysical (earthquakes), and anthropogenic (air guns) sounds. Acoustic data were obtained from the Preparatory Commission for the Comprehensive Nuclear-Test-Ban Treaty Organization station off Juan Fernandez, Chile (station HA03) and annotated by an analyst (498 h of audio data containing 30,873 events from 19 different classes), and then divided into training (60%), testing (20%), and tuning (20%) subsets. Each audio frame was represented as an observation vector obtained through a filterbank-based spectral feature extraction procedure. The HMM-DNN training procedure was carried out discriminatively by setting HMM states as targets. A model with Gaussian Mixtures Models and HMM (HMM-GMM) was trained to obtain an initial set of HMM target states. Feature transformation based on Linear Discriminant Analysis and Maximum Likelihood Linear Transform was also incorporated. The HMM-DNN system displayed good capacity for correctly detecting and classifying events, with high event-level accuracy (84.46%), high weighted average sensitivity (84.46%), and high weighted average precision (89.54%). Event-level accuracy increased with higher event signal-to-noise ratios. Event-level metrics per class also showed that our HMM-DNN system generalized well for most classes but performances were best for classes that either had a high number of training exemplars (e.g., generally above 50) and/or were for classes of signals that had low variability in spectral features, duration, and energy levels. Fin whale and Antarctic blue whale song and air guns performed particularly well.

A Deep Learning Framework for the Prediction and Diagnosis of Ovarian Cancer in Pre- and Post-Menopausal Women.

Ovarian cancer ranks as the fifth leading cause of cancer-related mortality in women. Late-stage diagnosis (stages III and IV) is a major challenge due to the often vague and inconsistent initial symptoms. Current diagnostic methods, such as biomarkers, biopsy, and imaging tests, face limitations, including subjectivity, inter-observer variability, and extended testing times. This study proposes a novel convolutional neural network (CNN) algorithm for predicting and diagnosing ovarian cancer, addressing these limitations. In this paper, CNN was trained on a histopathological image dataset, divided into training and validation subsets and augmented before training. The model achieved a remarkable accuracy of 94%, with 95.12% of cancerous cases correctly identified and 93.02% of healthy cells accurately classified. The significance of this study lies in overcoming the challenges associated with the human expert examination, such as higher misclassification rates, inter-observer variability, and extended analysis times. This study presents a more accurate, efficient, and reliable approach to predicting and diagnosing ovarian cancer. Future research should explore recent advances in this field to enhance the effectiveness of the proposed method further.

M4Raw: A multi-contrast, multi-repetition, multi-channel MRI k-space dataset for low-field MRI research

Recently, low-field magnetic resonance imaging (MRI) has gained renewed interest to promote MRI accessibility and affordability worldwide. The presented M4Raw dataset aims to facilitate methodology development and reproducible research in this field. The dataset comprises multi-channel brain k-space data collected from 183 healthy volunteers using a 0.3 Tesla whole-body MRI system, and includes T1-weighted, T2-weighted, and fluid attenuated inversion recovery (FLAIR) images with in-plane resolution of ~1.2 mm and through-plane resolution of 5 mm. Importantly, each contrast contains multiple repetitions, which can be used individually or to form multi-repetition averaged images. After excluding motion-corrupted data, the partitioned training and validation subsets contain 1024 and 240 volumes, respectively. To demonstrate the potential utility of this dataset, we trained deep learning models for image denoising and parallel imaging tasks and compared their performance with traditional reconstruction methods. This M4Raw dataset will be valuable for the development of advanced data-driven methods specifically for low-field MRI. It can also serve as a benchmark dataset for general MRI reconstruction algorithms.

Application of TVD-Net for sagittal alignment and instability measurements in cervical spine radiographs.

