Development Of Machine Learning Models Research Articles

Development and comparison of machine learning models for in-vitro drug permeation prediction from microneedle patch.

The field of machine learning (ML) is advancing to a larger extent and finding its applications across numerous fields. ML has the potential to optimize the development process of microneedle patch by predicting the drug release pattern prior to its fabrication and production. The early predictions could not only assist the in-vitro and in-vivo experimentation of drug release but also conserve materials, reduce cost, and save time. In this work, we have used a dataset gleaned from the literature to train and evaluate different ML models, such as stacking regressor, artificial neural network (ANN) model, and voting regressor model. In this study, models were developed to improve prediction accuracy of the in-vitro drug release amount from the hydrogel-type microneedle patch and the in-vitro drug permeation amount through the micropores created by solid microneedles on the skin. We compared the performance of these models using various metrics, including R-squared score (R2 score), root mean squared error (RMSE), and mean absolute error (MAE). Voting regressor model performed better with drug permeation percentage as an outcome feature having RMSE value of 3.24. In comparison, stacking regressor have a RMSE value of 16.54, and ANN model has shown a RMSE value of 14. The value of permeation amount calculated from the predicted percentage is found to be more accurate with RMSE of 654.94 than direct amount prediction, having a RMSE of 669.69. All our models have performed far better than the previously developed model before this research, which had a RMSE of 4447.23. We then optimized voting regressor model's hyperparameter and cross validated its performance. Furthermore, it was deployed in a webapp using Flask framework, showing a way to develop an application to allow other users to easily predict drug permeation amount from the microneedle patch at a particular time period. This project demonstrates the potential of ML to facilitate the development of microneedle patch and other drug delivery systems.

Binary dataset for machine learning applications to tropical cyclone formation prediction

Applications of machine learning (ML) in atmospheric science have been rapidly growing. To facilitate the development of ML models for tropical cyclone (TC) research, this binary dataset contains a specific customization of the National Center for Environmental Prediction (NCEP)/final analysis (FNL) data, in which key environmental conditions relevant to TC formation are extracted for a range of lead times (0–72 hours) during 1999–2023. The dataset is designed as multi-channel images centered on TC formation locations, with a positive and negative directory structure that can be readily read from any ML applications or common data interface. With its standard structure, this dataset provides users with a unique opportunity to conduct ML application research on TC formation as well as related predictability at different forecast lead times.

Permeability modelling in a highly heterogeneous tight carbonate reservoir using comparative evaluating learning-based and fitting-based approaches

Permeability modelling is considered a complex task in reservoir characterization and a key component of reservoir simulation. A common method for permeability modelling involves performing static rock typing (SRT) using routine core analysis data and developing simple fitting-based mathematical relations that link permeability to reservoir rock porosity. In the case of carbonate reservoirs, which are associated with high heterogeneities, fitting-based approaches may fail due to porosity–permeability data scattering. Accurate modelling of permeability using petrophysical well log data seems more promising since they comprise a vast array of information about the intrinsic properties of the geological formations. Furthermore, well log data exhibit continuity throughout the entire reservoir interval, whereas core data are discrete and limited in availability and coverage. In this research work, porosity, permeability and log data of two oil wells from a tight carbonate reservoir were used to predict permeability at un-cored intervals. Machine learning (ML) and fitting models were used to develop predictive models. Then, the developed ML models were compared to exponential and statistical fitting modelling approaches. The integrated ML permeability model based on Random Forest method performed significantly superior to exponential and statistical fitting-based methods. Accordingly, for horizontal and vertical permeability of test samples, the Root Mean Squared Error (RMSE) values were 3.7 and 4.5 for well 2, and 1.7 and 0.86 for well 4, respectively. Hence, using log data, permeability modelling was improved as it incorporates more comprehensive reservoir rock physics. The outcomes of this reach work can be used to improve the distribution of both horizontal and vertical permeability in the 3D model for future dynamic reservoir simulations in such a complex and heterogeneous reservoir system.

Radiomics-based machine learning approach for the prediction of grade and stage in upper urinary tract urothelial carcinoma-a step towards virtual biopsy.

Upper tract urothelial carcinoma is a rare, aggressive lesion, with early detection a key to its management. This study aimed to utilise computed tomographic urogram data to develop machine learning models for predicting tumour grading and staging in upper urothelial tract carcinoma patients and to compare these predictions with histopathological diagnosis used as reference standards. Protocol-based computed tomographic urogram data from 106 patients were obtained and visualised in 3D. Digital segmentation of the tumours was conducted by extracting textural radiomics features. They were further classified using 11 predictive models. The predicted grades and stages were compared to the histopathology of radical nephroureterectomy specimens. Classifier models worked well in mining the radiomics data and delivered satisfactory predictive machine learning models. The MultiLayer Panel showed 84% sensitivity and 93% specificity while predicting upper tract urothelial carcinoma grades. The Logistic Regression model showed a sensitivity of 83% and a specificity of 76% while staging. Similarly, other classifier algorithms (e.g., Support Vector classifier) provided a highly accurate prediction while grading upper tract urothelial carcinoma compared to clinical features alone or ureteroscopic biopsy histopathology. Data mining tools could handle medical imaging datasets from small (<2cm) tumours for upper tract urothelial carcinoma. The radiomics-based machine learning algorithms provide a potential tool to model tumour grading and staging with implications for clinical practice and the upgradation of current paradigms in cancer diagnostics. Machine learning based on radiomics features can predict upper tract urothelial cancer grading and staging with significant improvement over ureteroscopic histopathology. The study showcased the prowess of such emerging tools in the set objectives with implications towards virtual biopsy.

