Stratified Fivefold Cross-validation Research Articles

Frontoparietal and salience network synchronizations during nonsymbolic magnitude processing predict brain age and mathematical performance in youth.

The development and refinement of functional brain circuits crucial to human cognition is a continuous process that spans from childhood to adulthood. Research increasingly focuses on mapping these evolving configurations, with the aim to identify markers for functional impairments and atypical development. Among human cognitive systems, nonsymbolic magnitude representations serve as a foundational building block for future success in mathematical learning and achievement for individuals. Using task-based frontoparietal (FPN) and salience network (SN) features during nonsymbolic magnitude processing alongside machine learning algorithms, we developed a framework to construct brain age prediction models for participants aged 7-30. Our study revealed differential developmental profiles in the synchronization within and between FPN and SN networks. Specifically, we observed a linear increase in FPN connectivity, concomitant with a decline in SN connectivity across the age span. A nonlinear U-shaped trajectory in the connectivity between the FPN and SN was discerned, revealing reduced FPN-SN synchronization among adolescents compared to both pediatric and adult cohorts. Leveraging the Gradient Boosting machine learning algorithm and nested fivefold stratified cross-validation with independent training datasets, we demonstrated that functional connectivity measures of the FPN and SN nodes predict chronological age, with a correlation coefficient of .727 and a mean absolute error of 2.944 between actual and predicted ages. Notably, connectivity within the FPN emerged as the most contributing feature for age prediction. Critically, a more matured brain age estimate is associated with better arithmetic performance. Our findings shed light on the intricate developmental changes occurring in the neural networks supporting magnitude representations. We emphasize brain age estimation as a potent tool for understanding cognitive development and its relationship to mathematical abilities across the critical developmental period of youth. PRACTITIONER POINTS: This study investigated the prolonged changes in the brain's architecture across childhood, adolescence, and adulthood, with a focus on task-state frontoparietal and salience networks. Distinct developmental pathways were identified: frontoparietal synchronization strengthens consistently throughout development, while salience network connectivity diminishes with age. Furthermore, adolescents show a unique dip in connectivity between these networks. Leveraging advanced machine learning methods, we accurately predicted individuals' ages based on these brain circuits, with a more mature estimated brain age correlating with better math skills.

Classification of subtask types and skill levels in robot-assisted surgery using EEG, eye-tracking, and machine learning.

Objective and standardized evaluation of surgical skills in robot-assisted surgery (RAS) holds critical importance for both surgical education and patient safety. This study introduces machine learning (ML) techniques using features derived from electroencephalogram (EEG) and eye-tracking data to identify surgical subtasks and classify skill levels. The efficacy of this approach was assessed using a comprehensive dataset encompassing nine distinct classes, each representing a unique combination of three surgical subtasks executed by surgeons while performing operations on pigs. Four ML models, logistic regression, random forest, gradient boosting, and extreme gradient boosting (XGB) were used for multi-class classification. To develop the models, 20% of data samples were randomly allocated to a test set, with the remaining 80% used for training and validation. Hyperparameters were optimized through grid search, using fivefold stratified cross-validation repeated five times. Model reliability was ensured by performing train-test split over 30 iterations, with average measurements reported. The findings revealed that the proposed approach outperformed existing methods for classifying RAS subtasks and skills; the XGB and random forest models yielded high accuracy rates (88.49% and 88.56%, respectively) that were not significantly different (two-sample t-test; P-value = 0.9). These results underscore the potential of ML models to augment the objectivity and precision of RAS subtask and skill evaluation. Future research should consider exploring ways to optimize these models, particularly focusing on the classes identified as challenging in this study. Ultimately, this study marks a significant step towards a more refined, objective, and standardized approach to RAS training and competency assessment.

Predictive Modeling of COVID-19 Readmissions: Insights from Machine Learning and Deep Learning Approaches.