Cervical spinal malalignment and instability are frequently occurring pathological conditions involving neck pain, radiculopathy, and myelopathy, often requiring surgical intervention. Accurate assessment of cervical alignment and instability are essential in surgical planning and evaluating postoperative outcomes. To automatically measure the sagittal alignment and instability of the cervical spine, we develop a novel deep-learning model by detecting landmarks on cervical radiographs. We introduce the transformer-embedded residual network (ResNet) as the network's core to automatically identify vertebral landmarks on digital and film-transformed cervical radiographs, and simultaneously measure the segmental Cobb angle and horizontal displacement. A Transformer Module was embedded into the latent space to extract the relationship between different vertebrae. Then a Rotating Attention Module was integrated between the encoder-decoder pairs to highlight the key points and maintain more details. Finally, a Vector Loss Module was proposed to restrain the orientation of the adjacent vertebra to reduce misdetection. All images were obtained from local hospital. Digital images were split into training, validation, and test subsets (896, 225, and 353 images, respectively). Likewise, film-transformed images were split into 404, 115, and 150 images, respectively. The results of the model were compared with manual measurements. Our deep learning algorithm achieved mean absolute difference (MAD) at a level of 2.20° and 2.33°, symmetric mean absolute error(SMAPE)at 16.63% and 19.35%, respectively, when measuring Cobb angle on digital images and films. On evaluating cervical instability, the diagnostic accuracy, sensitivity, specificity, precision, and F1-score evaluation metrics were calculated. The corresponding values were 89.80%, 86.49%, 90.68%, 71.11%, and 78.05% on digital images, and 90.00%, 83.78%, 91.15%, 75.61%, and 79.49% on film-transformed images, which were comparable to experienced surgeons. Visualization results demonstrated robust effectiveness in subjects with severe osteophytes or artifacts. This study presents a novel and efficient deep-learning model to assist landmarks identification and angulation and displacement calculation on lateral cervical spine radiographs, and demonstrates excellent accuracy in measuring cervical alignment and sound sensitivity and specificity in cervical instability diagnosis. It should be helpful for future research and clinical applications.

Thermal Bridges on Building Rooftops

Thermal Bridges on Building Rooftops (TBBR) is a multi-channel remote sensing dataset. It was recorded during six separate UAV fly-overs of the city center of Karlsruhe, Germany, and comprises a total of 926 high-resolution images with 6927 manually-provided thermal bridge annotations. Each image provides five channels: three color, one thermographic, and one computationally derived height map channel. The data is pre-split into training and test data subsets suitable for object detection and instance segmentation tasks. All data is organized and structured to comply with FAIR principles, i.e. being findable, accessible, interoperable, and reusable. It is publicly available and can be downloaded from the Zenodo data repository. This work provides a comprehensive data descriptor for the TBBR dataset to facilitate broad community uptake.

The system of self-consistent models for pesticide toxicity to Daphnia magna

Quantitative structure-property/activity relationships (QSPRs/QSARs) are a tool of modern theoretical and computational chemistry. The self-consistent model system is both a method to build up a group of QSPR/QSAR models and an approach to checking the reliability of these models. Here, a group of models of pesticide toxicity toward Daphnia magna for different distributions into training and test sub-sets is compared. This comparison is the basis for formulating the system of self-consistent models. The so-called index of the ideality of correlation (IIC) has been used to improve the above models’ predictive potential of pesticide toxicity. The predictive potential of the suggested models should be classified as high since the average value of the determination coefficient for the validation sets is 0.841, and the dispersion is 0.033 (on all five models). The best model (number 4) has an average determination coefficient of 0.89 for the external validation sets (related to all five splits).

Comparison of machine learning techniques in prediction of mortality following cardiac surgery: analysis of over 220000 patients from a large national database.

To perform a systematic comparison of in-hospital mortality risk prediction post-cardiac surgery, between the predominant scoring system-European System for Cardiac Operative Risk Evaluation (EuroSCORE) II, logistic regression (LR) retrained on the same variables and alternative machine learning techniques (ML)-random forest (RF), neural networks (NN), XGBoost and weighted support vector machine. Retrospective analyses of prospectively routinely collected data on adult patients undergoing cardiac surgery in the UK from January 2012 to March 2019. Data were temporally split 70:30 into training and validation subsets. Mortality prediction models were created using the 18 variables of EuroSCORE II. Comparisons of discrimination, calibration and clinical utility were then conducted. Changes in model performance, variable-importance over time and hospital/operation-based model performance were also reviewed. Of the 227087 adults who underwent cardiac surgery during the study period, there were 6258 deaths (2.76%). In the testing cohort, there was an improvement in discrimination [XGBoost (95% confidence interval (CI) area under the receiver operator curve (AUC), 0.834-0.834, F1 score, 0.276-0.280) and RF (95% CI AUC, 0.833-0.834, F1, 0.277-0.281)] compared with EuroSCORE II (95% CI AUC, 0.817-0.818, F1, 0.243-0.245). There was no significant improvement in calibration with ML and retrained-LR compared to EuroSCORE II. However, EuroSCORE II overestimated risk across all deciles of risk and over time. The calibration drift was lowest in NN, XGBoost and RF compared with EuroSCORE II. Decision curve analysis showed XGBoost and RF to have greater net benefit than EuroSCORE II. ML techniques showed some statistical improvements over retrained-LR and EuroSCORE II. The clinical impact of this improvement is modest at present. However the incorporation of additional risk factors in future studies may improve upon these findings and warrants further study.

