Best-performing Model Research Articles

Predictive score for response to neoadjuvant chemotherapy in early-stage HR + /HER2- breast cancer.

Neoadjuvant chemotherapy (NACT) is a treatment option for early-stage hormone receptor-positive human epidermal growth factor receptor 2-negative (HR + /HER2-) breast cancer. Despite its use, pathological response rates in this subtype are often lower, and the impact of individual risk factors remains unclear. This study aimed to identify biomarkers and create a predictive score for NACT response. This retrospective, single-center study included patients with stage IIA-IIIC HR + /HER2- breast cancer treated with NACT and surgery (2019-2023). Multiple logistic regression analyzed associations between clinicopathological variables and pathologic response (partial/complete vs. absent) (p < 0.05). The best-performing model was used to develop a risk score. The study included 101 patients. Significant predictors of pathological response included tumor grade (G2/3 vs. G1), menopausal status (pre- vs. post-menopausal), and intrinsic subtype (luminal B vs. A). A dynamic calculator was created incorporating grade, hormonal status, intrinsic subtype, and Ki-67. This tool provides real-time input, facilitating personalized therapeutic decision-making.

Machine Learning and Deep Learning Models for Dengue Diagnosis Prediction: A Systematic Review

The global crisis triggered by the dengue outbreak has increased mortality and placed significant pressure on healthcare services worldwide. In response to this crisis, there has been a notable increase in research employing machine learning and deep learning algorithms to anticipate diagnosis in patients with suspected dengue. To conduct a comprehensive systematic review, a detailed analysis was carried out to explore and examine the machine learning methodologies applied in diagnosing this disease. An exhaustive search was conducted across numerous scientific databases, including Scopus, IEEE Xplore, PubMed, ACM, ScienceDirect, Wiley, and Sage, encompassing studies up to May 2024. This extensive search yielded a total of 2723 relevant articles. Following a rigorous evaluation, 32 scientific studies were selected for the final review, meeting the established criteria. A comprehensive analysis of these studies revealed the implementation of 48 distinct machine learning and deep learning algorithms, showcasing the heterogeneity of methodological approaches employed in the research domain. The results indicated that, in terms of performance, the support vector machine (SVM) algorithm was the most efficient, being reported in 25% of the analyzed studies. The Random Forest algorithm was the second most frequently used, appearing in 15.62% of the 32 reviewed articles. The PCA-SVM algorithm (poly-5), a variant of SVM, emerged as the best-performing model, achieving 99.52% accuracy, 99.75% sensitivity, and 99.09% specificity. These findings offer significant insights into the potential of machine learning techniques in the early diagnosis of dengue, underscoring the necessity to persist in exploring and refining these methodologies to enhance clinical care in cases of this disease.

Performance of Machine Learning Models in Predicting BRAF Alterations Using Imaging Data in Low-Grade Glioma: A Systematic Review and Meta-Analysis.

Understanding the BRAF alterations preoperatively could remarkably assist in predicting tumor behavior, which leads to a more precise prognostication and management strategy. Recent advances in artificial intelligence (AI) have resulted in effective predictive models. Therefore, for the first time, this study aimed to review the performance of machine learning (ML) and deep learning (DL) models in predicting the BRAF alterations in LGGs using imaging data. PubMed/MEDLINE, Embase, and the Cochrane Library were systematically searched for studies published up to June 1, 2024, and evaluated the performance of AI models in predicting BRAF alterations in LGGs using imaging data. Pooled sensitivity, specificity, and area under the receiver operating characteristics (AUROC) were meta-analyzed. A total of 6 studies with 951 patients were included in this systematic review. The pooled AUROC of internal validation cohorts for their best-performing model was 84.44 %, with models detecting BRAF mutation, BRAF fusion, BRAF fusion from mutation, and BRAF wild type producing similar AUROCs of 90.75%, 84.59%, 82.33%, and 82%, respectively. The best-performing models had pooled sensitivities of 80.3%, 87.51%, and 74.14% and pooled specificities of 88.57%, 70.41%, and 83.98% for detection of BRAF fusion from mutation, BRAF fusion, and BRAF mutation, respectively. AI models may perform relatively well in predicting BRAF alterations in LGG using imaging data and appear to be capable of high sensitivities and specificities. However, future studies with larger sample sizes implementing different ML or DL algorithms are required to reduce imprecision.

