Performance Of Model Research Articles

Abstract 4119992: 12-lead electrocardiograms predict adverse cardiovascular outcomes of emergency department patients

Introduction: Chest pain is a common cause for presentation to the emergency department (ED). Identifying patients at increased risk of adverse cardiovascular outcomes may help guide testing. Aims: We sought to determine if deep learning enabled 12-lead electrocardiogram (ECG) analysis may improve prediction of adverse cardiovascular outcomes of ED patients presenting with chest pain compared to conventional approaches. Methods: A pretrained convolutional neural network was applied to derive numerical representations of the index ECG waveforms from patients presenting to the Brigham and Women’s Hospital (BWH) ED with chest pain. We trained a neural network model (‘CP-AI’) to predict 30-day major adverse cardiovascular events (MACE) defined as myocardial infarction (MI), pulmonary embolism (PE), aortic dissection (AD), or mortality using age, sex, cardiac biomarkers (troponin and d-dimer), and the numerical ECG representations as inputs ( Figure 1 ). In a holdout sample of Massachusetts General Hospital (MGH) ED patients, we compared the performance of the CP-AI model to alternative models fit on (1) demographics (‘Age and Sex’) (2) demographics + cardiac biomarkers (‘Biomarker Model’), and (3) demographics + numerical ECG representations (‘ECG Model’). Results: The BWH derivation sample included 18,811 individuals (51% female, mean age 65 ± 16 years). The MGH validation sample included 14,476 individuals (45% female, mean age 65 ± 15 years). The CP-AI model significantly outperformed the comparison models for prediction of 30-day MACE with an area under the receiver operating characteristic curve (AUROC) of 0.82 [95% CI 0.81-0.83] ( Figure 1 ). CP-AI also significantly outperformed the comparison models for classification of the 30-day MACE components including MI (AUROC 0.91 [95% CI 0.90-0.92]), PE (AUROC 0.74 [95% CI 0.72-0.76]), and mortality (AUROC 0.86 [95% CI 0.85-0.87]), but not aortic dissection (0.71 [95% CI 0.68-0.75]). Conclusions: The integration of deep learning ECG analysis with conventional assessment improved prediction of adverse cardiovascular of ED patients presenting with chest pain.

Film image processing and production based on high-performance calculation

In film and television production, efficient and precise image processing is vital for achieving realistic visual effects. Therefore, exploring and applying advanced image processing technologies has become an essential method for elevating the production quality of film and television projects. This work investigates the application of artificial intelligence (AI) technology in the processing and production of animated images in film and television scenarios. By comparing the performance of standard Generative Adversarial Network (GAN), DenseNet, and CycleGAN models under different noise conditions, it is found that CycleGAN performs the best in image denoising and detail restoration. Experimental results demonstrate that CycleGAN achieves a Peak Signal-to-noise Ratio (PSNR) of 30.1dB and a Structural Similarity Index Measure (SSIM) of 0.88 under Gaussian noise conditions. Moreover, CycleGAN achieves a PSNR of 29.5dB and an SSIM of 0.85 under salt-and-pepper noise conditions. It outperforms the other models in both conditions. Additionally, CycleGAN’s mean absolute error is significantly lower than that of the other models. This work demonstrates that CycleGAN can more effectively handle complex noise and generate high-quality images under unsupervised learning conditions. These findings provide new directions for future image processing research and offer important references for model selection in practical applications. This work not only offers new perspectives on the development of animation image processing technology but also establishes a theoretical foundation for applying advanced AI techniques in film and television production. Through comparative analysis of various deep learning models, this work highlights the superior performance of CycleGAN under complex noise conditions. This advancement not only drives progress in image processing technology but also provides effective solutions for efficient production and quality enhancement of future film and television works.