This project employs artificial intelligence, including machine learning and deep learning, to assess COVID-19 readmission risk in Malaysia. It offers tools to mitigate healthcare resource strain and enhance patient outcomes. This study outlines a methodology for classifying COVID-19 readmissions. It starts with dataset description and pre-processing, while the data balancing was computed through Random Oversampling, Borderline SMOTE, and Adaptive Synthetic Sampling. Nine machine learning and ten deep learning techniques are applied, with five-fold cross-validation for evaluation. Optuna is used for hyperparameter selection, while the consistency in training hyperparameters is maintained. Evaluation metrics encompass accuracy, AUC, and training/inference times. Results were based on stratified five-fold cross-validation and different data-balancing methods. Notably, CatBoost consistently excelled in accuracy and AUC across all tables. Using ROS, CatBoost achieved the highest accuracy (0.9882 ± 0.0020) with an AUC of 1.0000 ± 0.0000. CatBoost maintained its superiority in BSMOTE and ADASYN as well. Deep learning approaches performed well, with SAINT leading in ROS and TabNet leading in BSMOTE and ADASYN. Decision Tree ensembles like Random Forest and XGBoost consistently showed strong performance.

Machine learning model comparison for freezing of gait prediction in advanced Parkinson's disease.

Freezing of gait (FOG) is a paroxysmal motor phenomenon that increases in prevalence as Parkinson's disease (PD) progresses. It is associated with a reduced quality of life and an increased risk of falls in this population. Precision-based detection and classification of freezers are critical to developing tailored treatments rooted in kinematic assessments. This study analyzed instrumented stand-and-walk (SAW) trials from advanced PD patients with STN-DBS. Each patient performed two SAW trials in their OFF Medication-OFF DBS state. For each trial, gait summary statistics from wearable sensors were analyzed by machine learning classification algorithms. These algorithms include k-nearest neighbors, logistic regression, naïve Bayes, random forest, and support vector machines (SVM). Each of these models were selected for their high interpretability. Each algorithm was tasked with classifying patients whose SAW trials MDS-UPDRS FOG subscore was non-zero as assessed by a trained movement disorder specialist. These algorithms' performance was evaluated using stratified five-fold cross-validation. A total of 21 PD subjects were evaluated (average age 64.24 years, 16 males, mean disease duration of 14 years). Fourteen subjects had freezing of gait in the OFF MED/OFF DBS. All machine learning models achieved statistically similar predictive performance (p < 0.05) with high accuracy. Analysis of random forests' feature estimation revealed the top-ten spatiotemporal predictive features utilized in the model: foot strike angle, coronal range of motion [trunk and lumbar], stride length, gait speed, lateral step variability, and toe-off angle. These results indicate that machine learning effectively classifies advanced PD patients as freezers or nonfreezers based on SAW trials in their non-medicated/non-stimulated condition. The machine learning models, specifically random forests, not only rely on but utilize salient spatial and temporal gait features for FOG classification.

Prediction of cervix cancer stage and grade from diffusion weighted imaging using EfficientNet

Purpose. This study aims to introduce an innovative noninvasive method that leverages a single image for both grading and staging prediction. The grade and the stage of cervix cancer (CC) are determined from diffusion-weighted imaging (DWI) in particular apparent diffusion coefficient (ADC) maps using deep convolutional neural networks (DCNN). Methods. datasets composed of 85 patients having annotated tumor stage (I, II, III, and IV), out of this, 66 were with grade (II and III) and the remaining patients with no reported grade were retrospectively collected. The study was IRB approved. For each patient, sagittal and axial slices containing the gross tumor volume (GTV) were extracted from ADC maps. These were computed using the mono exponential model from diffusion weighted images (b-values = 0, 100, 1000) that were acquired prior to radiotherapy treatment. Balanced training sets were created using the Synthetic Minority Oversampling Technique (SMOTE) and fed to the DCNN. EfficientNetB0 and EfficientNetB3 were transferred from the ImageNet application to binary and four-class classification tasks. Five-fold stratified cross validation was performed for the assessment of the networks. Multiple evaluation metrics were computed including the area under the receiver operating characteristic curve (AUC). Comparisons with Resnet50, Xception, and radiomic analysis were performed. Results. for grade prediction, EfficientNetB3 gave the best performance with AUC = 0.924. For stage prediction, EfficientNetB0 was the best with AUC = 0.931. The difference between both models was, however, small and not statistically significant EfficientNetB0-B3 outperformed ResNet50 (AUC = 0.71) and Xception (AUC = 0.89) in stage prediction, and demonstrated comparable results in grade classification, where AUCs of 0.89 and 0.90 were achieved by ResNet50 and Xception, respectively. DCNN outperformed radiomic analysis that gave AUC = 0.67 (grade) and AUC = 0.66 (stage). Conclusion. the prediction of CC grade and stage from ADC maps is feasible by adapting EfficientNet approaches to the medical context.