Abstract PO2-01-14: Development of Artificial Intelligence-based Machine Learning Models for Predicting Survival In Hormone-Receptor-Positive/HER2-Negative Early Breast Cancer undergoing Neoadjuvant Chemotherapy

Abstract Background. In hormone-receptor positive (HoR)/HER2 negative early breast cancer (BC) multiple efforts have been made to predict disease recurrence and survival. Machine learning techniques have been used but few studies have looked into their applicability in predicting survival on the basis of clinico-pathological characteristics. The aim of this study was to evaluate a random survival forest model to predict the prognosis of this specific BC subtype. Methods. In this multicenter, retrospective study, patients who fulfilled the following inclusion criteria were included: (1) diagnosis of pathologically confirmed HoR-positive/HER2-negative invasive BC; (2) early or locally advanced, stage I-II-III (3) patients receiving neoadjuvant anthracycline and/or taxane-based chemotherapy, concurrently or sequentially; (4) patients undergoing surgery for primary BC. Survival endpoints were disease-free survival (DFS) and overall survival (OS). A random survival forest algorithm was used to develop the predictive model. Ten-fold cross-validation was performed. The C-index and the continuous rank probability score (CRPS) were used to evaluate the discrimination of the predictive model, and a ROC curve was used to evaluate model precision. A cut point analysis based on maximally selected rank statistics was conducted to evaluate the best cut-off in the out-of-bag (OOB) mortality that could maximize DFS and OS prediction. Variable importance was assessed using Breiman-Cutler permutation importance. Results. Overall, 572 patients with HoR-positive/HER2-negative early BC were included. At univariate analysis age, T stage, N stage, grading, ER, PR and Ki67 were found to be significantly associated with DFS and ER, PgR, HER2, pCR ypT and ypN retained statistical significance at multivariate analysis. ER, pCR, pathological T and N stage were found to be significantly associated with OS at univariate and multivariate analysis. The following variables were included in the final model: menopausal status, age, histology, grade, clinical T/N, ER/PgR, Ki67 and HER2, pathological complete response (pCR), ypT and ypN. For DFS, the cross-validated C-index was 0.68 (95% CI 0.63-0.73) and the OOB CRPS was 0.15, with a OOB performance error of 0.31. The AUC calculated at 60 months was 0.91 (95% CI 0.88-0.94). Out-of bag RSF-based risk scores for individual patients were calculated and an optimal cut-off of 22.58 was identified. The HR between high risk and low risk group was 3.08 (95% CI 2.16-4.39, p&amp;lt; 0.001). For OS, the cross-validated C-index was 0.66 (95% CI 0.60-0.71) and the OOB CRPS was 0.17, with a OOB performance error of 0.35. The AUC calculated at 60 months 0.91 (95% CI 0.88-0.95). An optimal cut-off of 21.54 was identified and the HR between low risk and high risk group was 2.50 (95% CI 1.73-3.61, p&amp;lt; 0.001). For DFS, the most important variables were cN (9.00), ypN (7.35), ER (7.21), ypT (5.38), PgR (4.11), pCR (4.01) and age (2.82) while for OS the most important variables were cN (7.88), ER (5.37), PgR (3.61), age (3.37), pCR (2.80), ypT (2.64) and ypN (2.60). Discussion. We analyzed the performance of RSF in the prediction of DFS and OS based on the contribution of clinico-pathological features commonly available at the baseline of the NCT and post-surgery. By selecting patients with HoR-positive HER2-negative disease, we were able to show which clinico-pathological features within a ML model have greater predictive importance. The model could be integrated with image-based tumor and radiology features to improve its predictive accuracy. This retrospective multicenter study suggests that the combination of easily accessible clinico-pathological features within a ML model may reliably predict DFS and OS in the context of HoR-positive/HER2-negative BCs. Citation Format: Luca Mastrantoni, Giovanna Garufi, Noemi Maliziola, Elena Di Monte, Giorgia Arcuri, Valentina Frescura, Angelachiara Rotondi, Giulia Giordano, Luisa Carbognin, Alessandra Fabi, Ida Paris, Gianluca Franceschini, Armando Orlandi, Antonella Palazzo, Giovanni Scambia, Giampaolo Tortora, Emilio Bria. Development of Artificial Intelligence-based Machine Learning Models for Predicting Survival In Hormone-Receptor-Positive/HER2-Negative Early Breast Cancer undergoing Neoadjuvant Chemotherapy [abstract]. In: Proceedings of the 2023 San Antonio Breast Cancer Symposium; 2023 Dec 5-9; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2024;84(9 Suppl):Abstract nr PO2-01-14.

