Development Of Machine Learning Models Research Articles

Precision forecasting of fertilizer components’ concentrations in mixed variable-rate fertigation through machine learning

Accurate monitoring of fertilizer concentration and variable irrigation components is crucial for achieving precision irrigation through variable-rate fertigation. However, technological and cost constraints pose challenges in effectively managing mixed variable-rate fertigation. This study addresses these challenges by integrating machine learning (ML) with easily monitored physical parameters, such as electrical conductivity (EC), pH, and temperature, to predict fertilizer solution components and concentrations in mixed variable-rate fertigation. Cubic spline interpolation (CSI) was used to enhance the training dataset. Six ML algorithms—Multivariate Linear Regression (MLR), Support Vector Machines (SVM), K-Nearest Neighbors (KNN), Extremely Randomized Trees (ETs), Multilayer Perceptron (MLP), and Extreme Gradient Boosting (XGB)—were used to develop prediction models, with their performance evaluated using the coefficient of determination (R2), root mean square error (RMSE), and Akaike information criterion (AIC). The Extended Fourier Amplitude Sensitivity Test (EFAST) assessed the sensitivity of the ML models to physical parameters. All of the ML models, except for MLR, particularly SVM, demonstrated superior performance in predicting fertilizer solution components’ concentrations, with R2 values between 0.989 and 0.997 and RMSE values between 0.089 and 0.210. The CSI significantly enhanced model performance, resulting in larger R2 values and smaller RMSE values. The AIC results and sensitivity analysis confirmed the exceptional performance of SVM, emphasizing its suitability for predicting fertilizer components using easily measured physical parameters. The developed ML models offer valuable insights for decision-makers managing irrigation and fertilization in mixed variable-rate fertigation.

Advancing hydrogen storage predictions in metal-organic frameworks: A comparative study of LightGBM and random forest models with data enhancement

The escalating consumption of fossil fuels has given rise to a substantial upsurge in greenhouse gas concentrations and global temperatures, which, in turn, has triggered severe climate-related consequences. The critical imperative to reduce CO2 emissions and combat global warming has spurred extensive investigations into clean energy alternatives, with hydrogen emerging as a compelling zero-emission energy source. As a pivotal component of clean energy strategies, hydrogen requires designing compact, lightweight, and efficient storage systems. This study focuses on the development and evaluation of machine learning models for predicting the efficiency of Metal-Organic Frameworks (MOFs) in hydrogen storage, a key aspect of advancing clean energy technologies. MOFs, a class of nanoporous materials, show remarkable potential for hydrogen storage due to their high surface area and porosity. However, selecting the most suitable MOF for this application from a vast array of possible structures is a daunting task. In this context, machine learning algorithms offer an efficient alternative for predicting MOF suitability by considering their structural and chemical properties. We used ensemble learning methods, specifically Light Gradient Boosting Machine (LightGBM) and Random Forest (RF), to predict hydrogen uptake of MOFs based on a dataset of 219 experimentally tested samples. Two modeling scenarios were considered: one using the entire dataset, and the other involving strategic data pre-processing, including outlier removal and feature engineering. The results demonstrate that the measures taken to refine the dataset significantly enhance the predictive performance of the developed models, reducing prediction errors and improving overall goodness of fit. Specifically, the Mean Absolute Error (MAE) values for both the LightGBM and random forest models were reduced from 0.48 and 0.94, respectively, to 0.16 for both models, and the coefficients of determination (R2) increased substantially from 0.84 and 0.72 to 0.95, in both cases. Moreover, feature importance analysis unveiled that pressure-related features make the most significant contributions to the formation of tree ensembles during the model training process. A parametric sensitivity analysis was conducted revealing that H2 uptake in MOFs is most sensitive to changes in adsorption enthalpy, followed by surface area and temperature, while showing lower sensitivity to variations in pressure, consistent with established literature. These results underscore the pivotal role of data enhancement methods in refining machine learning models and can be instrumental in accelerating the development and optimization of MOF materials for clean energy applications.

Identification of patients' smoking status using an explainable AI approach: a Danish electronic health records case study.