Application of Artificial Intelligence to Predict CO2 Emissions: Critical Step towards Sustainable Environment

Prediction of carbon dioxide (CO2) emissions is a critical step towards a sustainable environment. In any country, increasing the amount of CO2 emissions is an indicator of the increase in environmental pollution. In this regard, the current study applied three powerful and effective artificial intelligence tools, namely, a feed-forward neural network (FFNN), an adaptive network-based fuzzy inference system (ANFIS) and long short-term memory (LSTM), to forecast the yearly amount of CO2 emissions in Saudi Arabia up to the year 2030. The data were collected from the “Our World in Data” website, which offers the measurements of the CO2 emissions from the years 1936 to 2020 for every country on the globe. However, this study is only concerned with the data related to Saudi Arabia. Due to some missing data, this study considered only the measurements in the years from 1954 to 2020. The 67 data samples were divided into 2 subsets for training and testing with the optimal ratio of 70:30, respectively. The effect of different input combinations on prediction accuracy was also studied. The inputs were combined to form six different groups to predict the next value of the CO2 emissions from the past values. The group of inputs that contained the past value in addition to the year as a temporal index was found to be the best one. For all the models, the performance accuracies were assessed using the root mean squared errors (RMSEs) and the coefficient of determination (R2). Every model was trained until the smallest RMSE of the testing data was reached throughout the entire training run. For the FFNN, ANFIS and LSTM, the averages of the RMSEs were 19.78, 20.89505 and 15.42295, respectively, while the averages of the R2 were found to be 0.990985, 0.98875 and 0.9945, respectively. Every model was applied individually to forecast the next value of the CO2 emission. To benefit from the powers of the three artificial intelligence (AI) tools, the final forecasted value was considered the average (ensemble) value of the three models’ outputs. To assess the forecasting accuracy, the ensemble was validated with a new measurement for the year 2021, and the calculated percentage error was found to be 6.8675% with an accuracy of 93.1325%, which implies that the model is highly accurate. Moreover, the resulting forecasting curve of the ensembled models showed that the rate of CO2 emissions in Saudi Arabia is expected to decrease from 9.4976 million tonnes per year based on the period 1954–2020 to 6.1707 million tonnes per year in the period 2020–2030. Therefore, the finding of this work could possibly help the policymakers in Saudi Arabia to take the correct and wise decisions regarding this issue not only for the near future but also for the far future.

PO-03-076 MACHINE LEARNING METHODS IDENTIFYING RISK FOR 30-DAY HEART FAILURE READMISSION USING DIAGNOSTICS PARAMETERS FROM IMPLANTED CARDIOVASCULAR DEVICES

Heart failure (HF) related diagnostics parameters measured daily from implantable cardioverter defibrillators (ICD) and cardiac resynchronization therapy defibrillators (CRTD) have shown to change before, during and after HF hospitalization events. We assessed the ability to identify patients at risk of a recurrent HF admission with different machine learning techniques leveraging statistical and temporal behavior of the cardiac compass parameters prior and during HF hospitalization events. We linked Optum® deidentified Electronic Health Record dataset during the period from 2007-2021 to the Medtronic CareLinkTM data warehouse with device based continuous diagnostic monitoring data. Patients with ICD/CRTD implants with intra-thoracic impedance diagnostic feature were included for this study. Retrospective analysis was completed identifying HF admissions with diagnostic parameters from 30-days prior admission day ending on discharge day. Analyzed events included such HF admissions within the 2 to 30 days for length of stay. Admission day parameters were evaluated with the following criteria: 7-days to 30-days activity, night heart rate and day heart rate ratios, plus 3-day impedance change. In addition, diagnostic data during hospitalization was summarized into range, median, mean, and standard deviation. Length of stay was also used as an input for the models. Two groups were created based on hospitalization events with or without 30-day readmission. Dataset was split into training and validation subsets. Multiple machine learning models were trained to classify events indicating those followed by a close HF readmission. Models’ results were compared in terms of sensitivity, specificity, and false positive rates on validation set. 1745 HF hospitalization events were assessed in this study, including 486 preceding a re-admission event within 30-days post discharge. The average length of stay on this cohort was 6.2 days. Seven machine learning models were trained using 28 features derived from diagnostic data and length of stay as input. Models’ sensitivity and specificity ranged from 0.11 to 0.84 and 0.89 to 0.93 respectively. Table 1 depicts area under the curve (AUC) for each trained model. Mean model accuracy on validation data was 0.84. A machine learning derived HF diagnostic criteria evaluating diagnostic parameters on HF admission day plus while in hospital may be useful to identify patients at higher risk of 30-days HF readmission at time of discharge.Tabled 1Table 1. AUC for each Classification ModelMODEL NAMEAUCLogistic Regression0.92Decision Tree0.82RUSBoosted Trees0.94Support Vector Machine0.83Bi-layered Neural Network0.86Three-layered Neural Network0.86Optimizable Neural Network0.94 Open table in a new tab