Predicting persistent fever in patients with cancer using machine learning.

e13552 Background: Sepsis is the leading cause of hospital readmission and death in the developed world. Cancer patients are at high risk due to their underlying disease and therapies. Meanwhile, antimicrobial resistance increases steadily, warranting a more careful administration of antibiotics. Predicting fever persistence 48 hours after initiation of antibiotics could help reduce antibiotic escalation and chest CT scans in patients where fever is predicted to subside within the next 24 hours. Methods: All cancer patients of the University Hospital Essen (UHE) between 1/1/2020-7/1/2023, receiving broad-spectrum i.v. antibiotics, and having persistent fever after 48 hours of first administration were extracted from UHE’s Fast Health Interoperability Resources clinical database. Vital signs, clinically important covariates, as well as lag features of vital signs and laboratory values were included. The binary target to predict was prevalence of body temperature ≥38 °C after 48-72 hours. Gradient boosting machines, a deep learning neural network, logistic regression (LR), random forest (RF) and a voting ensemble classifier were implemented. Hyperparameter tuning and model validation was performed through five-fold stratified cross validation (80/20% train/test split). Results: Analyzing 2.9 million patient records of UHE we obtained approx. 27,000 hospitalized adults with cancer. Restriction to hematology/oncology wards and intravenous broad-spectrum antibiotics resulted in 1,267 cancer patients. Of these, 264 patients had persistent fever after 48 hours, of which 111 patients had solid cancers, 66 lymphomas, and 87 other hematologic malignancies. 72 hours after antibiotic therapy 160 patients (60.2%) had persistent fever. Based on the area under the precision-recall curve (AUPRC) of 81.7%, the voting classifier model had the highest predictive performance. The model shows a sensitivity of 85%, a specificity of 63%, a Negative Predictive Value (NPV) of 75% and a Positive Predictive Value (PPV) of 73% on a cross-validated test set (Table). Conclusions: Machine learning-based algorithms can predict fever evolution and assist in clinical decision-making, potentially contributing to better patient care by preventing unnecessary antibiotics escalations and CT scans. Based on current model performance, 25% of chest CT scans could be obsolete as patients will not have persistent fever after 48 hours. Additional results will be presented at the meeting. [Table: see text]

A CT-based integrated model for preoperative prediction of occult lymph node metastasis in early tongue cancer.

Occult lymph node metastasis (OLNM) is an essential prognostic factor for early-stage tongue cancer (cT1-2N0M0) and a determinant of treatment decisions. Therefore, accurate prediction of OLNM can significantly impact the clinical management and outcomes of patients with tongue cancer. The aim of this study was to develop and validate a multiomics-based model to predict OLNM in patients with early-stage tongue cancer. The data of 125 patients diagnosed with early-stage tongue cancer (cT1-2N0M0) who underwent primary surgical treatment and elective neck dissection were retrospectively analyzed. A total of 100 patients were randomly assigned to the training set and 25 to the test set. The preoperative contrast-enhanced computed tomography (CT) and clinical data on these patients were collected. Radiomics features were extracted from the primary tumor as the region of interest (ROI) on CT images, and correlation analysis and the least absolute shrinkage and selection operator (LASSO) method were used to identify the most relevant features. A support vector machine (SVM) classifier was constructed and compared with other machine learning algorithms. With the same method, a clinical model was built and the peri-tumoral and intra-tumoral images were selected as the input for the deep learning model. The stacking ensemble technique was used to combine the multiple models. The predictive performance of the integrated model was evaluated for accuracy, sensitivity, specificity, and the area under the receiver operating characteristic curve (AUC-ROC), and compared with expert assessment. Internal validation was performed using a stratified five-fold cross-validation approach. Of the 125 patients, 41 (32.8%) showed OLNM on postoperative pathological examination. The integrated model achieved higher predictive performance compared with the individual models, with an accuracy of 84%, a sensitivity of 100%, a specificity of 76.5%, and an AUC-ROC of 0.949 (95% CI [0.870-1.000]). In addition, the performance of the integrated model surpassed that of younger doctors and was comparable to the evaluation of experienced doctors. The multiomics-based model can accurately predict OLNM in patients with early-stage tongue cancer, and may serve as a valuable decision-making tool to determine the appropriate treatment and avoid unnecessary neck surgery in patients without OLNM.