Evaluation of Drug–Drug Interactions Between Clarithromycin and Direct Oral Anticoagulants Using Physiologically Based Pharmacokinetic Models

Objective: This study assessed the pharmacokinetic (PK) interactions between clarithromycin (a P-glycoprotein [P-gp] inhibitor) and four direct oral anticoagulants (DOACs) (P-gp substrates) using physiologically based PK (PBPK) models to elucidate the influence of P-gp in the interaction between them. Methods: PBPK models for clarithromycin, DABE–dabigatran (DAB), rivaroxaban, apixaban, and edoxaban were constructed using GastroPlus™ (version 9.9), based on physicochemical data and PK parameters from the literature. The models were optimized and validated in healthy subjects. We evaluated the predictive performance of the established model and further assessed the impact of P-gp on the PK of the four DOACs. Successfully validated models were then used to evaluate potential drug–drug interactions (DDIs) between clarithromycin and the DOACs. Results: The established PBPK models accurately described the PK of clarithromycin, DABE–DAB, rivaroxaban, apixaban, and edoxaban. The predicted PK parameters (Cmax, Tmax, AUC0-t) were within 0.5–2 times the observed values. A sensitivity analysis of P-gp parameters indicated that an increase in P-gp expression was reduced by in vivo exposure to DOACs. The models demonstrated good predictive ability for DDIs between clarithromycin and the anticoagulants, and the ratio of the predicted values to the observed values of Cmax and the area under the curve (AUC) in the DDI state was within the range of 0.5–2. Conclusions: Comprehensive PBPK models for clarithromycin, DABE–DAB, rivaroxaban, apixaban, and edoxaban were developed, which can effectively predict DDIs mediated by P-gp’s function. These models provide theoretical support for clinical dose adjustments and serve as a foundation for future PBPK model development for DOACs under specific pathological conditions.

Diversifying Multi-Head Attention in the Transformer Model

Recent studies have shown that, due to redundancy, some heads of the Transformer model can be pruned without diminishing the efficiency of the model. In this paper, we propose a constrained optimization algorithm based on Hebbian learning, which trains specific layers in the Transformer architecture in order to enforce diversification between the different heads in the multi-head attention module. The diversification of the heads is achieved through a single-layer feed-forward neural network that is added to the Transformer architecture and is trained with the proposed algorithm. We utilize the algorithm in three different architectural variations of the baseline Transformer model. In addition to the diversification of the heads, the proposed methodology can be used to prune the heads that capture redundant information. Experiments on diverse NLP tasks, including machine translation, text summarization, question answering and large language modeling, show that our proposed approach consistently improves the performance of baseline Transformer models.

Intelligent water quality prediction system with a hybrid CNN–LSTM model

ABSTRACT Water pollution remains a longstanding challenge globally, prompting substantial investment in water quality protection. The integration of advanced machine learning models offers promising avenues for accurate water quality prediction, enabling proactive measures to safeguard water sources. Presently, water quality assessment relies predominantly on physical and chemical metrics. This study developed MultiLayer Perceptron (MLP), eXtreme Gradient Boosting (XGBoost), long short-term memory (LSTM), and a hybrid CNN–LSTM model to forecast pH and dissolved oxygen (DO) levels. Results demonstrated the hybrid model's superior performance, with mean squared errors (MSEs) of 0.0015 and 0.0361 for pH and DO prediction, respectively. For Water Quality Classification (WQC), Random Forest (RF), k-Nearest Neighbors (kNNs), Support Vector Machine (SVM), and Light gradient Boosting Machine (Light GBM) were employed, with SVM achieving the highest accuracy at 88.75%. The research underscores the effectiveness of the CNN–LSTM model in predicting pH and DO levels. Leveraging these predictions as inputs to the SVM model offers valuable insights, particularly in regions where conventional monitoring methods face limitations. This streamlined approach, requiring only two parameters, signifies a significant advancement in accurate water quality prediction.

Analysing crash severity on expressways in India using statistical and ML approaches

The high injury severity of traffic crashes on Indian expressways is a significant concern for road safety experts, although studies dedicated to this critical issue are limited. The present study investigates the factors influencing crash severity using multinomial logit (MNL), decision tree (DT) and random forest (RF) models on a dataset of 2,747 crashes on three selected expressways. The dependent variable, crash severity, had four injury severity categories: fatal, severe, minor, and property damage only (PDO). Various explanatory variables, included traffic and speed characteristics, temporal and geometric characteristics, primary contributing factors, crash type, and the vehicles involved. Synthetic minority oversampling (SMOTE) and randomise class balancing (RCB) techniques were also employed to tackle the class imbalance issue in the dataset. The predictive performance of models was evaluated using classification accuracy and Kappa value. The RF model showed the highest predictive accuracy on the RCB dataset. The key findings highlight the critical need for enforcement of speed limits and entry restrictions, lane discipline, and improvement of design deficiencies such as provisions of truck laybys and bus bays. These measures can help in policy-making and engineering improvements to enhance road safety on expressways in India.