Smoking is a critical risk factor responsible for over eight million annual deaths worldwide. It is essential to obtain information on smoking habits to advance research and implement preventive measures such as screening of high-risk individuals. In most countries, including Denmark, smoking habits are not systematically recorded and at best documented within unstructured free-text segments of electronic health records (EHRs). This would require researchers and clinicians to manually navigate through extensive amounts of unstructured data, which is one of the main reasons that smoking habits are rarely integrated into larger studies. Our aim is to develop machine learning models to classify patients' smoking status from their EHRs. This study proposes an efficient natural language processing (NLP) pipeline capable of classifying patients' smoking status and providing explanations for the decisions. The proposed NLP pipeline comprises four distinct components, which are; (1) considering preprocessing techniques to address abbreviations, punctuation, and other textual irregularities, (2) four cutting-edge feature extraction techniques, i.e. Embedding, BERT, Word2Vec, and Count Vectorizer, employed to extract the optimal features, (3) utilization of a Stacking-based Ensemble (SE) model and a Convolutional Long Short-Term Memory Neural Network (CNN-LSTM) for the identification of smoking status, and (4) application of a local interpretable model-agnostic explanation to explain the decisions rendered by the detection models. The EHRs of 23,132 patients with suspected lung cancer were collected from the Region of Southern Denmark during the period 1/1/2009-31/12/2018. A medical professional annotated the data into 'Smoker' and 'Non-Smoker' with further classifications as 'Active-Smoker', 'Former-Smoker', and 'Never-Smoker'. Subsequently, the annotated dataset was used for the development of binary and multiclass classification models. An extensive comparison was conducted of the detection performance across various model architectures. The results of experimental validation confirm the consistency among the models. However, for binary classification, BERT method with CNN-LSTM architecture outperformed other models by achieving precision, recall, and F1-scores between 97% and 99% for both Never-Smokers and Active-Smokers. In multiclass classification, the Embedding technique with CNN-LSTM architecture yielded the most favorable results in class-specific evaluations, with equal performance measures of 97% for Never-Smoker and measures in the range of 86 to 89% for Active-Smoker and 91-92% for Never-Smoker. Our proposed NLP pipeline achieved a high level of classification performance. In addition, we presented the explanation of the decision made by the best performing detection model. Future work will expand the model's capabilities to analyze longer notes and a broader range of categories to maximize its utility in further research and screening applications.

AntiMicro.ai: An Artificial Intelligence powered web application for predicting antibacterial/antifungal susceptibility and constructing personalized machine learning models

Background To train and develop machine learning models on the Pfizer antibacterial and antifungal datasets with subsequent deployment to an interactive Web Application. Methods We utilized R version 4.3.1 to perform descriptive analysis to obtain features/predictors. Python 3.10 libraries NumPy, Pandas, Scikit learn, Pycaret were used to train machine learning models. All these models were scored using area under the curve, recall, precision, F1, Kappa and the Mathews correlation coefficient. The best performing model was then deployed into a web application built on Streamlit after which deployment was done using GitHub and Streamlit cloud. A prototype android app was also developed using GoNative. Results The exploratory data analysis, S Aureus (17.2%) was the most common species however in the sub group analysis of the isolates with genotypic values, K Pneumoniae(48.2%) and E Coli (36.9%) were dominant. Amongst the fungi, Candida albicans (38.3%) and Candida glabrata (15.5%) were dominant. Feature selection was done using Shapley additive explanation plots. Using Extreme Gradient Boosting (XGBoost), we were able to achieve 99% and 97.8% AUC in the prediction of antibacterial and antifungal susceptibility respectively with minimal overfitting. Conclusions Decision tree based methods are a viable options of predicting antibacterial and antifungal drug resistance. When presented in visually appealing modes like web applications and android apps, it can be a useful guide to clinicians on initial treatment while awaiting definitive phenotypic testing. It can also be a surveillance tool that can craft antimicrobial resistance strategies.

Development of interpretable machine learning models to predict in-hospital prognosis of acute heart failure patients.