Application of k-Nearest Neighbors Method for Drug Concentration and Cardiotoxicity Classification Using Extracellular Field Potentials and Reconstructed Action Potentials of Cardiac Cells

Micro-electrode array (MEA) systems are important for measuring extracellular field potentials (FP) of cardiac cells, which is a crucial step in cardiotoxicity assessment. However, without modification, the MEA system is only capable of recording FPs. This limits the number of parameters for cardiotoxicity assessment only to FP parameters, while the action potential (AP) parameters remain unused. To address this issue the MEA systems are often modified to use electro- or optoporation to record the local extracellular APs (LEAPs), which allows to reliably quantify the AP morphology. As an alternative to MEA modification and cell membrane stimulation the AP can be reconstructed mathematically.This study explores how using additional parameters from reconstructed action potentials (RAPs), derived from FPs, can improve the accuracy of k-NN machine learning models for drug concentration and potential cardiotoxicity classification. The k-NN classifier was trained using combinations of FP and RAP parameters. The k-NN models were evaluated using five-fold stratified cross-validation and cross-channel validation. Their performances were compared using error rate, macro precision, macro recall and macro F1 score accuracy metrics. The results indicated that ncorporating RAP parameters into the feature set increased the F1 score of k-NN model for DMSO concentration classification by up to 10.78% compared to the training set with only FP features.

Multi-parametric assessment of cardiac magnetic resonance images to distinguish myocardial infarctions: A tensor-based radiomics feature.

This study assessed the myocardial infarction (MI) using a novel fusion approach (multi-flavored or tensor-based) of multi-parametric cardiac magnetic resonance imaging (CMRI) at four sequences; T1-weighted (T1W) in the axial plane, sense-balanced turbo field echo (sBTFE) in the axial plane, late gadolinium enhancement of heart short axis (LGE-SA) in the sagittal plane, and four-chamber views of LGE (LGE-4CH) in the axial plane. After considering the inclusion and exclusion criteria, 115 patients (83 with MI diagnosis and 32 as healthy control patients), were included in the present study. Radiomic features were extracted from the whole left ventricular myocardium (LVM). Feature selection methods were Least Absolute Shrinkage and Selection Operator (Lasso), Minimum Redundancy Maximum Relevance (MRMR), Chi-Square (Chi2), Analysis of Variance (Anova), Recursive Feature Elimination (RFE), and SelectPersentile. The classification methods were Support Vector Machine (SVM), Logistic Regression (LR), and Random Forest (RF). Different metrics, including receiver operating characteristic curve (AUC), accuracy, F1- score, precision, sensitivity, and specificity were calculated for radiomic features extracted from CMR images using stratified five-fold cross-validation. For the MI detection, Lasso (as the feature selection) and RF/LR (as the classifiers) in sBTFE sequences had the best performance (AUC: 0.97). All features and classifiers of T1 + sBTFE sequences with the weighted method (as the fused image), had a good performance (AUC: 0.97). In addition, the results of the evaluated metrics, especially mean AUC and accuracy for all models, determined that the T1 + sBTFE-weighted fused method had strong predictive performance (AUC: 0.93±0.05; accuracy: 0.93±0.04), followed by T1 + sBTFE-PCA fused method (AUC: 0.85±0.06; accuracy: 0.84±0.06). Our selected CMRI sequences demonstrated that radiomics analysis enables to detection of MI accurately. Among the investigated sequences, the T1 + sBTFE-weighted fused method with the highest AUC and accuracy values was chosen as the best technique for MI detection.