Explorative Short-Term Predictive Models for the Belgian (Energy) Renovation Market Incorporating Macroeconomic and Sector-Specific Variables

Retrofitting existing buildings is a cornerstone of Europe’s strategy for a sustainable built environment. Therefore, accurate short-term forecasts to evaluate policy impacts and inform future strategies are needed. This study investigates the short-term predictive modelling of renovation activity in Belgium, focusing on overall renovation activity (RA) and energy-specific renovation activity (EA). Using data from 2012 to 2023, linear modelling was employed to analyze the relationships between RA/EA and macroeconomic indicators, market confidence, building permits, and loan data, with model performance evaluated using Mean Absolute Percentage Error (MAPE). Monthly data and time lags of up to 24 months were considered. The three best-performing models for RA achieved MAPE values between 2.9% and 3.1%, with validated errors ranging from 0.1% to 4.1%. For EA, the best models yielded MAPE values between 4.4% and 4.6% and validated errors between 8.9% and 14%. Renovation loans and building permits emerged as strong predictors for RA, while building material prices and loans were more relevant for EA. The time lag analysis highlighted that renovation processes typically span 15–24 months following loan approval. However, the low accuracy observed for EA underscores the need for further refinement. This explorative effort forms a solid base, inviting additional research to enhance our predictive capabilities and improve short-term modelling of the (green) residential renovation market.

Optimizing high-strength concrete compressive strength with explainable machine learning

This study leverages machine learning to enhance the prediction of high-strength concrete (HSC) compressive strength, addressing the limitations of conventional methods, which are often tedious, less reliable, and time-consuming. Extreme Gradient Boosting (XGB) serves as the primary model, with hyperparameter optimization via metaheuristic algorithms such as Cuckoo Search (CSA), Water Strider (WS), Leopard Seal (LS), Harris Hawk (HH), Invasive Weed (IW), and Forest Optimization (FO). A total of 681 data sets were collected from existing literature. The models underwent tenfold cross-validation, with the LS-XGB model achieving an almost ideal performance in testing sets. Other models, including CSA-XGB, WS-XGB, HH-XGB, IW-XGB, and FO-XGB, also demonstrated strong performance, each with R2 > 0.96. For model explainability, Shapley's Additive Explanation (SHAP) analysis has been applied to the best-performing LS-XGB model. The analysis revealed that cement and superplasticizer (SP) are the most crucial features contributing to HSC development, with optimal ranges identified at 600–900 kg/m3 for cement and 8–10 kg/m3 for SP. The study demonstrates on how feature interactions contribute to concrete materials compressive strength, providing better and above all sustainable constructions. Furthermore, the LS-XGB model's optimal performance depicts the strongly nonlinear nature of HSC materials, validated through a set of derived graphs. Additionally, 30 concrete cubes were prepared for experimental validation, and the datasets demonstrated an accuracy of 92% showcasing the ability of models to make well informed decision.

Elevated serum levels of GPX4, NDUFS4, PRDX5, and TXNRD2 as predictive biomarkers for castration resistance in prostate cancer patients: an exploratory study.