In recent years, there has been remarkable development in machine learning (ML) models, showing a trend towards high prediction performance. ML models with high prediction performance often become structurally complex and are frequently perceived as black boxes, hindering intuitive interpretation of the prediction results. We aimed to develop ML models with high prediction performance, interpretability, and superior risk stratification to predict in-hospital mortality and worsening heart failure (WHF) in patients with acute heart failure (AHF). Based on the Kyoto Congestive Heart Failure registry, which enrolled 4056 patients with AHF, we developed prediction models for in-hospital mortality and WHF using information obtained on the first day of admission (demographics, physical examination, blood test results, etc.). After excluding 16 patients who died on the first or second day of admission, the original dataset (n=4040) was split 4:1 into training (n=3232) and test datasets (n=808). Based on the training dataset, we developed three types of prediction models: (i) the classification and regression trees (CART) model; (ii) the random forest (RF) model; and (iii) the extreme gradient boosting (XGBoost) model. The performance of each model was evaluated using the test dataset, based on metrics including sensitivity, specificity, area under the receiver operating characteristic curve (AUC), Brier score, and calibration slope. For the complex structure of the XGBoost model, we performed SHapley Additive exPlanations (SHAP) analysis, classifying patients into interpretable clusters. In the original dataset, the proportion of females was 44.8% (1809/4040), and the average age was 77.9±12.0. The in-hospital mortality rate was 6.3% (255/4040) and the WHF rate was 22.3% (900/4040) in the total study population. In the in-hospital mortality prediction, the AUC for the XGBoost model was 0.816 [95% confidence interval (CI): 0.815-0.818], surpassing the AUC values for the CART model (0.683, 95% CI: 0.680-0.685) and the RF model (0.755, 95% CI: 0.753-0.757). Similarly, in the WHF prediction, the AUC for the XGBoost model was 0.766 (95% CI: 0.765-0.768), outperforming the AUC values for the CART model (0.688, 95% CI: 0.686-0.689) and the RF model (0.713, 95% CI: 0.711-0.714). In the XGBoost model, interpretable clusters were formed, and the rates of in-hospital mortality and WHF were similar among each cluster in both the training and test datasets. The XGBoost models with SHAP analysis provide high prediction performance, interpretability, and reproducible risk stratification for in-hospital mortality and WHF for patients with AHF.

Identifying potential (re)hemorrhage among sporadic cerebral cavernous malformations using machine learning

The (re)hemorrhage in patients with sporadic cerebral cavernous malformations (CCM) was the primary aim for CCM management. However, accurately identifying the potential (re)hemorrhage among sporadic CCM patients in advance remains a challenge. This study aims to develop machine learning models to detect potential (re)hemorrhage in sporadic CCM patients. This study was based on a dataset of 731 sporadic CCM patients in open data platform Dryad. Sporadic CCM patients were followed up 5 years from January 2003 to December 2018. Support vector machine (SVM), stacked generalization, and extreme gradient boosting (XGBoost) were used to construct models. The performance of models was evaluated by area under receiver operating characteristic curves (AUROC), area under the precision-recall curve (PR-AUC) and other metrics. A total of 517 patients with sporadic CCM were included (330 female [63.8%], mean [SD] age at diagnosis, 42.1 [15.5] years). 76 (re)hemorrhage (14.7%) occurred during follow-up. Among 3 machine learning models, XGBoost model yielded the highest mean (SD) AUROC (0.87 [0.06]) in cross-validation. The top 4 features of XGBoost model were ranked with SHAP (SHapley Additive exPlanations). All-Elements XGBoost model achieved an AUROCs of 0.84 and PR-AUC of 0.49 in testing set, with a sensitivity of 0.86 and a specificity of 0.76. Importantly, 4-Elements XGBoost model developed using top 4 features got a AUROCs of 0.83 and PR-AUC of 0.40, a sensitivity of 0.79, and a specificity of 0.72 in testing set. Two machine learning-based models achieved accurate performance in identifying potential (re)hemorrhages within 5 years in sporadic CCM patients. These models may provide insights for clinical decision-making.

The influence of manual segmentation strategies and different phases selection on machine learning-based computed tomography in renal tumors: a systematic review and meta-analysis.