Abstract PO1-07-01: Artificial intelligence-based prediction of Oncotype DX Score from whole slide images using human-interpretable features and breast biomarkers

Abstract Background The Oncotype DX Breast Recurrence Score assay (ODX) is a commonly used genomic test for patients with estrogen receptor (ER)-positive, HER2-negative, early-stage invasive breast cancer. While ODX predicts patients’ recurrence risk and benefit from chemotherapy, it is tissue and time-consuming, and expensive. Previous deep-learning or non-linear ODX prediction models achieved promising performance using whole-slide images (WSI) or with other covariates (e.g., ER, progesterone receptor (PR), HER-2, Ki-67 scores and tumor stage) but systematic quantification of the contribution of individual histological features to the ODX score remained challenging. Here, we extracted a rich set of human-interpretable features (HIFs) quantifying nuclear morphology and the distribution of cells and tissues in the tumor microenvironment. We used these HIFs, along with manually assessed ER, PR, HER-2 scores, and tumor stage, to predict ODX scores. We also explored the role of Ki-67 features in augmenting our model predictions Methods We developed machine learning models to extract cell, tissue, and nuclear HIFs from WSI stained with hematoxylin and eosin (H&amp;E) and immunohistochemistry (IHC) against Ki-67. These models were deployed on 353 H&amp;E WSI to extract 276 HIFs quantifying cell densities, tissue areas, relative cell counts, and nuclei size, shape and color. Hierarchical agglomerative clustering was applied to cluster these features into 11 groups. Univariate regression was performed to identify correlations between ODX scores and feature clusters. To predict ODX score using HIFs, a multivariable regression model was fitted to the data where regressors are representative features from each cluster, in addition to ER, PR, HER-2 scores and stage as covariates. The model was fitted on the training set (N=266 slides) and evaluated on the held-out test set (N=87 slides). To examine the utility of Ki-67 IHC features in ODX prediction, a Ki-67 model was deployed on a subset of 194 matched Ki-67 IHC WSI to extract count proportions and densities of Ki-67 positive cells in cancer epithelium and stroma. Results Significant positive correlations were identified between ODX scores and feature clusters corresponding to cancer cell density (p &amp;lt; 10-7), macrophage density (p &amp;lt; 10-4), immune cell density (p &amp;lt; 10-16), and variations in cancer nuclear size (p &amp;lt; 10-5) and color (p &amp;lt; 10-3). Significant negative correlations were observed between ODX scores and clusters related to fibroblast density (p &amp;lt; 10-3), and variations of non-cancer cell nuclear color (p = 0.02). Evaluation of our model’s ability to predict ODX scores revealed an association between predicted and observed scores (Pearson r = 0.74). The AUROC for model predictions of high/low classifications (with reference to a cutoff ODX score of 20) was 0.80. Ki-67 features were clustered together with H&amp;E cancer cell density features and showed a stronger correlation with ODX scores compared to the H&amp;E features alone. Combining Ki-67 cell density features with H&amp;E features led to increased performance of the multivariable regression model on the test set compared to H&amp;E features only (r = 0.62 vs. r = 0.59). However, a model combining Ki-67, H&amp;E and covariates performed similarly to a model with H&amp;E and covariates (r = 0.73), suggesting that Ki-67 features may not be needed for ODX prediction when H&amp;E and covariate information is available. Conclusions Our model predicts ODX recurrence scores with comparable performance to other black-box approaches using WSI and breast biomarker status and without the need for, Ki-67 features. Using readily available information, our model has the potential to be a convenient screening tool for patient stratification, which may lead to better patient care, lower costs, and faster treatment. Citation Format: Nhat Le, John Van Arnam, Christian Kirkup, Ylaine Gerardin, Michael Drage, Neha Khaitan, Anurag Sharma, Georgi Galev, Nicholas Indorf, Emily Sansevere, Ilan Wapinski, Stephanie Hennek, Gloria Zhang. Artificial intelligence-based prediction of Oncotype DX Score from whole slide images using human-interpretable features and breast biomarkers [abstract]. In: Proceedings of the 2023 San Antonio Breast Cancer Symposium; 2023 Dec 5-9; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2024;84(9 Suppl):Abstract nr PO1-07-01.

Abstract PO1-14-08: Real-world clinical genomics study of HR+/HER2- metastatic breast cancers treated by CDK4/6i plus endocrine therapies revealed a drug resistant tumor segment characterized by ER independence