Using advanced machine learning algorithms to predict academic major completion: A cross-sectional study

BackgroundExisting prediction methods for academic majors based on personality traits have notable gaps, including limited model complexity and generalizability.The current study aimed to utilize advanced Machine Learning (ML) algorithms with smoothing functions to predict academic majors completed based on personality subscales. MethodsWe used reports from 59,413 individuals to perform the current study. All advanced algorithms implemented in this article were based on R software (version 4.1.3, R Core Team, 2021). All model parameters were optimized based on resampling and cross-validation (CV). In addition, pseudo-R2 as a robust metric has been used to compare the performance of models, which, unlike most studies, considers the quality of model-predicted probabilities. ResultThe results indicated that advanced ML models' performance on training and test data was superior to logistic regression. Pseudo-R2 and AUC results showed that advanced models such as kNN, GBE, and RF had the highest scores based on test data compared to other models. The pseudo-R2 values for the models used in this study varied across the test dataset; the lowest value belonged to the logistic regression algorithm at .022, and the highest value was recorded for the kNN algorithm at .099. The agreeableness subscale is the most influential component in predicting the completion of university education, followed by conscientiousness and emotional stability. ConclusionThe potential of advanced methods to enhance the accuracy and validity of predictions is a promising development in our field. Their performance, particularly in handling large data sets with complex patterns, is a reason for optimism about the future of research in this area.

Randomising spatial patterns supports the integration of intraspecific variation in ecological niche models

Ecological niche models (ENMs) are an essential modelling technique in biodiversity prediction and conservation and are frequently used to forecast species responses to global changes. Classic species‐level models may show limitations as they assume species homogeneity, neglecting intraspecific variation. Composite ENMs allow the integration of intraspecific variation by combining intraspecific‐level ENMs, capturing individual environmental responses over the species' geographic range. While recent studies suggest that accounting for intraspecific variation improves model predictions, we currently lack methods to test the significance of the improvement. Here, we propose a null model approach that randomises observed intraspecific structures as an appropriate baseline for comparison. We illustrate this approach by comparing predictive performance of a species‐level ENM to composite ENMs for European beech Fagus sylvatica. To investigate the influence of spatial lineage structure, we tested all models against the same withheld data to allow comparison across models based on five common performance metrics. We found that the species‐level ENM expressed higher overall performance (i.e. AUC, TSS, and Boyce index) and specificity (ability to predict absences), while the composite ENMs achieved higher sensitivity (ability to predict presences). In line with this, the composite ENMs also showed increased sensitivity and decreased specificity compared to the null models that randomised lineage structure. We showed that the assessment of model performance strongly varies based on the used measures, emphasising a careful investigation of multiple measures for evaluation. The application of null models allowed us to disentangle the effect of observed patterns of intraspecific variation in ENMs. Further, we highlight the validation and use of well‐founded subgroups for modelling. Although intraspecific variation improves the prediction of occurrences of European beech, it did not fully outcompete the classic species‐level model and should be used with care and rather to improve understanding and to supplement, not replace, species‐level models.

Empowering Data Sharing in Neuroscience: A Deep Learning De-identification Method for Pediatric Brain MRIs.