Prostate cancer (PCa) is a heterogeneous disease affecting over 14% of the male population worldwide. Although patients often respond positively to initial treatments within the first 2-3 years, many eventually develop a more lethal form of the disease known as castration-resistant PCa (CRPC). At present, no biomarkers that predict the onset of CRPC are available. This study aims to provide insights into the diagnosis and prediction of CRPC emergence. Protein expression dynamics were analysed in drug (androgen receptor inhibitor)-tolerant persister (DTP) and drug withdrawal cells using proteomics to identify potential biomarkers. These biomarkers were subsequently validated using a mouse model, 180-paired carcinoma/benign tissues, and 482 serum samples. Five machine learning algorithms were employed to build clinical prediction models, wherein the SHapley Additive exPlanation (SHAP) framework was used to interpret the best-performing model. Moreover, three regression models were developed to determine the Time from initial PCa diagnosis to CRPC development (TPC) in patients. We identified that the protein expression levels of GPX4, NDUFS4, PRDX5, and TXNRD2 were significantly upregulated in PCa patients, particularly in those with CRPC. Among the tested machine learning models, the random forest and extreme gradient boosting models performed best on tissue and serum cohorts, achieving AUCs of 0.958 and 0.988, respectively. In addition, a significant inverse correlation was observed between TPC and serum levels of these four biomarkers. This correlation was formulated in three regression models, which achieved the smallest mean absolute error of 1.903 on independent datasets for predicting CRPC emergence. Our study provides new insights into the role of DTP cells in CRPC development. The quad protein panel identified in our study, along with the post hoc and intrinsically explainable prediction models, may serve as a convenient and real-time prognostic tool, addressing the current lack of clinical biomarkers for CRPC.

Smartphone pupillometry with machine learning differentiates ischemic from hemorrhagic stroke: A pilot study.

Similarities between acute ischemic and hemorrhagic stroke make diagnosis and triage challenging. We studied a smartphone-based quantitative pupillometer for differentiation of acute ischemic and hemorrhagic stroke. Stroke patients were recruited prior to surgical or interventional treatment. Smartphone pupillometry was used to quantify components of the pupillary light reflex (PLR). A synthetic minority oversampling technique (SMOTE) was applied to correct sample size imbalance. Four binary classification model types were trained using all possible combinations of the PLR components with 10-fold cross validation stratified by cohort. Models were evaluated for accuracy, sensitivity, specificity, area under the curve (AUC), and F1 score. The three best-performing models were selected based on AUC. Shapley additive explanation plots were produced to explain PLR parameter impacts on model predictions. Eleven subjects with intraparenchymal hemorrhage and 22 subjects with acute ischemic stroke were enrolled. One way ANOVA demonstrated significant differences between healthy control data, AIS, and IPH in five out of seven PLR parameters. After SMOTE, each class had n=22 PLR recordings for model training. The best-performing model was random forest using a combination of latency, mean and maximum constriction velocity, and mean dilation velocity to discriminate between stroke types with 91.5% (95% confidence interval: 84.1-98.9) accuracy, 90% (82.9-97.1) sensitivity, 93.3% (83-100) specificity, 0.917 (0.847-0.987) AUC, and 90.7% (84.1-97.3) F1 score. Smartphone-based quantitative pupillometry could be useful in differentiating between acute ischemic and hemorrhagic stroke.

A Personalized Periodontitis Risk Based on Nonimage Electronic Dental Records by Machine Learning

Objective: This study aimed to develop a machine-learning model to predict the risk for Periodontal Disease (PD) based on nonimage electronic dental records (EDRs).Methods: By using EDRs collected in the BigMouth repository, dental patients from the US were included. Patients were labeled as cases or controls, based on PD diagnosis, treatment and pocketing. By learning from their data, a model was trained. The ability of the developed model to predict PD was evaluated by the accuracy, sensitivity, specificity and area under the curve (AUROC) and the most important features were determined. The best-performing model was applied to the validation set.Results: The final study population included 43,331 participants. Based on the development set, the Random Forest model performed with high sensitivity (81%) and had an excellent AUROC (94%), compared to four other ML and deep learning techniques. The most important predictors were bleeding proportion, age, the number of visits, prior preventive treatment, smoking and drugs usage. When the model was applied to the validation set, the model could detect almost all cases (91%), but overestimated controls (specificity=0.54). When EDRs were retrieved 3 years before the PD diagnosis, the predictions for PD were still sensitive (89%).Conclusion: Based on consistent and complete EDR, ML has an excellent ability to assist with the early detection and prevention of PD cases. Further research is required to follow-up high-risk controls and improve the model's internal and external validation. Improved EDR documentation is an important first step.Clinical significance: If such ML models become clinically applied, clinicians can be assisted with personalized risk predictions based on the individual. If the key riskcontributing factors for the individual are revealed/provided, ML can suggest targeted prevention interventions. These advancements can contribute to a reduced workload, sustainable EDRs, data-based dental care, and, ultimately, improved patient outcomes.