Abstract Background CDK4/6 inhibitors (CDK4/6i) plus endocrine therapies (ET) are the standard-of-care for hormone receptor–positive/human epidermal receptor 2–negative metastatic breast cancer (HR+/HER2− mBC). However, drug resistance remains a major unmet need. Investigations of drug resistance mechanisms has been hampered by a dearth of tumor molecular profiling data from the post-treatment setting. To address this challenge, we have conducted a real-world clinical genomics study to better understand the molecular mechanism of CDK4/6i resistance as well as to stratify patients based on integrated multi-omics profiles. Methods We retrospectively analyzed a multi-omics dataset of 400 HR+/HER2- mBC patients who had received CDK4/6i plus ET and developed progressive disease (PD) from the de-identified Tempus database. Pre-treatment and post-progression biopsies were taken  1 year prior to starting the CDK4/6i treatment or following PD respectively. Tempus xT next-generation sequencing (DNA-seq of 648 genes) and RNA sequencing assays were performed on 427 tumor FFPE samples, including 200 pre-treatment, 227 post-progression and 26 longitudinal pairs. Results The median age of the patients was 57 (54.9-57.4) and median progression free survival (PFS) is 379 (341-433) days. Two genes were found to harbor a significant increase in genomic alteration frequencies (GAF) after adjusting for FDR at post-progression vs. pre-treatment – ESR1 (41.9% vs. 15%, p=5.4e-10), RB1 (13.2% vs. 3%, p=8.5e-05). ESR1 and RB1 also harbored high frequencies of acquired genomic alterations among 26 paired samples at 34.6% and 11.5% respectively. TP53 mutation at baseline was significantly associated with shorter PFS at baseline (p=4.23e-05, HR=2.081) and TP53 GAF significantly increased after PD (37% vs. 28.5%, p=0.039). BRCA1/2 pathogenic mutations (p=1.63e-04, HR=3.066), APOBEC mutation signature S13 (p=0.0125, HR=1.55) and CCNE1 gene expression (p=0.024, HR=1.46) were significantly associated with shorter PFS. APOBEC signature (p=0.0035) and CCNE1 expression (p=1.33e-06) also significantly increased post-progression. Among the top molecular features associated with longer PFS were markers of estrogen signaling such as PGR gene expression (p=6.76e-04, HR=0.565) and the Hallmark estrogen response signature (p=0.021, HR=0.679). Applying a multi-omics pattern recognition algorithm, we identified a molecularly distinct cluster (IC1) characterized by down-regulation of estrogen signaling. IC1 is significantly associated with shorter PFS (p=3.72e-05, HR=0.22) and increased from 4% pre-treatment to 23% post-progression (p=7.3e-08). Further, IC1 is strongly enriched in markers previously implicated in CDK4/6i resistance including CCNE1 expression, RB1 mutation and MYC/E2F activation. We then developed machine learning models to predict gene-level dependency trained on cancer cell line expression and CRISPR-KO screen data. These models predicted decreased dependency on ESR1 and CDK4 and increased dependency on CDK2 in IC1, strengthening the association between ER independence and CDK4/6i resistance. Conclusions Our real-world clinical genomics study identified a comprehensive list of biomarkers associated with resistance to CDK4/6i plus ET and estimated patient prevalence for these markers in the post-treatment setting. Integrated and machine-learning analyses identified a subset of aggressive tumors with estrogen independence characteristics that are implicated in CDK4/6i resistance and suggested new therapeutic strategies. Citation Format: Zhengyan Kan, Ji Wen, Jennifer Webster, Vinicius Bonato, Whijae Roh, Xinmeng Jasmine Mu, Paul Rejto, Jadwiga Bienkowska. Real-world clinical genomics study of HR+/HER2- metastatic breast cancers treated by CDK4/6i plus endocrine therapies revealed a drug resistant tumor segment characterized by ER independence [abstract]. In: Proceedings of the 2023 San Antonio Breast Cancer Symposium; 2023 Dec 5-9; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2024;84(9 Suppl):Abstract nr PO1-14-08.

Developing a Machine-Learning Predictive Model for Retention of Posterior Cruciate Ligament in Patients Undergoing Total Knee Arthroplasty.

Predicting whether the posterior cruciate ligament (PCL) should be preserved during total knee arthroplasty (TKA) procedures is a complex task in the preoperative phase. The choice to either retain or excise the PCL has a substantial effect on the surgical outcomes and biomechanical integrity of the knee joint after the operation. To enhance surgeons' ability to predict the removal and retention of the PCL in patients before TKA, we developed machine learning models. We also identified significant feature factors that contribute to accurate predictions during this process. Patients' data on TKA continuously performed by a single surgeon who had intended initially to undergo implantation of cruciate-retaining (CR) prostheses was collected. During the sacrifice of PCL, we utilized anterior-stabilized (AS) tibial bearings. The dataset was split into CR and AS categories to form distinct groups. Relevant information regarding age, gender, body mass index (BMI), the affected side, and preoperative diagnosis was extracted by reviewing the medical records of the patients. To ensure the authenticity of the research, an initial step involved capturing X-ray images before the surgery. These images were then analyzed to determine the height of the medial condyle (MMH) and lateral condyle (LMH), as well as the ratios between MLW and MMH and MLW and LMH. Additionally, the insall-salvati index (ISI) was calculated, and the severity of any varus or valgus deformities was assessed. Eight machine-learning methods were developed to predict the retention of PCL in TKA. Risk factor analysis was performed using the SHApley Additive exPlanations method. A total of 307 knee joints from 266 patients were included, among which there were 254 females and 53 males. A stratified random sampling technique was used to split patients in a 70:30 ratio into a training dataset and a testing dataset. Eight machine-learning models were trained using data feeding. Except for the AUC of the LGBM Classifier, which is 0.70, the AUCs of other machine learning models are all lower than 0.70. In importance-based analysis, ISI, MMH, LMH, deformity, and age were confirmed as important predictive factors for PCL retention in operations. The LGBM Classifier model achieved the best performance in predicting PCL retention in TKA. Among the potential risk factors, ISI, MMH, LMH, and deformity played essential roles in the prediction of PCL retention.