Privacy concerns, such as identifiable facial features within brain scans, have hindered the availability of pediatric neuroimaging datasets for research. Consequently, pediatric neuroscience research lags adult counterparts, particularly in rare disease and under-represented populations. The removal of face regions (image defacing) can mitigate this, however existing defacing tools often fail with pediatric cases and diverse image types, leaving a critical gap in data accessibility. Given recent NIH data sharing mandates, novel solutions are a critical need. To develop an AI-powered tool for automatic defacing of pediatric brain MRIs, deep learning methodologies (nnU-Net) were employed using a large, diverse multi-institutional dataset of clinical radiology images. This included multi-parametric MRIs (T1w, T1w-contrast enhanced, T2w, T2w-FLAIR) with 976 total images from 208 brain tumor patients (Children's Brain Tumor Network, CBTN) and 36 clinical control patients (Scans with Limited Imaging Pathology, SLIP) ranging in age from 7 days to 21 years old. Face and ear removal accuracy for withheld testing data was the primary measure of model performance. Potential influences of defacing on downstream research usage were evaluated with standard image processing and AI-based pipelines. Group-level statistical trends were compared between original (non-defaced) and defaced images. Across image types, the model had high accuracy for removing face regions (mean accuracy, 98%; N=98 subjects/392 images), with lower performance for removal of ears (73%). Analysis of global and regional brain measures (SLIP cohort) showed minimal differences between original and defaced outputs (mean r S=0.93, all p < 0.0001). AI-generated whole brain and tumor volumes (CBTN cohort) and temporalis muscle metrics (volume, cross-sectional area, centile scores; SLIP cohort) were not significantly affected by image defacing (all r S>0.9, p<0.0001). The defacing model demonstrates efficacy in removing facial regions across multiple MRI types and exhibits minimal impact on downstream research usage. A software package with the trained model is freely provided for wider use and further development (pediatric-auto-defacer; https://github.com/d3b-center/pediatric-auto-defacer-public). By offering a solution tailored to pediatric cases and multiple MRI sequences, this defacing tool will expedite research efforts and promote broader adoption of data sharing practices within the neuroscience community. AI = artificial intelligence; CBTN = Children's Brain Tumor Network; CSA = cross-sectional area; SLIP = Scans with Limited Imaging Pathology; TMT = temporalis muscle thickness; NIH = National Institute of Health; LH = left hemisphere; RH = right hemisphere.

Abstract Su706: New risk markers of sudden cardiac death in the young : insights from machine learning on EHR data

Background: Sudden Cardiac Death (SCD) in young adults is a significant public health issue, with survival rates often less than 15%. Effective prevention is critical as many SCD cases go undetected due to nonspecific symptoms and ineffective risk stratification methods, especially among the youth. Objective: To explore risk markers specific to SCD in younger populations using Artificial Intelligence (AI) trained on systematically collected Electronic Health Records (EHR) data, and to compare these markers with those in older populations. Methods: We conducted a retrospective cohort study using data from 46,060 participants in Paris and its inner suburbs, spanning 2011 to 2020. The cohort included 1,100 SCD cases and 1,100 controls aged 40 years or younger, and 21,930 SCD cases and 21,930 controls aged over 40. A machine learning model was developed, trained, and validated on these cohorts. We used the Catboost algorithm and evaluated model performance using AUC, sensitivity, and specificity. The SHAP algorithm was employed to elucidate the relationship between variables and predicted risk. Results: The model achieved on validation sets an AUC of 0.75 (95% CI 0.67-0.83) for the younger cohort and AUC of 0.82 (95% CI: 0.70-0.83), for the older cohort. Neuropsychiatric drugs were the most influential predictors for the younger cohort, whereas age and cardiovascular drugs were most predictive for the older cohort. The results indicate that traditional cardiovascular risk factors play a secondary role in predicting SCD among younger individuals. Conclusion: In analyzing predictors of sudden death in individuals under 40, we developed a personalized SCD prediction model using EHR data tailored to this demographic. Psychiatric medications, and possibly underlying psychiatric conditions, emerged as significant predictors. Our findings emphasize the need for a multidisciplinary approach to SCD prevention, integrating neuropsychiatric and cardiovascular considerations.