Machine learning models for quantitatively prediction of toxicity in macrophages induced by metal oxide nanoparticles.

As nanotechnology advances, metal oxide nanoparticles (MeONPs) increasingly come into contact with humans. The inhaled MeONPs cannot be effectively cleared by cilia or lung mucus. In the last decade, potential immune toxicity arising from exposure to MeONPs has been extensively debated, as lung macrophage is the main pathway for cleaning inhaled exogenous particles. However, their toxicity on lung macrophages has rarely been quantitatively predicted in silico due to the complexity of responses in macrophages and the intricate properties of MeONPs. Here, machine learning (ML) methods were used to establish models for quantitatively predicting the toxicity of MeONPs in macrophages. A multidimensional dataset including 240 data points covering the lethality, biochemical behaviors, and physicochemical properties of 30 MeONPs was obtained. ML models based on different algorithms with high prediction accuracy were constructed by addressing the issue of class imbalance during the training process. The models were verified by 10-fold cross-validation and external validation. The best-performed model has an R2 of 0.85 and 0.90 in the 10-fold cross-validation and external test set, respectively; and Q2 of 0.88 and 0.90 in the 10-fold cross-validation and test set, respectively. Five parameters that impact toxicity were identified and the toxicity mechanisms were elucidated by ML analysis. The prediction results can be used to fill the data gap in the risk assessment of nanomaterials. The framework offers valuable insights for designing and utilizing safe nanoparticles, as well as aiding in decision-making processes aimed at protecting the environment and public health.

Revolutionizing Diagnostic Insights: Exploring Advanced Image Processing Techniques and Neural Networks in Traditional Indian Medicine

The Siddha and Ayurveda traditional Indian medicine practices utilize non-invasive diagnostic methods, such as Neikuri and Taila Bindu Pariksha, for patient diagnosis through urine analysis. While these methods have proven effective for centuries, their accuracy highly depends on the subjective experience of practitioners. To address this limitation, this study explores the use of advanced image processing techniques and deep learning, specifically Convolutional Neural Networks (CNNs), to automate and enhance diagnostic image analysis. This study utilized five pre-trained CNN models, namely DenseNet, ResNet, VGG-19, Inception, and EfficientNet, on a dataset of Neikuri images acquired from a Siddha medical institute, to standardize and improve the accuracy of patient diagnosis. The comparative evaluation revealed DenseNet as the best-performing model, achieving a classification accuracy of 93.33%, while Inception v3 followed with 90.5%. This study highlights the potential of integrating modern neural networks with traditional diagnostic practices, paving the way for more objective, efficient, and accessible healthcare solutions in traditional Indian medicine.