Developing machine learning models with metaheuristic algorithms for droplet size prediction in a microfluidic microchannel

Droplet-based microfluidics holds promise for advanced biomedical applications due to its enhanced accuracy, throughput, rapid reaction time, and minimal reagent consumption. However, achieving droplets with the desired radius often entails costly design iterations. In this paper, we propose a novel approach that combines machine learning algorithms with metaheuristic algorithms to leverage their respective strengths in speed and accuracy. Specifically, we demonstrate the effectiveness of the Rain Optimization Algorithm (ROA) in enhancing the accuracy of five machine learning models. Additionally, we introduce a hybrid model, optimized by ROA, for predicting droplet generation in a cross-junction microfluidic channel, achieving a coefficient of determination (R2) of 0.937, reflecting a 3.6 % increase in accuracy. We employ symbolic regression to validate the hybrid model, yielding an R2 value of 0.987, indicating strong agreement with real data. Furthermore, we utilize the SHAP method in Explainable AI to interpret the hybrid model's performance. Notably, the proposed hybrid model requires only four seconds to solve the prediction problem, significantly reducing computational time compared to the numerical model and experimental method. This approach facilitates the study of each input parameter's impact on droplet size prediction, offering otherwise challenging and costly insights to obtain experimentally.

A Study on the Predictions of Wave Breaker Index in a Gravel Beach Using Linear Machine Learning Model

To date, numerous empirical formulas have been proposed through hydraulic model experiments to predict the wave breaker index, including wave height and depth of wave breaking, due to the inherent complexity of generation mechanisms. Unfortunately, research on the characteristics of wave breaking and the prediction of the wave breaker index for gravel beaches has been limited. This study aims to forecast the wave breaker index for gravel beaches using representative linear-based machine learning techniques known for their high predictive performance in regression or classification problems across various research fields. Initially, the applicability of existing empirical formulas for wave breaker indices to gravel seabeds was assessed. Various linear-based machine learning algorithms were then employed to build prediction models, aiming to overcome the limitations of existing empirical formulas in predicting wave breaker indices for gravel seabeds. Among the developed machine learning models, a new calculation formula for easily computable wave breaker indices based on the model was proposed, demonstrating high predictive performance for wave height and depth of wave breaking on gravel beaches. The study validated the predictive capabilities of the proposed wave breaker indices through hydraulic model experiments and compared them with existing empirical formulas. Despite its simplicity as a polynomial, the newly proposed empirical formula for wave breaking indices in this study exhibited exceptional predictive performance for gravel beaches.

Machine learning-based screening of fuel properties for SI and CI engines using a hybrid group extraction method

The design of high-performance fuels is crucial for achieving clean and efficient combustion of engines. In the current study, a framework for machine learning (ML) model-based fuel design is presented to identify compounds with desired properties for internal combustion engine (ICE) applications. Descriptors computed from newly proposed structural and positional-based group extraction method in this study are used as input in ML models due to the simplicity and computational-efficiency. The ML models were trained and validated using publically available experimental data for up to 1135 compounds. The results demonstrated a high level of predictive accuracy, with R2 values generally exceeding 0.99 for 11 physicochemical properties including melting point, boiling point, enthalpy of vaporization, surface tension, dynamic viscosity, low heating value, liquid density, yield sooting index, cetane number, research octane number and motor octane number. Furthermore, by employing the developed ML models to predict new data, a Fuel Property Database containing 1135 fuels with 12 fuel properties is established, enhanced by an interface for a user-friendly Fuel Physicochemical Properties Prediction Tool that facilitates swift property predictions and database expansion. The Pearson correlation coefficient (PCC) approach is subsequently employed to assess the correlation between descriptors and predicted or experimental properties. The strong agreement between the correlations affirms the effective predictive performance of the ML models within the calibrated data range. Additionally, a PCC analysis is conducted to reveal the individual effects of influential factors and their interactions, highlighting significant features like molecular weight size, degree of branching, and functional group content that influence physical and chemical properties, thereby providing valuable insights for fuel design and optimization. Finally, a initial comprehensive data-driven screening was carried out to identify potential fuel candidates that meet the key property limits for combustion applications in both spark-ignition (SI) and compression-ignition (CI) engines.

An open-access simultaneous electrocardiogram and phonocardiogram database.

