Comparable Accuracy Research Articles

Validation of a Compact and Self-Contained Pyrosequencing Platform for Clinical Screening of RAS Mutations in Thyroid Cancers

Background and Objectives: Accurate screening of clinically significant tumor mutations is critical for precision medicine in oncology. This requires genotyping platforms with high accuracy and compatibility with varying DNA yields from challenging sample types. Here, we have validated a new, improved, compact, and self-contained pyrosequencing platform (Pyromark Q48 Autoprep; Q48) for screening N-, K- and H-RAS mutations in thyroid cancers. Methods: A set of 73 thyroid cancer and 16 non-thyroid cancer samples (fine needle aspirates and formalin-fixed paraffin-embedded) with known mutation status of RAS genes were tested using the Q48 platform. Performance parameters such as accuracy, precision, and limit-of-detection were established. Q48 workflow was compared to an older Q96 pyrosequencing platform to highlight the differences and advantages. RAS testing by pyrosequencing was also compared to a clinically validated next-generation sequencing platform using 56 thyroid cancer samples. Results: The Q48 Pyromark was found to be a very reliable platform suited for quick testing of RAS genes with complete accuracy, high precision, and high detection sensitivity. It had comparable accuracy, with higher sequencing success rates than NGS. The hands-on time, workflow ease, and efficiency were also significantly improved in comparison with the Q96 platform. Conclusions: Overall, the Q48 platform was found to be a well-suited and agile clinical sequencing platform to rapidly screen RAS mutations.

Enhanced Multilinear PCA for Efficient Image Analysis and Dimensionality Reduction: Unlocking the Potential of Complex Image Data

This paper presents an Enhanced Multilinear Principal Component Analysis (EMPCA) algorithm, an improved variant of the traditional Multilinear Principal Component Analysis (MPCA) tailored for efficient dimensionality reduction in high-dimensional data, particularly in image analysis tasks. EMPCA integrates random singular value decomposition to reduce computational complexity while maintaining data integrity. Additionally, it innovatively combines the dimensionality reduction method with the Mask R-CNN algorithm, enhancing the accuracy of image segmentation. Leveraging tensors, EMPCA achieves dimensionality reduction that specifically benefits image classification, face recognition, and image segmentation. The experimental results demonstrate a 17.7% reduction in computation time compared to conventional methods, without compromising accuracy. In image classification and face recognition experiments, EMPCA significantly enhances classifier efficiency, achieving comparable or superior accuracy to algorithms such as Support Vector Machines (SVMs). Additionally, EMPCA preprocessing exploits latent information within tensor structures, leading to improved segmentation performance. The proposed EMPCA algorithm holds promise for reducing image analysis runtimes and advancing rapid image processing techniques.

Predicting Antibody Affinity Changes upon Mutation Based on Unbound Protein Structures

Antibodies are key proteins in the immune system that can reversibly and non-covalently bind specifically to their corresponding antigens, forming antigen–antibody complexes. They play a crucial role in recognizing foreign or self-antigens during the adaptive immune response. Monoclonal antibodies have emerged as a promising class of biological macromolecule therapeutics with broad market prospects. In the process of antibody drug development, a key engineering challenge is to improve the affinity of candidate antibodies, without experimentally resolved structures of the antigen–antibody complexes as input for computer-aided predictive methods. In this work, we present an approach for predicting the effect of residue mutations on antibody affinity without the structures of the antigen–antibody complexes. The method involves the graph representation of proteins and utilizes a pre-trained encoder. The encoder captures the residue-level microenvironment of the target residue on the antibody along with the antigen context pre- and post-mutation. The encoder inherently possesses the potential to identify paratope residues. In addition, we curated a benchmark dataset specifically for mutations of the antibody. Compared to baseline methods based on complex structures and sequences, our approach achieves superior or comparable average accuracy on benchmark datasets. Additionally, we validate its advantage of not requiring antigen–antibody complex structures as input for predicting the effects of mutations in antibodies against SARS-CoV-2, influenza, and human cytomegalovirus. Our method shows its potential for identifying mutations that improve antibody affinity in practical antibody engineering applications.