Enhancing Cover Management Factor Classification Through Imbalanced Data Resolution

This study addresses the persistent challenge of class imbalance in land use and land cover (LULC) classification within the Shihmen Reservoir watershed in Taiwan, where LULC is used to map the Cover Management factor (C-factor). The dominance of forests in the LULC categories leads to an imbalanced dataset, resulting in poor prediction performance for minority classes when using machine learning techniques. To overcome this limitation, we applied the Synthetic Minority Over-sampling Technique (SMOTE) and the 90-model SMOTE-variants package in Python to balance the dataset. Due to the multi-class nature of the data and memory constraints, 42 models were successfully used to create a balanced dataset, which was then integrated with a Random Forest algorithm for C-factor classification. The results show a marked improvement in model accuracy across most SMOTE variants, with the Selected Synthetic Minority Over-sampling Technique (Selected_SMOTE) emerging as the best-performing method, achieving an overall accuracy of 0.9524 and a sensitivity of 0.6892. Importantly, the previously observed issue of poor minority class prediction was resolved using the balanced dataset. This study provides a robust solution to the class imbalance issue in C-factor classification, demonstrating the effectiveness of SMOTE variants and the Random Forest algorithm in improving model performance and addressing imbalanced class distributions. The success of Selected_SMOTE underscores the potential of balanced datasets in enhancing machine learning outcomes, particularly in datasets dominated by a majority class. Additionally, by addressing imbalance in LULC classification, this research contributes to Sustainable Development Goal 15, which focuses on the protection, restoration, and sustainable use of terrestrial ecosystems.

Abstract 4135726: Prospective validation and implementation pilot study of an emergency department heart failure risk stratification tool: STRIDE-HF

Background: We previously found the STRIDE-HF emergency department (ED) risk tool accurately predicted risk of a 30-day serious adverse event (SAE), including 30-day mortality, cardiopulmonary resuscitation, intra-aortic balloon pump insertion, intubation, new dialysis, myocardial infarction, or coronary revascularization. Research question: We sought to prospectively validate STRIDE-HF and describe safety of the deployed model in a clinical pilot. We hypothesized that performance would decrease slightly with prospective evaluation and that real-time risk display would contribute to safe disposition decisions. Goals: 1) To prospectively validate KPCARES-HF across 21 community EDs among ED patients with acute heart failure (AHF) from January 1, 2023 – December 31 st , 2023; and 2) to describe outcomes among patients predicted to be very low risk and discharged home in the clinical pilot. Methods: For prospective validation, we report 30-day mortality rates by risk strata. We assess model performance using area under the receiver operator curve (AUROC) and sensitivity at a key clinical threshold. We describe rates of 30-day SAE among patients predicted to be very low risk and discharged after ED or observation unit care in the clinical pilot. Results: There were 13,274 ED patients in the prospective validation; median age was 76, 50.8% were female, and 44.5% were non-White. We found 11.4%, 24.8%, 31.9%, and 31.9% of patients were very low, low, moderate, and high risk, respectively, while 21.6%, 15.2% and 63.0% were discharged, observed, and admitted, respectively. Among discharged patients, 29.2% and 13.3% were moderate or high risk, respectively, and 30-day mortality rates among these patients were 2.2% and 8.4%, respectively. Among 8,363 admitted patients, 27.7% were very low or low risk. Conclusions: STRIDE-HF maintained high predictive accuracy with prospective validation in this diverse, multi-center cohort. Our findings suggest that use of STRIDE-HF might help better align risk with admission decision and can be safely used to assist providers in identifying patients who may be stable for outpatient care.

Performance Analysis of ML and DL Models: Impact of Linear and Non-Linear Optimizers on Model Efficiency

Introduction: Emphasizing the effect of linear and non-linear optimizers on the performance of machine learning (ML) and deep learning (DL) models, this work addresses the CIFAR-10 image classification task. The work explores over multiple data sets how optimizers influence model behavior and efficiency. Objectives: With a view on five primary assessment metrics—accuracy, precision, recall, F1-score, and AUC the main purpose is to evaluate on SVM and CNN models the performance of linear optimizers (gradient descent, SGD) and non-linear optimizers (Adam, RMSProp). Methods: The study offers fair and complete comparison by way of a consistent evaluation strategy. Analyzing how different optimizer influences model convergence rates, computational cost, and training stability takes front stage in the experimental setting. Designed on CIFAR-10, optimizers are used on ML and DL models whose performance on all metrics is noted. Results: Results show that non-linear optimizers especially Adam much increase CNN model performance by means of faster convergence, higher classification accuracy, and better model stability. While helpful for simpler models, linear optimizers such SGD show slower convergence and limited adaptation with more complicated data. Conclusions: This study provides direction on selecting appropriate optimizers depending on model complexity, job needs, and computing restrictions coupled with important information on optimizing model training in image classification and other areas.