A semantic fidelity interpretable-assisted decision model for lung nodule classification.

Early diagnosis of lung nodules is important for the treatment of lung cancer patients, existing capsule network-based assisted diagnostic models for lung nodule classification have shown promising prospects in terms of interpretability. However, these models lack the ability to draw features robustly at shallow networks, which in turn limits the performance of the models. Therefore, we propose a semantic fidelity capsule encoding and interpretable (SFCEI)-assisted decision model for lung nodule multi-class classification. First, we propose multilevel receptive field feature encoding block to capture multi-scale features of lung nodules of different sizes. Second, we embed multilevel receptive field feature encoding blocks in the residual code-and-decode attention layer to extract fine-grained context features. Integrating multi-scale features and contextual features to form semantic fidelity lung nodule attribute capsule representations, which consequently enhances the performance of the model. We implemented comprehensive experiments on the dataset (LIDC-IDRI) to validate the superiority of the model. The stratified fivefold cross-validation results show that the accuracy (94.17%) of our method exceeds existing advanced approaches in the multi-class classification of malignancy scores for lung nodules. The experiments confirm that the methodology proposed can effectively capture the multi-scale features and contextual features of lung nodules. It enhances the capability of shallow structure drawing features in capsule networks, which in turn improves the classification performance of malignancy scores. The interpretable model can support the physicians' confidence in clinical decision-making.

OPTIMIZING ULTRASOUND IMAGE CLASSIFICATION THROUGH TRANSFER LEARNING: FINE-TUNING STRATEGIES AND CLASSIFIER IMPACT ON PRE-TRAINED INNER-LAYERS

Transfer Learning (TL) is a popular deep learning technique used in medical image analysis, especially when data is limited. It leverages pre-trained knowledge from State-Of-The-Art (SOTA) models and applies it to specific applications through Fine-Tuning (FT). However, fine-tuning large models can be time-consuming, and determining which layers to use can be challenging. This study explores different fine-tuning strategies for five SOTA models (VGG16, VGG19, ResNet50, ResNet101, and InceptionV3) pre-trained on ImageNet. It also investigates the impact of the classifier by using a linear SVM for classification. The experiments are performed on four open-access ultrasound datasets related to breast cancer, thyroid nodules cancer, and salivary glands cancer. Results are evaluated using a five-fold stratified cross-validation technique, and metrics like accuracy, precision, and recall are computed. The findings show that fine-tuning 15% of the last layers in ResNet50 and InceptionV3 achieves good results. Using SVM for classification further improves overall performance by 6% for the two best-performing models. This research provides insights into fine-tuning strategies and the importance of the classifier in transfer learning for ultrasound image classification.

Machine Learning Model with Computed Tomography Radiomics and Clinicobiochemical Characteristics Predict the Subtypes of Patients with Primary Aldosteronism