Development and External Validation of [18F]FDG PET-CT-Derived Radiomic Models for Prediction of Abdominal Aortic Aneurysm Growth Rate

Objective (1): To develop and validate a machine learning (ML) model using radiomic features (RFs) extracted from [18F]FDG PET-CT to predict abdominal aortic aneurysm (AAA) growth rate. Methods (2): This retrospective study included 98 internal and 55 external AAA patients undergoing [18F]FDG PET-CT. RFs were extracted from manual segmentations of AAAs using PyRadiomics. Recursive feature elimination (RFE) reduced features for model optimisation. A multi-layer perceptron (MLP) was developed for AAA growth prediction and compared against Random Forest (RF), XGBoost, and Support Vector Machine (SVM). Accuracy was evaluated via cross-validation, with uncertainty quantified using dropout (MLP), standard deviation (RF), and 95% prediction intervals (XGBoost). External validation used independent data from two centres. Ground truth growth rates were calculated from serial ultrasound (US) measurements or CT volumes. Results (3): From 93 initial RFs, 29 remained after RFE. The MLP model achieved an MAE ± SEM of 1.35 ± 3.2e−4 mm/year with the full feature set and 1.35 ± 2.5e−4 mm/year with RFE. External validation yielded 1.8 ± 8.9e−8 mm/year. RF, XGBoost, and SVM models produced comparable accuracies internally (1.4–1.5 mm/year) but showed higher errors during external validation (1.9–1.97 mm/year). The MLP model demonstrated reduced uncertainty with the full feature set across all datasets. Conclusions (4): An MLP model leveraging [18F]FDG PET-CT radiomics accurately predicted AAA growth rates and generalised well to external data. In the future, more sophisticated stratification could guide individualised patient care, facilitating risk-tailored management of AAAs.

Quantifying and combining uncertainty for improving the behavior of Digital Twin Systems

Abstract Uncertainty is an inherent property of any complex system, especially those that incorporate physical parts or operate in real environments. In this paper, we focus on the Digital Twins of adaptive systems, which are particularly complex to design, verify, and optimize. One of the problems of having two systems (the physical one and its digital replica) is that their behavior may not always be consistent. In addition, both twins are normally subject to different types of uncertainties, which complicates their comparison. In this paper we propose the explicit representation and treatment of the uncertainty of both twins, and show how this enables a more accurate comparison of their behaviors. Furthermore, this allows us to reduce the overall system uncertainty and improve its behavior by properly averaging the individual uncertainties of the two twins. An exemplary incubator system is used to illustrate and validate our proposal.

Integrating simulation and machine learning for accurate preform charge prediction in Sheet Molding Compound manufacturing

The Sheet Molding Compound (SMC) process is essential in high-volume manufacturing of composite structures due to its scalability and efficiency. A primary challenge, however, lies in determining the initial charge shape that ensures complete mold filling without excessive overflow, typically resolved through labor-intensive trial and error. While simulations can anticipate the mold filling outcome, they often lack the capability to fine-tune the initial preform configuration, leading to inefficiencies in both time and material. This study presents an innovative, simulation-driven approach for accurately predicting initial charge shapes for two-dimensional (2D) mold designs. By employing Darcy’s Law and a fixed mesh grid framework, the methodology simulates a reverse material flow to trace the optimal preform shape. A complementary machine learning (ML) model was then developed to predict the preform shapes based on mold geometry, final thickness, and initial charge thickness. Serving as a digital twin of the SMC process, this ML model delivers results with comparable accuracy to simulations, significantly enhancing computational efficiency and avoiding common convergence issues in traditional simulations. This ML-driven digital twin approach also provides a robust proof of concept for addressing initial charge shapes in complex three-dimensional (3D) molds, where the computational demands of reverse flow simulations may present challenges. This combined simulation and ML framework equips manufacturers with a more precise and efficient tool for optimizing SMC processes, minimizing material waste, and reducing production time.

Anisotropic resistivity estimation and uncertainty quantification from borehole triaxial electromagnetic induction measurements: Gradient-based inversion and physics-informed neural network