Abstract 4137796: Novel Deep Learning Aligned Strain Techniques in Repaired Tetralogy of Fallot

Introduction: Multicenter studies show that demographic and CMR-based volumetrics and strain predict death, ventricular tachycardia and fibrillation (DVTF) in repaired tetralogy of Fallot (rTOF). We developed a novel deep learning method to calculate radial (RS) and circumferential strain (CS) from end-diastole (ED) to other key frames: mid systole (MS), end-systole (ES), peak flow (PF) in diastole, and mid-diastole (MD) (called ED2K), as well as from each key frame to the next (called K2K strain), for each left ventricle (LV) segment. Hypothesis: We hypothesized that: H1) ED-ES strain; H2) septal strain; and H3) diastolic strain would be additionally predictive of DVTF as compared to a traditional model. Approach: 704 patients combined from the German Competence Network and INDICATOR cohorts had ED2K and K2K strain values calculated using CMR short axis stack. We first created a 4-variable “traditional” logistic regression model including age at CMR, RVEF%, LVEF% and RVESVi. We then separately added ED2K and K2K parameters to assess increased ability to discriminate DVTF, measured by the c-statistic. Results: In univariate analyses, H1: ED to ES RS and CS; H2: multiple systolic septal and non-septal variables; and H3: overall diastolic RS and CS were significantly associated with DVTF, with c-statistics between 0.65 and 0.73 (n=56, Table 1). When added to the traditional model, H1: ED to ES CS; and H2: septal and non-septal CS slightly improved model performance (Table 2). Model calibration was adequate for all models. Conclusion: Our key-frame specific strain recapitulates the CS prediction of adverse events in rTOF, and slightly improves prediction in a multiparameter model. Future work includes external validation.

Multiple feature selection based on an optimization strategy for causal analysis of health data.

Recent advancements in information technology and wearable devices have revolutionized healthcare through health data analysis. Identifying significant relationships in complex health data enhances healthcare and public health strategies. In health analytics, causal graphs are important for investigating the relationships among health features. However, they face challenges owing to the large number of features, complexity, and computational demands. Feature selection methods are useful for addressing these challenges. In this paper, we present a framework for multiple feature selection based on an optimization strategy for causal analysis of health data. We select multiple health features based on an optimization strategy. First, we define a Weighted Total Score (WTS) index to assess the feature importance after the combination of different feature selection methods. To explore an optimal set of weights for each method, we design a multiple feature selection algorithm integrated with the greedy algorithm. The features are then ranked according to their WTS, enabling selection of the most important ones. After that, causal graphs are constructed based on the selected features, and the statistical significance of the paths is assessed. Furthermore, evaluation experiments are conducted on an experiment dataset collected for this study and an open dataset for diabetes. The results demonstrate that our approach outperforms baseline models by reducing the number of features while improving model performance. Moreover, the statistical significance of the relationships between features uncovered through causal graphs is validated for both datasets. By using the proposed framework for multiple feature selection based on an optimization strategy for causal analysis, the number of features is reduced and the causal relationships are uncovered and validated.

Abstract 4136932: Impact of Different Socioeconomic Metrics on Heart Failure-Related Admission and Short-Term Outcomes in Maryland