Adrenal venous sampling (AVS) is the primary method for differentiating between primary aldosterone (PA) subtypes. The aim of study is to develop prediction models for subtyping of patients with PA using computed tomography (CT) radiomics and clinicobiochemical characteristics associated with PA. This study retrospectively enrolled 158 patients with PA who underwent AVS between January 2014 and March 2021. Neural network machine learning models were developed using a two-stage analysis of triple-phase abdominal CT and clinicobiochemical characteristics. In the first stage, the models were constructed to classify unilateral or bilateral PA; in the second stage, they were designed to determine the predominant side in patients with unilateral PA. The final proposed model combined the best-performing models from both stages. The model's performance was evaluated using repeated stratified five-fold cross-validation. We employed paired t-tests to compare its performance with the conventional imaging evaluations made by radiologists, which categorize patients as either having bilateral PA or unilateral PA on one side. In the first stage, the integrated model that combines CT radiomic and clinicobiochemical characteristics exhibited the highest performance, surpassing both the radiomic-alone and clinicobiochemical-alone models. It achieved an accuracy and F1 score of 80.6%±3.0% and 74.8%±5.2% (area under the receiver operating curve [AUC] = 0.778±0.050). In the second stage, the accuracy and F1 score of the radiomic-based model were 88%±4.9% and 81.9%±6.2% (AUC=0.831±0.087). The proposed model achieved an accuracy and F1 score of 77.5%±3.9% and 70.5%±7.1% (AUC=0.771±0.046) in subtype diagnosis and lateralization, surpassing the accuracy and F1 score achieved by radiologists' evaluation (p<.05). The proposed machine learning model can predict the subtypes and lateralization of PA. It yields superior results compared toconventional imaging evaluation and has potential to supplement the diagnostic process in PA.

Development of postoperative delirium prediction models in patients undergoing cardiovascular surgery using machine learning algorithms

Associations between delirium and postoperative adverse events in cardiovascular surgery have been reported and the preoperative identification of high-risk patients of delirium is needed to implement focused interventions. We aimed to develop and validate machine learning models to predict post-cardiovascular surgery delirium. Patients aged ≥ 40 years who underwent cardiovascular surgery at a single hospital were prospectively enrolled. Preoperative and intraoperative factors were assessed. Each patient was evaluated for postoperative delirium 7 days after surgery. We developed machine learning models using the Bernoulli naive Bayes, Support vector machine, Random forest, Extra-trees, and XGBoost algorithms. Stratified fivefold cross-validation was performed for each developed model. Of the 87 patients, 24 (27.6%) developed postoperative delirium. Age, use of psychotropic drugs, cognitive function (Mini-Cog < 4), index of activities of daily living (Barthel Index < 100), history of stroke or cerebral hemorrhage, and eGFR (estimated glomerular filtration rate) < 60 were selected to develop delirium prediction models. The Extra-trees model had the best area under the receiver operating characteristic curve (0.76 [standard deviation 0.11]; sensitivity: 0.63; specificity: 0.78). XGBoost showed the highest sensitivity (AUROC, 0.75 [0.07]; sensitivity: 0.67; specificity: 0.79). Machine learning algorithms could predict post-cardiovascular delirium using preoperative data.Trial registration: UMIN-CTR (ID; UMIN000049390).

DecOM: similarity-based microbial source tracking of ancient oral samples using k-mer-based methods

BackgroundThe analysis of ancient oral metagenomes from archaeological human and animal samples is largely confounded by contaminant DNA sequences from modern and environmental sources. Existing methods for Microbial Source Tracking (MST) estimate the proportions of environmental sources, but do not perform well on ancient metagenomes. We developed a novel method called decOM for Microbial Source Tracking and classification of ancient and modern metagenomic samples using k-mer matrices.ResultsWe analysed a collection of 360 ancient oral, modern oral, sediment/soil and skin metagenomes, using stratified five-fold cross-validation. decOM estimates the contributions of these source environments in ancient oral metagenomic samples with high accuracy, outperforming two state-of-the-art methods for source tracking, FEAST and mSourceTracker.ConclusionsdecOM is a high-accuracy microbial source tracking method, suitable for ancient oral metagenomic data sets. The decOM method is generic and could also be adapted for MST of other ancient and modern types of metagenomes. We anticipate that decOM will be a valuable tool for MST of ancient metagenomic studies.6i9WtiAouxwxZ9D7DN1ayhVideo

Machine learning model for predicting late recurrence of atrial fibrillation after catheter ablation