The EPHNOGRAM project aimed to develop a low-cost, low-power device for simultaneous ECG and PCG recording, with additional channels for environmental audio to enhance PCG through active noise cancellation. The objective was to study multimodal electro-mechanical activities of the heart, offering insights into the differences and synergies between these modalities during various cardiac activity levels.&#xD;&#xD;Approach: We developed and tested several hardware prototypes of a simultaneous ECG-PCG acquisition device. Using this technology, we collected simultaneous ECG and PCG data from 24 healthy adults during different physical activities, including resting, walking, running, and stationary biking, in an indoor fitness center. The data were annotated using a robust software that we developed for detecting ECG R-peaks and PCG S1 and S2 components, and overseen by a human expert. We also developed machine learning models using ECG-based, PCG-based, and joint ECG-PCG features, like R-R and S1-S2 intervals, to classify physical activities and analyze electro-mechanical dynamics.&#xD;&#xD;Main Results: The results show a significant coupling between ECG and PCG components, especially during high-intensity exercise. Notable micro-variations in S2-based heart rate show differences in the heart's electrical and mechanical functions. The Lomb-Scargle periodogram and approximate entropy analyses confirm the higher volatility of S2-based heart rate compared to ECG-based heart rate. Correlation analysis shows stronger coupling between R-R and R-S1 intervals during high-intensity activities. Hybrid ECG-PCG features, like the R-S2 interval, were identified as more informative for physical activity classification through mRMR feature selection and SHAP value analysis.&#xD;&#xD;Significance: The EPHNOGRAM database, is available on PhysioNet. The database enhances our understanding of cardiac function, enabling future studies on the heart's mechanical and electrical interrelationships. The results of this study can contribute to improved cardiac condition diagnoses. Additionally, the designed hardware has the potential for integration into wearable devices and the development of multimodal stress test technologies.

Development and validation of machine learning models and nomograms for predicting the surgical difficulty of laparoscopic resection in rectal cancer

BackgroundThe objective of this study is to develop and validate a machine learning (ML) prediction model for the assessment of laparoscopic total mesorectal excision (LaTME) surgery difficulty, as well as to identify independent risk factors that influence surgical difficulty. Establishing a nomogram aims to assist clinical practitioners in formulating more effective surgical plans before the procedure.MethodsThis study included 186 patients with rectal cancer who underwent LaTME from January 2018 to December 2020. They were divided into a training cohort (n = 131) versus a validation cohort (n = 55). The difficulty of LaTME was defined based on Escal’s et al. scoring criteria with modifications. We utilized Lasso regression to screen the preoperative clinical characteristic variables and intraoperative information most relevant to surgical difficulty for the development and validation of four ML models: logistic regression (LR), support vector machine (SVM), random forest (RF), and decision tree (DT). The performance of the model was assessed based on the area under the receiver operating characteristic curve(AUC), sensitivity, specificity, and accuracy. Logistic regression-based column-line plots were created to visualize the predictive model. Consistency statistics (C-statistic) and calibration curves were used to discriminate and calibrate the nomogram, respectively.ResultsIn the validation cohort, all four ML models demonstrate good performance: SVM AUC = 0.987, RF AUC = 0.953, LR AUC = 0.950, and DT AUC = 0.904. To enhance visual evaluation, a logistic regression-based nomogram has been established. Predictive factors included in the nomogram are body mass index (BMI), distance between the tumor to the dentate line ≤ 10 cm, radiodensity of visceral adipose tissue (VAT), area of subcutaneous adipose tissue (SAT), tumor diameter >3 cm, and comorbid hypertension.ConclusionIn this study, four ML models based on intraoperative and preoperative risk factors and a nomogram based on logistic regression may be of help to surgeons in evaluating the surgical difficulty before operation and adopting appropriate responses and surgical protocols.

Asynchronous Federated Learning with Grey Wolf Optimization for the Heterogeneity IoT Devices

The Internet of Things (IoT) creates new options for real-time data collection and Machine Learning (ML) model development. Nonetheless, it's feasible that a particular IoT device does not have sufficient computational power to train and implement a complete learning model. However, there are significant communication costs and data security and privacy concerns associated with sending actual data to a centralised server with a lot of computational power. Federated Learning (FL) is a potential way to train ML models using low-powered devices and Edge Servers (ES) since it is a distributed ML architecture. However, the vast majority of the works in existence make the unsustainable assumption of a synchronized parameters update manner with similar IoT nodes and reliable communications networks. To increase training efficiency and accelerate the speed for heterogeneous IoT devices in an unreliable network environment, we designed an Asynchronous Federated Learning strategy with Grey Wolf Optimization (AFL-GWO) in this research. In particular, we develop a Lightweight Node Selection (LNS) technique and propose an AFL-GWO model to efficiently complete learning tasks. To ensure that diverse IoT nodes with varying computational capabilities and network connectivity are represented in the global learning aggregate, the proposed technique makes node selections on an iterative basis. We show through extensive experiments that our suggested AFL-GWO system outperforms the state-of-the-art techniques on identically and independently distributed (IID) and non-IID data distribution in a variety of contexts.