FIND BEAR: An explainable artificial intelligence platform for stages automatic recognition in pool boiling

Pool boiling heat transfer plays a critical role as a high heat flux thermal control technology in a variety of scientific and application scenarios. However, automatic identifications of pool boiling stages with non-invasive technology achieving outstanding performance is barely reported. In this paper, we develop a novel artificial intelligence (AI) platform called Boiling-stages Explainable and Automatic Recognition (BEAR) and integrate it into our own FIND Multiphysics software. Inspired by the top-down and bottom-up attentional control mechanism of the human brain, BEAR aims to mimic the cognitive and discriminative processes of human experts while giving full play to the computational advantages of personal computers. This aim is decomposed into three goals and realized by the corresponding modules forming the architecture. BEAR is effective with an identification accuracy of up to 99.1 %. It is cost-efficient as the training time of the best-performing model embedded is 9.06 s with the model’s size of 65kB. It takes only about 0.048 ∼ 0.197 s per image to perform the entire automated identification process on a personal computer with an 8-core processor. Even a small dataset with low spatial and temporal resolution, comprising 33 images of 256 × 256 pixels, can yield results that are comparable to state-of-the-art literature in the field. Moreover, BEAR provides a platform with the scalability to carry a variety of algorithms for different application purposes, which can be beneficial in facilitating innovative modeling and further deployment in future application devices.

Abstract TP335: Machine learning model for prediction of functional outcome in patients with recent small subcortical infarction.

Objective: Approximately 10%-17% of patients with a recent small subcortical infarction (RSSI) experience an adverse functional prognosis at three months. Identifying risk factors for poor prognosis in these patients and developing a predictive model are of significant clinical relevance. Methods: We conducted a prospective cohort study involving 630 patients with recent small subcortical infarction (RSSI). These patients were divided into training (70%) and internal validation (30%) cohorts. Additionally, for external validation, 103 patients from two other hospitals were included. To predict functional outcomes at three months, we employed eight different machine learning algorithms: Support Vector Machine, Decision Tree, Random Forest, Gradient Boosting, Extreme Gradient Boosting, Light Gradient Boosting Machine, Cat Boosting, and Logistic Regression. The interpretability of the best-performing model was enhanced with SHapley Additive exPlanations(SHAP). Results: Among the 630 patients included in the study, the average age of participants was 59.0 ±13.0 years and 65 (10.3%) exhibited poor functional outcomes at three months. The Cat Boosting model showed the highest performance among eight machine learning models. The final model could accurately predict outcome in both internal (AUC = 0.968) and external (AUC = 0.856) validations. SHAP analysis indicated that the initial NIHSS score was the most important variable. The web application is available at https://yrssiml.streamlit. app. Conclusions: Our research developed a predictive model using the Cat Boosting algorithm to forecast poor functional outcomes at three months in patients with RSSI. This model enables clinicians to identify patients at high risk and implement targeted interventions more effectively.

Location-Based Technology for Real-Time Artifact Recognition in Businesses

The recognition of historical artifacts play a crucial role in sustaining cultural heritage and advancing tourism. Despite advancements in object detection technologies, accurately identifying artifacts in diverse geographical and environmental contexts remains a significant challenge. Existing models often struggle to adapt to region-specific features and the complexity of historical artifacts, limiting their practical applications. To address these limitations, this study evaluates the potential of YOLOv4, YOLOv7-X, and YOLOv9c models for historical artifact recognition, with a particular focus on location-based segmentation. Geographically distinct datasets were utilized for training and evaluation, enabling the models to achieve higher accuracy in region-specific artifact detection. Among the tested models, YOLOv9c demonstrated superior performance, achieving the highest metrics across accuracy (96%), precision (93%), recall (95%), and mean average precision (mAP, 71%), making it the best-performing model. These results highlight YOLOv9c’s robustness and adaptability to complex datasets and diverse artifact characteristics. A user-friendly application interface was also developed, allowing real-time detection and providing detailed historical information about the artifacts. However, challenges such as the high computational cost of training YOLOv9c on high-resolution datasets were observed, particularly when compared to YOLOv4, which was computationally efficient but less accurate. YOLOv7-X offered a balance between performance and computational efficiency. The results demonstrate that location-based segmentation significantly enhances detection accuracy, making this approach highly effective for real-world applications in cultural heritage preservation and tourism.