Introduction: Annually, over 500,000 Americans are hospitalized due to heart failure (HF), marking it as a major contributor to morbidity and mortality. It also poses a significant financial burden and leads to considerable losses in productivity. Objective: This study investigates the predictive accuracy of different socioeconomic metrics on the risk and outcomes of HF in Maryland. Methodology: A retrospective analysis of the Maryland State Inpatient Database (2016-2020) was conducted to assess the predictive accuracy of race/ethnicity, insurance status, household median income, and neighborhood poverty level (measured by the Distressed Communities Index) on the risk of heart failure-related hospital admissions and outcomes. Multivariate logistic regression models were also used to adjust for confounders. Result: During the study period, a total of 389,220 cases of HF were reported in the Maryland SID. The majority of these patients were white (56.8%) and female (51.1%), with a median age of 73 years (interquartile range [IQR] 62-82 years). The in-hospital mortality rate was 5.1%, while rates of atrial fibrillation, cardiac arrest and prolonged hospital stay were 34.4%, 0.3%, and 48.4%, respectively. Multivariate analysis revealed a substantial area under the ROC curve (AUC) indicating good model performance: 0.88 for predicting HF, 0.64 for atrial fibrillation, 0.64 for cardiac arrest 0.57 for prolonged hospital stays, 0.63 for mortality. Subgroup analyses showed variable predictiveness by race (AUC = 0.4378), payment method (AUC = 0.5754), income quartile (AUC = 0.5202), and deprivation composite score (AUC = 0.4751). Patients with private insurance had the highest risk of stress cardiomyopathy (odds ratio [OR] = 1.98; 95% confidence interval [CI] 1.70-2.29). Socioeconomic metrics, including neighborhood distress, showed varying predictive accuracy for the HF-related admissions and selected short-term outcomes, with the highest predictive accuracy for neighborhood distress on the risk of HF (AUC = 0.50, std: 0.006), atrial fibrillation (AUC = 0.48, std: 0.0007), cardiac arrest (AUC = 0.51, std: 0.007), and prolonged hospital stays (AUC = 0.53, std: 0.0005) and mortality (AUC = 0.50, std: 0.0015). Conclusion: Neighborhood poverty level demonstrates significant predictive power for assessing the risk of HF-related hospital admissions and the short-term outcomes among Maryland residents, exceeding factors like insurance and race/ethnicity.

Early prediction of Mycobacterium tuberculosis transmission clusters using supervised learning models.

Identifying individuals with tuberculosis (TB) with a high risk of onward transmission can guide disease prevention and public health strategies. Here, we train classification models to predict the first sampled isolates in Mycobacterium tuberculosis transmission clusters from demographic and disease data. We find that supervised learning, in particular balanced random forests, can be used to develop predictive models to identify people with TB that are more likely associated with TB cluster growth, with good model performance and AUCs of ≥ 0.75. We also identified the most important patient and disease characteristics in the best performing classification model, including host demographics, site of infection, TB lineage, and age at diagnosis. This framework can be used to develop predictive tools for the early assessment of potential cluster growth to prioritise individuals for enhanced follow-up with the aim of reducing transmission chains.

RADT-12. DERIVING IMAGING BIOMARKERS FOR PRIMARY CENTRAL NERVOUS SYSTEM LYMPHOMA USING DEEP LEARNING

Abstract PURPOSE Primary central nervous system lymphoma (PCNSL) is typically treated with chemotherapy, steroids, and/or whole brain radiotherapy (WBRT). Determining which patients benefit from WBRT after chemotherapy and which are adequately treated with chemotherapy alone remains a clinical challenge. Although WBRT is associated with improved outcomes, it also carries a risk of neurocognitive side effects. This study aims to refine PCNSL patient phenotyping by leveraging deep learning (DL) extracted imaging biomarkers to enable personalized therapy. METHODS Our study included 71 patients treated between 2009-2021. The primary outcome of interest was overall survival (OS) assessed at one- and two-year cutoffs. The DL model leveraged an 8-layer 2D convolutional neural network analyzing individual slices of post-contrast T1-weighted pre-treatment MRI scans. Survival predictions utilized a weighted voting system related to tumor size. Model performance was assessed with accuracy, sensitivity, specificity, and F1 scores. Time-dependent AUCs were calculated and C-statistics were computed to summarize the results. Kaplan-Meier (KM) survival analysis assessed differences between low and high-risk groups and were evaluated using the log-rank test. RESULTS The cohort’s average age was 65.6 years with average OS of 2.80 years. For one-year OS, the model achieved an AUC of 0.73 (0.60-0.85), accuracy of 0.73 (0.61-0.82), sensitivity of 0.72 (0.54–0.85), and specificity of 0.73 (0.58–0.84). For two-year OS, the model achieved an AUC of 0.70 (0.58-0.82), accuracy of 0.70 (0.58-0.79), sensitivity of 0.68 (0.55–0.81), and specificity of 0.71 (0.55–0.84). Model performance is presented with respective 95% confidence intervals. KM curves showed the one and two-year models effectively discriminated between low- and high-risk group, reporting c-statistics of 0.73 (p&amp;lt;.001) and 0.71 (p&amp;lt;.001) respectively. A sub-analysis confirmed consistent model performance across different tumor volumes and focality. CONCLUSION DL classifiers of PCNSL MRIs can stratify patient phenotypes beyond traditional risk paradigms. Given dissensus surrounding PCNSL treatment, DL can augment risk stratification and treatment personalization, especially with regards to WBRT decision making.