Late recurrence of atrial fibrillation (LRAF) in the first year following catheter ablation is a common and significant clinical problem. Our study aimed to create a machine-learning model for predicting arrhythmic recurrence within the first year since catheter ablation. The study comprised 201 consecutive patients (age: 61.8 ± 8.1; women 36%) with paroxysmal, persistent, and long-standing persistent atrial fibrillation (AF) who underwent cryoballoon (61%) and radiofrequency ablation (39%). Five different supervised machine-learning models (decision tree, logistic regression, random forest, XGBoost, support vector machines) were developed for predicting AF recurrence. Further, SHapley Additive exPlanations were derived to explain the predictions using 82 parameters based on clinical, laboratory, and procedural variables collected from each patient. The models were trained and validated using a stratified fivefold cross-validation, and a feature selection was performed with permutation importance. The XGBoost model with 12 variables showed the best performance on the testing cohort, with the highest AUC of 0.75 [95% confidence interval 0.7395, 0.7653]. The machine-learned model, based on the easily available 12 clinical and laboratory variables, predicted LRAF with good performance, which may provide a valuable tool in clinical practice for better patient selection and personalized AF strategy following the procedure.

An optimized training approach for meteor detection with an attention mechanism to improve robustness on limited data

Researchers have extensively used convolutional neural networks to detect meteor falls on Earth. However, when dealing with limited available data, these networks may need more robustness to classify new real-world images correctly. This study proposes an optimized training approach of a pre-trained model with an attention mechanism to achieve better generalization results in such a scenario. We compare two architectures, an optimized base model and another version with an attention mechanism. Furthermore, we present a new and publicly available optical meteor dataset that merges several public data sources. We used the merged dataset to train classification models combined with a stratified five-fold cross-validation strategy to determine the reliability of the prediction. The experimental results from both architectures showed good and similar performance. To further determine the best architecture, we performed an additional analysis with visual explanations in new observations. The architecture with an attention mechanism was the best model achieving a false alarm ratio of 2.6% and an accuracy of 97%.

An explainable AI approach for diagnosis of COVID-19 using MALDI-ToF mass spectrometry

Current artificial intelligence (AI) applications for the diagnosis of coronavirus disease 2019 (COVID-19) often lack a biological foundation in the decision-making process. In this study, we have employed AI for COVID-19 diagnosis using mass spectrometry (MS) data and leveraged explainable AI (X-AI) to explain the decision-making process on a local (per-sample) and global (all samples) basis. We first assessed eight machine learning models with five feature engineering techniques using a five-fold stratified cross-validation. The best accuracy was achieved by Random Forest (RF) classifier using the ratio of areas under the curve (AUC) from the MS data as features. These features were chosen on the basis of tentatively representing both human and viral proteins in human gargle samples. We evaluated the RF classifier on a 70%−30% train-test split strategy of 152 human gargle samples, yielding an accuracy of 94.12% on the test dataset. Employing X-AI, we further interpreted the RF model using shapely additive explanations (SHAP) and feature importance techniques, including permutation and impurity-based feature importances. With these interpretation models offering a local and global explanation for the machine learning model decisions, we devised a straightforward, three-stage X-AI framework that can enable medical practitioners to understand the mechanisms of a black-box AI model. To the medical practitioner, this instills trust in the AI model by providing the rationales for its decisions.

Machine learning based model to diagnose obstructive coronary artery disease using calcium scoring, PET imaging, and clinical data

IntroductionAccurate risk stratification in patients with suspected stable coronary artery disease is essential for choosing an appropriate treatment strategy. Our aim was to develop and validate a machine learning (ML) based model to diagnose obstructive CAD (oCAD). MethodWe retrospectively have included 1007 patients without a prior history of CAD who underwent CT-based calcium scoring (CACS) and a Rubidium-82 PET scan. The entire dataset was split 4:1 into a training and test dataset. An ML model was developed on the training set using fivefold stratified cross-validation. The test dataset was used to compare the performance of expert readers to the model. The primary endpoint was oCAD on invasive coronary angiography (ICA). ResultsROC curve analysis showed an AUC of 0.92 (95% CI 0.90-0.94) for the training dataset and 0.89 (95% CI 0.84-0.93) for the test dataset. The ML model showed no significant differences as compared to the expert readers (p ≥ 0.03) in accuracy (89% vs. 88%), sensitivity (68% vs. 69%), and specificity (92% vs. 90%). ConclusionThe ML model resulted in a similar diagnostic performance as compared to expert readers, and may be deployed as a risk stratification tool for obstructive CAD. This study showed that utilization of ML is promising in the diagnosis of obstructive CAD.