Actual evapotranspiration and energy balance estimation from vineyards using micro-meteorological data and machine learning modeling

Actual evapotranspiration (ETa) can be commonly estimated using numerical models based on i) weather and plant-based parameters, ii) from remotely sensed data and energy balance algorithms, and lately, iii) through the development and implementation of machine learning (ML) modeling techniques. In this work, supervised ML models were developed from a vineyard located in Talca, Chile, (i) to estimate actual evapotranspiration (ETa) (Model 1; M1) using the micrometeorological approach [Eddy Covariance; EC; sensible (H), latent (LE), soil heat fluxes (G) and net radiation (Rn)] and data from an automatic meteorological station (AMS) in reference conditions as ground-truth (inputs); (ii) to estimate energy balance components (Model 2; M2) from AMS data (inputs) and EC energy balance data as targets; (iii) to estimate ETa from the EC’s measured ETa data as target and thermal time data (degree hours; DH) calculated from air temperature with a base of 5 °C increments from 5 – 45 °C as inputs (Model 3; M3) and iv) to estimate energy balance components (targets from EC) from the same inputs of Model 3 (Model 4; M4). Results showed that the developed ML models had high accuracy and performance with no signs of over or under-fitting with a high correlation (R) and slope (b) close to unity (M1; R=0.94; b=0.89; M2; R=0.97; b=0.93; M3; R=0.97; b=0.89–0.95; M4; R=0.98; b=0.97). Furthermore, models were directly deployed over another vineyard located 22 km West of the modeled vineyard at 60 m lower over the sea level with significant performances and R values (R = 0.64–0.87; b = 0.66–1.00 for M1 to M4, respectively). These models could be used for precision irrigation to increase water use efficiency and better control canopy vigor, balance fruit and vegetative components, and ultimately improve berry and wine quality traits.

Towards inclusivity in AI: A comparative study of cognitive engagement between marginalized female students and peers

AbstractThis study addresses the need for inclusive AI education by focusing on marginalized female students who historically lack access to learning opportunities in computing. It applies the theoretical framework of intersectionality to understand how gender, race and ethnicity intersect to shape these students' learning experiences and outcomes. Specifically, this study investigated 27 high‐school students' cognitive engagement in machine learning practices. We conducted the Wilcoxon–Mann–Whitney test to explore differences in cognitive engagement between marginalized female students and their peers, employed comparative content analysis to delve into significant differences and analysed interview data thematically to gain deeper insights into students' machine learning model development processes. The findings indicated that, when engaging in machine learning practices requiring drawing diverse cultural perspectives, marginalized female students demonstrated significantly higher performance compared to their peers. In particular, marginalized female students exhibited strengths in holistic language analysis, paying attention to writers' intentions and recognizing cultural nuances in language. This study suggests that integrating language analysis and machine learning across subjects has the potential to empower marginalized female students and amplify their perspectives. Furthermore, it calls for a strengths‐based approach to reshape the narrative of underrepresentation and promote equitable participation in machine learning and AI. Practitioner notesWhat is already known about this topic Female students, particularly those from underrepresented groups such as African American and Latina students, often experience low levels of cognitive engagement in computing. Marginalized female students possess unique strengths that, when nurtured, have the potential to not only transform their own learning experiences but also contribute to the advancement of the computing field. It is critical to empower marginalized female students in K‐12 AI (ie, a subfield of computing) education, seeking to bridge the gender and racial disparity in AI. What this paper adds Marginalized female students outperformed their peers in responding to machine learning questions related to feature analysis and feature distribution interpretation. When responding to these questions, they demonstrated a holistic approach to analysing language by considering interactions between features and writers' intentions. They drew on knowledge about how language was used to convey meaning in different cultural contexts. Implications for practice and/or policy Educators should design learning environments that encourage students to draw upon their cultural backgrounds, linguistic insights and diverse experiences to enhance their engagement and performance in AI‐related activities. Educators should strategically integrate language analysis and machine learning across different subjects to create interdisciplinary learning experiences that support students' exploration of the interplay among language, culture and AI. Educational institutions and policy initiatives should adopt a strengths‐based approach that focuses on empowering marginalized female students by acknowledging their inherent abilities and diverse backgrounds.

Machine-learning-based models to predict cardiovascular risk using oculomics and clinic variables in KNHANES

BackgroundRecent researches have found a strong correlation between the triglyceride-glucose (TyG) index or the atherogenic index of plasma (AIP) and cardiovascular disease (CVD) risk. However, there is a lack of research on non-invasive and rapid prediction of cardiovascular risk. We aimed to develop and validate a machine-learning model for predicting cardiovascular risk based on variables encompassing clinical questionnaires and oculomics.MethodsWe collected data from the Korean National Health and Nutrition Examination Survey (KNHANES). The training dataset (80% from the year 2008 to 2011 KNHANES) was used for machine learning model development, with internal validation using the remaining 20%. An external validation dataset from the year 2012 assessed the model’s predictive capacity for TyG-index or AIP in new cases. We included 32122 participants in the final dataset. Machine learning models used 25 algorithms were trained on oculomics measurements and clinical questionnaires to predict the range of TyG-index and AIP. The area under the receiver operating characteristic curve (AUC), accuracy, precision, recall, and F1 score were used to evaluate the performance of our machine learning models.ResultsBased on large-scale cohort studies, we determined TyG-index cut-off points at 8.0, 8.75 (upper one-third values), 8.93 (upper one-fourth values), and AIP cut-offs at 0.318, 0.34. Values surpassing these thresholds indicated elevated cardiovascular risk. The best-performing algorithm revealed TyG-index cut-offs at 8.0, 8.75, and 8.93 with internal validation AUCs of 0.812, 0.873, and 0.911, respectively. External validation AUCs were 0.809, 0.863, and 0.901. For AIP at 0.34, internal and external validation achieved similar AUCs of 0.849 and 0.842. Slightly lower performance was seen for the 0.318 cut-off, with AUCs of 0.844 and 0.836. Significant gender-based variations were noted for TyG-index at 8 (male AUC=0.832, female AUC=0.790) and 8.75 (male AUC=0.874, female AUC=0.862) and AIP at 0.318 (male AUC=0.853, female AUC=0.825) and 0.34 (male AUC=0.858, female AUC=0.831). Gender similarity in AUC (male AUC=0.907 versus female AUC=0.906) was observed only when the TyG-index cut-off point equals 8.93.ConclusionWe have established a simple and effective non-invasive machine learning model that has good clinical value for predicting cardiovascular risk in the general population.