Spinal cord demyelination predicts neurological deterioration in patients with mild degenerative cervical myelopathy

BackgroundDegenerative cervical myelopathy (DCM) is the most common form of atraumatic spinal cord injury globally. Clinical guidelines regarding surgery for patients with mild DCM and minimal symptoms remain uncertain. This...

Exploring the assessment of post-cardiac valve surgery pulmonary complication risks through the integration of wearable continuous physiological and clinical data

BackgroundPostoperative pulmonary complications (PPCs) following cardiac valvular surgery are characterized by high morbidity, mortality, and economic cost. This study leverages wearable technology and machine learning algorithms to preoperatively identify high-risk individuals, thereby enhancing clinical decision-making for the mitigation of PPCs.MethodsA prospective study was conducted at the Department of Cardiovascular Surgery of West China Hospital, Sichuan University, from August 2021 to December 2022. We examined 100 cardiac valvular surgery patients, where wearable technology was utilized to collect and analyze nocturnal physiological data at the 24-hour admission, in conjunction with clinical data extraction from the Hospital Information System’s electronic records. We systematically evaluated three different input types (physiological, clinical, and both) and five classifiers (XGB, LR, RF, SVM, KNN) to identify the combination with strong predictive performance for PPCs. Feature selection was conducted using Recursive Feature Elimination with Cross-Validated (RFECV) for each model, yielding an optimal feature subset for each, followed by a grid search to tune hyperparameters. Stratified 5-fold cross-validation was used to evaluate the generalization performance. The significance of AUC differences between models was tested using the DeLong test to determine the optimal prognostic model comprehensively. Additionally, univariate logistic regression analysis was conducted on the features of the best-performing model to understand the impact of individual feature on PPCs.ResultsIn this study, 22 patients (22%) developed PPCs. Across classifiers, models combining both physiological and clinical features performed better than physiological or clinical features alone. Specifically, including physiological data in the classification model improved AUC, ACC, F1, and precision by an average of 8.32%, 1.80%, 3.28% and 6.06% compared to using clinical data only. The XGB classifier, utilizing both dataset, achieved the highest performance with an AUC of 0.82 (± 0.08) and identified eight significant features. The DeLong test indicated that the XGB model utilizing the both dataset significantly outperformed the XGB models trained on the physiological or clinical datasets alone. Univariate logistic regression analysis suggested that surgical methods, age, nni_50, and min_ven_in_mean are significantly associated with the occurrence of PPCs.ConclusionThe integration of continuous wearable physiological and clinical data significantly improves preoperative risk assessment for PPCs, which helps to optimize surgical management and reduce PPCs morbidity and mortality.

Postoperative fever following surgery for oral cancer: Incidence, risk factors, and the formulation of a machine learning-based predictive model.

Postoperative fever (POF) is a common occurrence in patients undergoing major surgery, presenting challenges and burdens for both patients and surgeons yet. This study endeavors to examine the incidence, identify risk factors, and establish a machine learning-based predictive model for POF following surgery of oral cancer. A total of seven hundred and twenty-seven consecutive patients undergoing radical resection of oral cancer were retrospectively investigated. The analysis encompassed 34 parameters, incorporating demographic and clinical characteristics, biochemical and hematological assay results, surgical-related data, hospitalization costs and stay in hospital. Six machine learning models were compared by the area under the receiver operating characteristic curve (AUC). The best-performing models were selected for further analyze, including feature importance evaluation and nomogram analysis, identifying key POF risk factors, and establish a comprehensive prediction model. A total of 466 patients with surgery for oral cancer met the criteria, with an average age of (54.2 ± 11.1) years, including an POF group (n = 197) and a non-POF group (n = 269). The fever group has greater hospitalization costs, longer lengths of stay, and higher infection biochemical indicators (leucocyte ratio and erythrocyte sedimentation rate). Furthermore, Among the 6 machine learning models, logistic regression models performed best, with the higher AUC and accuracy. In univariate and multivariate logistic analysis showed that age, sex, reoperation, Charlson Comorbidity Index score (CCI), leukocyte, bleeding and blood transfusion were independent risk factors for POF of patients following surgery in oral cancer. Then seven variables were selected to establish the nomogram for predict the probability of POF by nomogram algorithm. Postoperative fever patients following radical resection of oral cancer have greater burden. Machine learning algorithms can be effectively used to identify potential risk factors of POF, which may enhance individualized treatment plans in oral cancer patient during perioperative period.