Computed tomography radiomics reveals prognostic value of immunophenotyping in laryngeal squamous cell carcinoma: a comparison of whole tumor- versus habitats-based approaches.

To compare the performance of whole tumor and habitats-based computed tomography (CT) radiomics for predicting immunophenotyping in laryngeal squamous cell carcinomas (LSCC) and further evaluate the stratified effect of the radiomics model on disease-free survival (DFS) and overall survival (OS) of LSCC patients. In all, 106 LSCC patients (40 with inflamed and 66 with non-inflamed immunophenotyping) were randomly assigned into a training (n = 53) and testing (n = 53) cohort. Briefly, 750 radiomics features from contrast-enhanced CT images were respectively extracted from the whole tumor and two Otsu method-derived subregions. Intraclass correlation coefficients (ICCs) were calculated to evaluate the reproducibility. The radiomics models for predicting immunophenotyping were respectively created using K-nearest neighbors (KNN), logistic regression (LR), and Naive bayes (NB) classifiers. The performance of models in the testing cohort were compared using area under the curve (AUC). The prognostic value of the optimal model was determined by survival analysis. The radiomics features derived from whole tumor showed better reproducibility than those derived from habitats. The best model for the whole tumor (LR classifier) showed superior performance than that for the habitats (KNN classifier) in the testing cohort, but there were no significant differences (AUC: 0.741 vs. 0.611, p = 0.112). Multivariable Cox regression analysis showed that the immunophenotyping predicted by the optimal model was an independent risk factor of unfavorable DFS (p = 0.009) and OS (p = 0.008) in LSCC patients. Whole tumor-based CT radiomics could serve as a potential predictive biomarker of immunophenotyping and outcome prediction in LSCC patients.

DeepBP: Ensemble deep learning strategy for bioactive peptide prediction.

Bioactive peptides are important bioactive molecules composed of short-chain amino acids that play various crucial roles in the body, such as regulating physiological processes and promoting immune responses and antibacterial effects. Due to their significance, bioactive peptides have broad application potential in drug development, food science, and biotechnology. Among them, understanding their biological mechanisms will contribute to new ideas for drug discovery and disease treatment. This study employs generative adversarial capsule networks (CapsuleGAN), gated recurrent units (GRU), and convolutional neural networks (CNN) as base classifiers to achieve ensemble learning through voting methods, which not only obtains high-precision prediction results on the angiotensin-converting enzyme (ACE) inhibitory peptides dataset and the anticancer peptides (ACP) dataset but also demonstrates effective model performance. For this method, we first utilized the protein language model-evolutionary scale modeling (ESM-2)-to extract relevant features for the ACE inhibitory peptides and ACP datasets. Following feature extraction, we trained three deep learning models-CapsuleGAN, GRU, and CNN-while continuously adjusting the model parameters throughout the training process. Finally, during the voting stage, different weights were assigned to the models based on their prediction accuracy, allowing full utilization of the model's performance. Experimental results show that on the ACE inhibitory peptide dataset, the balanced accuracy is 0.926, the Matthews correlation coefficient (MCC) is 0.831, and the area under the curve is 0.966; on the ACP dataset, the accuracy (ACC) is 0.779, and the MCC is 0.558. The experimental results on both datasets are superior to existing methods, demonstrating the effectiveness of the experimental approach. In this study, CapsuleGAN, GRU, and CNN were successfully employed as base classifiers to implement ensemble learning, which not only achieved good results in the prediction of two datasets but also surpassed existing methods. The ability to predict peptides with strong ACE inhibitory activity and ACPs more accurately and quickly is significant, and this work provides valuable insights for predicting other functional peptides. The source code and dataset for this experiment are publicly available at https://github.com/Zhou-Jianren/bioactive-peptides .