Delineating the region/volume of interest (ROI/VOI) and selecting the phases are of importance in developing machine learning (ML). The results will change when choosing different methods of drawing the ROI/VOI and selecting different phases. However, there is no related standard for delineating the ROI/VOI and selecting the phases in renal tumors to develop ML based on computed tomography (CT). The PubMed and Web of Science were searched for related studies published until March 1, 2023. Inclusion criteria were studies that developed ML models in renal tumors from CT images. And the binary diagnostic accuracy data were extracted to obtain the outcomes, such as sensitivity (SE), specificity (SP), accuracy (ACC), and area under the curve (AUC). Twenty-three papers were included in the meta-analysis with a pooled SE of 87% (95% CI 85-88%), SP of 82% (95% CI 79-85%), and AUC of 91% (95% CI 89-93%) in phases; a pooled SE of 82% (95% CI 80-84%), SP of 85% (95% CI 83-86%), and AUC of 90% (95% CI 88-93%) in phases combined with delineating strategies, respectively. In all different combinations, the contour-focused and single phase produce the highest AUC of 93% (95% CI 90-95%). In subgroup analyses (sample size, year of publication, and geographical distribution), the performance was acceptable on phases and phases combined strategies. To explore the effect of manual segmentation strategies and different phases selection on ML-based CT, we find that the method of single phase (CMP or NP) combined with contour-focused was considered a better strategy compared to the other strategies.

Developing machine learning models to predict multi-class functional outcomes and death three months after stroke in Sweden.

Globally, stroke is the third-leading cause of mortality and disability combined, and one of the costliest diseases in society. More accurate predictions of stroke outcomes can guide healthcare organizations in allocating appropriate resources to improve care and reduce both the economic and social burden of the disease. We aim to develop and evaluate the performance and explainability of three supervised machine learning models and the traditional multinomial logistic regression (mLR) in predicting functional dependence and death three months after stroke, using routinely-collected data. This prognostic study included adult patients, registered in the Swedish Stroke Registry (Riksstroke) from 2015 to 2020. Riksstroke contains information on stroke care and outcomes among patients treated in hospitals in Sweden. Prognostic factors (features) included demographic characteristics, pre-stroke functional status, cardiovascular risk factors, medications, acute care, stroke type, and severity. The outcome was measured using the modified Rankin Scale at three months after stroke (a scale of 0-2 indicates independent, 3-5 dependent, and 6 dead). Outcome prediction models included support vector machines, artificial neural networks (ANN), eXtreme Gradient Boosting (XGBoost), and mLR. The models were trained and evaluated on 75% and 25% of the dataset, respectively. Model predictions were explained using SHAP values. The study included 102,135 patients (85.8% ischemic stroke, 53.3% male, mean age 75.8 years, and median NIHSS of 3). All models demonstrated similar overall accuracy (69%-70%). The ANN and XGBoost models performed significantly better than the mLR in classifying dependence with F1-scores of 0.603 (95% CI; 0.594-0.611) and 0.577 (95% CI; 0.568-0.586), versus 0.544 (95% CI; 0.545-0.563) for the mLR model. The factors that contributed most to the predictions were expectedly similar in the models, based on clinical knowledge. Our ANN and XGBoost models showed a modest improvement in prediction performance and explainability compared to mLR using routinely-collected data. Their improved ability to predict functional dependence may be of particular importance for the planning and organization of acute stroke care and rehabilitation.

Using spectral vegetation indices and machine learning models for predicting the yield of sugar beet (Beta vulgaris L.) under different irrigation treatments

Yield prediction is essential for production planning, resource management, and competitive advantages. The use of vegetation indices and machine learning for yield prediction is rapid, practical, and accurate and contributes to the development of digital agriculture. Although many studies have been carried out on yield estimation, the number of studies on sugar beet and irrigation applications is limited. To solve this problem, in this study yield prediction of sugar beet under different irrigation treatments was carried out using spectral reflection-based vegetation indices obtained by proximal measurements and developed machine learning models. The multilayer perceptron (MLP), random forest (RF), k-nearest (kNN), bagging (BAG), support vector regression (SVR), gaussian processes (GP), and CustomNet (CN) predictors were used in yield prediction. The irrigation treatments included I100 (application of irrigation to increase the available moisture to field capacity by around 45–50% of the water holding capacity at the root depth), I75 (75% of the water applied in the control group), I50 (50% of the water applied in the control group), and I125 (25% of the water applied in the control group). The combinations of vegetation indices were determined using Pearson correlation and principal component analysis. OSAVI, SAVI, and NDVI had the most significant influence on the yield prediction of sugar beet. Among the combinations used, the kNN1 model (including all attributes as input) showed high R2 values of 0.99 for training and 0.64 for testing. In addition, the kNN3 model (including Treatment, NDVI, NWI, and OSAVI as inputs) had R2 values of 0.97 for training and 0.65 for testing, indicating successful outcomes. However, the MLP1 and SVR1 models had lower results with test R2 values of 0.53 and 0.56, respectively. The results demonstrate the successful application of machine learning and vegetation indices for yield prediction. The proposed approaches could encourage agricultural experts and researchers in future work for yield mapping and identification.