Development and Validation of Machine Learning Models for Risk Prediction of Postpartum Stress Urinary Incontinence: A Prospective Observational Study.

This study aims to develop a postpartum stress urinary incontinence (PPSUI) risk prediction model based on an updated definition of PPSUI, using machine learning algorithms. The goal is to identify the best model for early clinical screening to improve screening accuracy and optimize clinical management strategies. This prospective study collected data from 1208 postpartum women, with the dataset randomly divided into training and testing sets (8:2). Five machine learning algorithms-logistic regression, decision trees, random forests, support vector machines (SVM), and eXtreme gradient boosting (XGBoost)-were used to construct the PPSUI risk prediction model. Model performance was evaluated using multiple metrics, and the best-performing model was selected and validated for generalizability with the testing set. The final model retained ten features: birth weight, weight gain during pregnancy, BMI before delivery, pre-pregnancy BMI, age of delivery, gestation, parity, pre-delivery uterine height, age of first delivery, and labor analgesia. Among the five algorithms, the random forest model performed best, with a test AUC of 0.995 (95% CI 0.990-0.999, P < 0.05), accuracy of 0.956, precision of 0.957, recall of 0.944, specificity of 0.966, and F1 score of 0.951. The model's high generalizability was confirmed with the testing set and further validated through bootstrapping and tenfold cross-validation. The random forest model shows strong clinical potential for PPSUI risk prediction and early screening. Future studies should expand the sample size and include multi-center data to further enhance the model's clinical applicability.

A robust auto-contouring and data augmentation pipeline for adaptive MRI-guided radiotherapy of pancreatic cancer with a limited dataset

Objective.This study aims to develop and evaluate a fast and robust deep learning-based auto-segmentation approach for organs at risk in MRI-guided radiotherapy of pancreatic cancer to overcome the problems of time-intensive manual contouring in online adaptive workflows. The research focuses on implementing novel data augmentation techniques to address the challenges posed by limited datasets.Approach.This study was conducted in two phases. In phase I, we selected and customized the best-performing segmentation model among ResU-Net, SegResNet, and nnU-Net, using 43 balanced 3DVane images from 10 patients with 5-fold cross-validation. Phase II focused on optimizing the chosen model through two advanced data augmentation approaches to improve performance and generalizability by increasing the effective input dataset: (1) a novel structure-guided deformation-based augmentation approach (sgDefAug) and (2) a generative adversarial network-based method using a cycleGAN (GANAug). These were compared with comprehensive conventional augmentations (ConvAug). The approaches were evaluated using geometric (Dice score, average surface distance (ASD)) and dosimetric (D2% and D50% from dose-volume histograms) criteria.Main results.The nnU-Net framework demonstrated superior performance (mean Dice: 0.78 ± 0.10, mean ASD: 3.92 ± 1.94 mm) compared to other models. The sgDefAug and GANAug approaches significantly improved model performance over ConvAug, with sgDefAug demonstrating slightly superior results (mean Dice: 0.84 ± 0.09, mean ASD: 3.14 ± 1.79 mm). The proposed methodology produced auto-contours in under 30 s, with 75% of organs showing less than 1% difference in D2% and D50% dose criteria compared to ground truth.Significance.The integration of the nnU-Net framework with our proposed novel augmentation technique effectively addresses the challenges of limited datasets and stringent time constraints in online adaptive radiotherapy for pancreatic cancer. Our approach offers a promising solution for streamlining online adaptive workflows and represents a substantial step forward in the practical application of auto-segmentation techniques in clinical radiotherapy settings.