Using a CT-based scale to evaluate disease extension and the resectability of locally advanced thyroid cancer.

To establish a computed tomography (CT)-based scale to evaluate the resectability of locally advanced thyroid cancer. This twin-centre retrospective study included 95 locally advanced thyroid cancer patients from the 1st centre as the training cohort and 31 patients from the 2nd centre as the testing cohort, who were categorised into the resectable and unresectable groups. Three radiologists scored the CT scans of each patient by evaluating the extension to the recurrent laryngeal nerve (RLN), trachea, oesophagus, artery, vein, soft tissue, and larynx. A 14-score scale (including all comprised structures) and a 12-score scale (excluding larynx) were developed. Receiver-operating characteristic (ROC) analysis was used to evaluate the performance of the scales. Stratified fivefold cross-validation and external verification were used to validate the scale. In the training cohort, compromised RLN (p < 0.001), trachea (p = 0.001), oesophagus (p = 0.002), artery (p < 0.001), vein (p = 0.005), and soft tissue (p < 0.001) were predictors for unresectability, while compromised larynx (p = 0.283) was not. The 12-score scale (AUC = 0.882, 95%CI: 0.812-0.952) was not inferior to the 14-score scale (AUC = 0.891, 95%CI: 0.823-0.960). In subgroup analysis, the AUCs of the 12-score scale were 0.826 for treatment-naïve patients and 0.976 for patients with prior surgery. The 12-score scale was further validated with a fivefold cross-validation analysis, with an overall accuracy of 78.9-89.4%. Finally, external validation using the testing cohort showed an AUC of 0.875. The researchers built a CT-based 12-score scale to evaluate the resectability of locally advanced thyroid cancer. Validation with a larger sample size is required to confirm the efficacy of the scale. This 12-score CT scale would help clinicians evaluate the resectability of locally advanced thyroid cancer. • The researchers built a 12-score CT scale (including recurrent laryngeal nerve, trachea, oesophagus, artery, vein, and soft tissue) to evaluate the resectability of locally advanced thyroid cancer. • This scale has the potential to help clinicians make treatment plans for locally advanced thyroid cancer.

Radiomics and artificial neural networks modelling for identification of high-risk carotid plaques.

In this study, we aimed to investigate the classification of symptomatic plaques by evaluating the models generated via two different approaches, a radiomics-based machine learning (ML) approach, and an end-to-end learning approach which utilized deep learning (DL) techniques with several representative model frameworks. We collected high-resolution magnetic resonance imaging (HRMRI) data from 104 patients with carotid artery stenosis, who were diagnosed with either symptomatic plaques (SPs) or asymptomatic plaques (ASPs), in two medical centers. 74 patients were diagnosed with SPs and 30 patients were ASPs. Sampling Perfection with Application-optimized Contrasts (SPACE) by using different flip angle Evolutions was used for MRI imaging. Repeated stratified five-fold cross-validation was used to evaluate the accuracy and receiver operating characteristic (ROC) of the trained classifier. The two proposed approaches were investigated to train the models separately. The difference in the model performance of the two proposed methods was quantitatively evaluated to find a better model to differentiate between SPs and ASPs. 3D-SE-Densenet-121 model showed the best performance among all prediction models (AUC, accuracy, precision, sensitivity, and F1-score of 0.9300, 0.9308, 0.9008, 0.8588, and 0.8614, respectively), which were 0.0689, 0.1119, 0.1043, 0.0805, and 0.1089 higher than the best radiomics-based ML model (MLP). Decision curve analysis showed that the 3D-SE-Densenet-121 model delivered more net benefit than the best radiomics-based ML model (MLP) with a wider threshold probability. The DL models were able to accurately differentiate between symptomatic and asymptomatic carotid plaques with limited data, which outperformed radiomics-based ML models in identifying symptomatic plaques.