Predicting blood transfusion following traumatic injury using machine learning models: A systematic review and narrative synthesis.

Haemorrhage is a leading cause of preventable death in trauma. Accurately predicting a patient's blood transfusion requirement is essential but can be difficult. Machine learning (ML) is a field of artificial intelligence that is emerging within medicine for accurate prediction modelling. This systematic review aimed to identify and evaluate all ML models that predict blood transfusion in trauma. This systematic review was registered on The International Prospective register of Systematic Reviews (CRD4202237110). MEDLINE, Embase and the Cochrane Central Register of Controlled Trials were systematically searched. Publications reporting a ML model that predicted blood transfusion in injured adult patients were included. Data extraction and risk of bias assessment was performed using validated frameworks. Data was synthesised narratively due to significant heterogeneity. Twenty-five ML models for blood transfusion prediction in trauma were identified. Models incorporated diverse predictors and varied ML methodologies. Predictive performance was variable but eight models achieved excellent discrimination (AUROC >0.9) and nine models achieved good discrimination (AUROC >0.8) in internal validation. Only two models reported measures of calibration. Four models have been externally validated in prospective cohorts: the Bleeding Risk Index, Compensatory Reserve Index, the Marsden model and the Mina model. All studies were considered at high risk of bias often due to retrospective datasets, small sample size and lack of external validation. This review identified twenty-five ML models developed to predict blood transfusion requirement after injury. Seventeen ML models demonstrated good to excellent performance in-silico but only four models were externally validated. To date ML models demonstrate the potential for early and individualised blood transfusion prediction but further research is critically required to narrow the gap between ML model development and clinical application. Systematic Review Without Meta-Analysis, Level IV.

Using ChatGPT to annotate a dataset: A case study in intelligent tutoring systems

Large language models like ChatGPT can learn in-context (ICL) from examples. Studies showed that, due to ICL, ChatGPT achieves impressive performance in various natural language processing tasks. However, to the best of our knowledge, this is the first study that assesses ChatGPT's effectiveness in annotating a dataset for training instructor models in intelligent tutoring systems (ITSs). The task of an ITS instructor model is to automatically provide effective tutoring instruction given a student's state, mimicking human instructors. These models are typically implemented as hardcoded rules, requiring expertise, and limiting their ability to generalize and personalize instructions. These problems could be mitigated by utilizing machine learning (ML). However, developing ML models requires a large dataset of student states annotated by corresponding tutoring instructions. Using human experts to annotate such a dataset is expensive, time-consuming, and requires pedagogical expertise. Thus, this study explores ChatGPT's potential to act as a pedagogy expert annotator. Using prompt engineering, we created a list of instructions a tutor could recommend to a student. We manually filtered this list and instructed ChatGPT to select the appropriate instruction from the list for the given student's state. We manually analyzed ChatGPT's responses that could be considered incorrectly annotated. Our results indicate that using ChatGPT as an annotator is an effective alternative to human experts. The contributions of our work are (1) a novel dataset annotation methodology for the ITS, (2) a publicly available dataset of student states annotated with tutoring instructions, and (3) a list of possible tutoring instructions.

Development, evaluation and validation of machine learning models to predict hospitalizations of patients with coronary artery disease within the next 12 months