Predictive model for a product without history using LightGBM. Pricing model for a new product

The article focuses on developing a predictive product pricing model using LightGBM. Also, the goal was to adapt the LightGBM method for regression problems and, especially, in the problems of forecasting the price of a product without history, that is, with a cold start.The article contains the necessary concepts to understand the working principles of the light gradient boosting machine, such as decision trees, boosting, random forests, gradient descent, GBM (Gradient Boosting Machine), GBDT (Gradient Boosting Decision Trees). The article provides detailed insights into the algorithms used for identifying split points, with a focus on the histogram-based approach.LightGBM enhances the gradient boosting algorithm by introducing an automated feature selection mechanism and giving special attention to boosting instances characterized by more substantial gradients. This can lead to significantly faster training and improved prediction performance. The Gradient-based One-Side Sampling (GOSS) and Exclusive Feature Bundling (EFB) techniques used as enhancements to LightGBM are vividly described. The article presents the algorithms for both techniques and the complete LightGBM algorithm.This work contains an experimental result. To test the lightGBM, a real dataset of one Japanese C2C marketplace from the Kaggle site was taken. In the practical part, a performance comparison between LightGBM and XGBoost (Extreme Gradient Boosting Machine) was performed. As a result, only a slight increase in estimation performance (RMSE, MAE, R-squard) was found by applying LightGBM over XGBoost, however, there exists a notable contrast in the training procedure’s time efficiency. LightGBM exhibits an almost threefold increase in speed compared to XGBoost, making it a superior choice for handling extensive datasets.This article is dedicated to the development and implementation of machine learning models for product pricing using LightGBM. The incorporation of automatic feature selection, a focus on highgradient examples, and techniques like GOSS and EFB demonstrate the model’s versatility and efficiency. Such predictive models will help companies improve their pricing models for a new product. The speed of obtaining a forecast for each element of the database is extremely relevant at a time of rapid data accumulation.

Biomolecular Adsorption on Nanomaterials: Combining Molecular Simulations with Machine Learning.

Adsorption free energies of 32 small biomolecules (amino acids side chains, fragments of lipids, and sugar molecules) on 33 different nanomaterials, computed by the molecular dynamics - metadynamics methodology, have been analyzed using statistical machine learning approaches. Multiple unsupervised learning algorithms (principal component analysis, agglomerative clustering, and K-means) as well as supervised linear and nonlinear regression algorithms (linear regression, AdaBoost ensemble learning, artificial neural network) have been applied. As a result, a small set of biomolecules has been identified, knowledge of adsorption free energies of which to a specific nanomaterial can be used to predict, within the developed machine learning model, adsorption free energies of other biomolecules. Furthermore, the methodology of grouping of nanomaterials according to their interactions with biomolecules has been presented.

An interpretable framework for modeling global solar radiation using tree-based ensemble machine learning and Shapley additive explanations methods

Machine learning techniques provide an effective and cost-efficient solution for estimating solar radiation for solar energy utilization. However, the reported machine learning-based solar radiation models fail to offer comprehensive explanations for their outputs. Therefore, this study aims to tackle this issue by developing machine learning models that are both accurate and interpretable. To achieve the objective, this study evaluated the performance of tree-based ensemble algorithms, using optimized combinations of model input parameters for different climate zones in China. The results showed that the extreme gradient boosting (XGBoost) models demonstrated the highest overall accuracy, model stability, and generalization ability. At the national scale, the developed XGBoost models yielded an average R2, MAE, and RMSE of 0.939, 1.226 MJ/m2, and 1.663 MJ/m2, respectively, showing significant improvements of 2.13–27.78% in RMSE compared to recently reported models. Most importantly, the state-of-the-art SHapley Additive exPlanations (SHAP) technique was integrated with the developed XGBoost models to enhance model interpretability in terms of global and local feature importance, as well as the interaction effects between model features. The results of the SHAP value analysis demonstrated the robustness of sunshine duration in modeling global solar radiation, revealing thresholds where its values undergo a shift from negative to positive effects on model output. SHAP interaction values illustrated the interaction effects among features in the developed solar radiation model, uncovering the model's complex non-linear relationships. Additionally, this study provided explanations for individual instances based on the SHAP method. Overall, this study provided an accurate, reliable, and transparent machine learning model and an enlightening framework for modeling global solar radiation at sites without observations.