Deep learning-based malaria parasite detection: convolutional neural networks model for accurate species identification of Plasmodium falciparum and Plasmodium vivax

Accurate malaria diagnosis with precise identification of Plasmodium species is crucial for an effective treatment. While microscopy is still the gold standard in malaria diagnosis, it relies heavily on trained personnel. Artificial intelligence (AI) advances, particularly convolutional neural networks (CNNs), have significantly improved diagnostic capabilities and accuracy by enabling the automated analysis of medical images. Previous models efficiently detected malaria parasites in red blood cells but had difficulty differentiating between species. We propose a CNN-based model for classifying cells infected by P. falciparum, P. vivax, and uninfected white blood cells from thick blood smears. Our best-performing model utilizes a seven-channel input and correctly predicted 12,876 out of 12,954 cases. We also generated a cross-validation confusion matrix that showed the results of five iterations, achieving 63,654 out of 64,126 true predictions. The model’s accuracy reached 99.51%, a precision of 99.26%, a recall of 99.26%, a specificity of 99.63%, an F1 score of 99.26%, and a loss of 2.3%. We are now developing a system based on real-world quality images to create a comprehensive detection tool for remote regions where trained microscopists are unavailable.

Utilizing echocardiographic findings and machine learning to predict elevated left ventricular filling pressures in patients with preserved ejection fraction

Abstract Background Estimation of left ventricular filling pressures (LVFP) is important for accurate diagnosis and prognosis of heart failure (HF). The current ASE/EACVI algorithm for identifying elevated LVFP shows modest performance in the setting of preserved ejection fraction (EF) (1). Left atrial reservoir strain (LASr) has been proposed as a novel and additional diagnostic marker to identify elevated LVFP in HF with preserved EF (HFpEF) (1). While machine learning has shown good diagnostic performance utilising conventional echocardiographic findings (2), the incremental value of LASr has not been adequately explored using ML models. Purpose This study aimed to evaluate the performance of a machine learning (ML) model that integrates clinical and echocardiographic data including LASr to identify elevated LVFP in the setting of suspected HF with preserved EF. Methods Patients with dyspnoea, EF ≥50% and sinus rhythm undergoing right heart catheterisation (RHC) and near-simultaneous echocardiography were selected from the KARUM database. Patients with atrial fibrillation, pacemakers, significant valvular disease or infiltrative cardiomyopathy were excluded. Elevated LVFP was defined as invasive pulmonary artery wedge pressure (PAWP) ≥15 mmHg. The dataset was split into 75% for model development (training set) and 25% for unbiased testing of model performance (test set). Missing numerical and categorical variables were imputed based on the training data using Bayesian Ridge regression and K-Nearest Neighbours, respectively. Recursive Feature Elimination (RFE) was used to reduce the number of variables. A comprehensive grid search with repeated 10-fold cross-validation was employed to optimise hyperparameters for an XGBoost (Xtreme Gradient Boosting) ML model. The model with the highest mean area under the receiver operating characteristic curve (AUC) across cross-validation folds was reported and selected as the final model, with its performance reported on the independent test set. Finally, we applied SHapley Additive exPlanations (SHAP) to evaluate and rank the individual variable importance. Results Of 210 patients, 157 were allotted for model development (training set) and 53 for reporting model performance (test set). Ten significant variables were identified through RFE. The best-performing XGBoost model achieved a mean AUC of 86.6 % across the cross-validation folds and maintained good diagnostic performance in the test set with an AUC of 89.4% (95% CI: 79.1-99.7%) (Fig 1). Applying SHAP ranked the variables and found LASr the most important variable (Fig 2). Conclusion An ML model based on clinical and echocardiographic data, including LASr, demonstrated good diagnostic performance in identifying elevated LVFP in the setting of suspected HF with preserved EF. ROC Curve for XGBoost Model SHAP summary plot