BackgroundImproved survival of patients after acute coronary syndromes, population growth, and overall life expectancy rise have led to a significant increase in the proportion of patients with stable coronary artery disease (CAD), creating a significant load on the entire healthcare system. The disease often progresses with the development of many complications while significantly increasing the likelihood of hospitalization. Developing and applying a machine learning model for predicting hospitalizations of patients with CAD to an inpatient medical facility will allow for close monitoring of high-risk patients, early preventive interventions, and optimized medical care. AimsDevelopment and external validation of personalized models for predicting the preventable hospitalizations of patients with stable CAD and its complications using ML algorithms and data of real-world clinical practice. Methods135,873 depersonalized electronic health records of 49,103 patients with stable CAD were included in the study. Anthropometric measurements, physical examination results, laboratory, instrumental, anamnestic, and socio-demographic data, widely used in routine medical practice, were considered as potential predictors, a total of 73 features. Logistic regression, decision tree-based methods including gradient boosting (AdaBoost, LightGBM, XGBoost, CatBoost) and bagging (RandomForest and ExtraTrees), discriminant analysis (LinearDiscriminant, QuadraticDiscriminant), and naive Bayes classifier were compared. External validation was performed on the data of a separate region. ResultsThe best results and stability to external validation data were shown by the CatBoost model with an AUC of 0.875 (95% CI 0.865–0.885) for the internal testing and 0.872 (95% CI 0.856–0.886) for the external validation. The best model showed good performance evaluated through AUROC, Brier score and standardized net benefit (for the target NPV threshold) for the validation dataset that was only slightly similar to the train data. ConclusionThe metrics of the best model were superior to previously published studies. The results of external validation demonstrated the relative stability of the model to new data from another region that confirms the possibility of the model’s application in real clinical practice.

Innovative integrated workflow for data-driven production forecasting and well completion optimization: A Montney Formation case study

With the use of advanced data analytics and machine learning (ML) techniques, the oil and gas industry has transformed significantly in recent years. These innovations have opened new opportunities in reservoir engineering, allowing engineers to leverage data for better decision-making and well completion optimization. A key challenge in hydrocarbon exploration and production is forecasting gas production rates from unconventional sources. This is a crucial factor that affects how resources are allocated and decisions are made, requiring the application of novel techniques to enhance hydrocarbon production. Here, a novel and comprehensive automated workflow is introduced covering data collection, data preparation, feature selection, ML model development, and operational parameter optimization. The workflow uses various ML and optimization methods, including a modified Differential Evolution (MDE) algorithm, to improve gas production forecasting and optimize well completion parameters. We applied the proposed workflow to 3144 horizontal gas wells in the Montney Formation, British Columbia (BC), Canada, to assess its performance. The results showed the advantages of incorporating the initial 24 months of cumulative gas production data alongside well and completion data. Also, following the fine-tuning of hyperparameters for the ML algorithms, we identified the most effective model for forecasting cumulative gas production within the Montney Formation as a two-layer Artificial Neural Network (ANN) model. Moreover, using the MDE algorithm, the optimization results revealed the possibility of increasing gas production by more than 10% for about 47% of the wells. This research makes a significant contribution to the field of unconventional gas exploration by introducing an automated workflow that simplifies data processing, model development, and decision-making processes. It also offers valuable insights for optimizing unconventional gas production in the Montney Formation and similar reservoirs, as the energy industry moves towards more sustainable and efficient practices.

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.

Development and validation of machine learning models to predict unplanned hospitalizations of patients with diabetes within the next 12 months

BACKGROUND: The incidence of diabetes mellitus (DM) both in the Russian Federation and in the world has been steadily increasing for several decades. Stable population growth and current epidemiological characteristics of DM lead to enormous economic costs and significant social losses throughout the world. The disease often progresses with the development of specific complications, while significantly increasing the likelihood of hospitalization. The creation and inference of a machine learning model for predicting hospitalizations of patients with DM to an inpatient medical facility will make it possible to personalize the provision of medical care and optimize the load on the entire healthcare system.AIM: Development and validation of models for predicting unplanned hospitalizations of patients with diabetes due to the disease itself and its complications using machine learning algorithms and data from real clinical practice.MATERIALS AND METHODS: 170,141 depersonalized electronic health records of 23,742 diabetic patients were included in the study. Anamnestic, constitutional, clinical, instrumental and laboratory data, widely used in routine medical practice, were considered as potential predictors, a total of 33 signs. Logistic regression (LR), gradient boosting methods (LightGBM, XGBoost, CatBoost), decision tree-based methods (RandomForest and ExtraTrees), and a neural network-based algorithm (Multi-layer Perceptron) were compared. External validation was performed on the data of the separate region of Russian Federation.RESULTS: The best results and stability to external validation data were shown by the LightGBM model with an AUC of 0.818 (95% CI 0.802–0.834) in internal testing and 0.802 (95% CI 0.773–0.832) in external validation.CONCLUSION: The metrics of the best model were superior to previously published studies. The results of external validation showed the relative stability of the model to new data from another region, that reflects the possibility of the model’s application in real clinical practice.

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.