Rapid and accurate petrophysical reservoir description and quantification is important for subsurface energy resource modeling and engineering. Triaxial borehole resistivity measurements enable the estimation of key in-situ rock properties, such as the volumetric concentration of shale and sandstone resistivity. However, traditional approaches for estimating these properties from triaxial, or anisotropic, borehole resistivity measurements rely on analytical solutions that solve a system of equations simultaneously, which can lead to numerical instability and inefficiency. By reformulating the system of anisotropic resistivity equations as an inverse problem, we can achieve a more stable, accurate, and efficient estimation of key petrophysical properties. We propose two methods, namely nonlinear gradient-based and physics-informed neural network (PINN) inversion, to estimate the volumetric concentration of shale and sandstone resistivity from the parallel- and perpendicular-to-bedding-plane resistivity logs, posed as the solution of an inverse problem. Furthermore, we compare the PINN, nonlinear gradient-based inversion and analytical solution methods in terms of accuracy, computational efficiency, and uncertainty quantification using four different datasets of anisotropic resistivity logs, i.e., two synthetic and two field cases. The PINN inversion technique estimates the petrophysical properties within 0.5 CPU milliseconds with an accuracy between 91% and 99%, while nonlinear gradient-based inversion can estimate the petrophysical properties with an accuracy between 98% and 99% but requires several CPU minutes depending on the size of the dataset. The proposed PINN technique is therefore capable of providing fast and accurate estimation and uncertainty quantification of key petrophysical properties from triaxial resistivity logs, with comparable accuracy and approximately 106 speedup compared to nonlinear gradient-based inversion, and can be used for real-time applications such as automatic shale properties estimation, logging-while-drilling measurement interpretation, automated well geosteering, and time-lapse reservoir monitoring.

Pseudo-MRI Engine for MRI-Free Electromagnetic Source Imaging.

Structural head MRIs are a crucial ingredient in MEG/EEG source imaging; they are used to define a realistically shaped volume conductor model, constrain the source space, and visualize the source estimates. However, individual MRIs are not always available, or they may be of insufficient quality for segmentation, leading to the use of a generic template MRI, matched MRI, or the application of a spherical conductor model. Such approaches deviate the model geometry from the true head structure and limit the accuracy of the forward solution. Here, we implemented an easy-to-use tool, pseudo-MRI engine, which utilizes the head-shape digitization acquired during a MEG/EEG measurement for warping an MRI template to fit the subject's head. To this end, the algorithm first removes outlier digitization points, densifies the point cloud by interpolation if needed, and finally warps the template MRI and its segmented surfaces to the individual head shape using the thin-plate-spline method. To validate the approach, we compared the geometry of segmented head surfaces, cortical surfaces, and canonical brain regions in the real and pseudo-MRIs of 25 subjects. We also tested the MEG source reconstruction accuracy with pseudo-MRIs against that obtained with the real MRIs from individual subjects with simulated and real MEG data. We found that the pseudo-MRI enables comparable source localization accuracy to the one obtained with the subject's real MRI. The study indicates that pseudo-MRI can replace the need for individual MRI scans in MEG/EEG source imaging for applications that do not require subcentimeter spatial accuracy.

Comparative analysis of three methods for estimating the compositions of construction waste.

Determination of the relative compositions of the mixed construction waste is crucial and an important step to enhance resource management. This information influences the design of construction waste recycling and sorting facilities, and aids in formulating effective management strategies for recycled and sorted waste products. However, different methods for waste sorting and composition recognition possess distinct characteristics and only apply to specific practical scenarios. In this study, three methods are compared: (i) manual sorting as a reference method, (ii) a manual image recognition method using infrared thermal imaging, and (iii) a deep learning-based image recognition method based on the SegFormer semantic segmentation model. The comparison focuses on accuracy, preferences, prerequisites, socio-environmental impacts, costs, and improvement potential. Results show that both manual and deep learning-based image recognition methods yield comparable accuracy to manual sorting for inert waste, with relative errors below 5.2%, but relatively higher recognition errors for non-inert waste. Overall, manual sorting remains the most cost-effective and fastest method, despite its high labor demand, spatial constraints, environmental impacts, and limited improvement potential. In comparison, manual image recognition requires approximately 9.2 times the processing time and 2.3 times the cost of manual sorting, while deep learning-based image recognition incurs about 9.9 times the time and 2.5 times the cost. Nevertheless, both image recognition methods offer potential environmental benefits and long-term efficiency gains.

Abstract WP283: Comparative Accuracy of Transcranial Doppler, Transthoracic Echocardiography, Transesophageal Echocardiography, and Cardiac CT in Detecting Right-to-Left Shunt in Embolic Stroke of Undetermined Source

Abstract WP283: Comparative Accuracy of Transcranial Doppler, Transthoracic Echocardiography, Transesophageal Echocardiography, and Cardiac CT in Detecting Right-to-Left Shunt in Embolic Stroke of Undetermined Source

Analysis of big data from New York taxi trip 2023: revenue prediction using ordinary least squares solution and limited-memory Broyden-Fletcher-Goldfarb-Shanno algorithms

This study explores the prediction of taxi trip fares using two linear regression methods: normal equations (ordinary least squares solution (OLS)) and limited-memory Broyden-Fletcher-Goldfarb-Shanno (L-BFGS). Utilizing a dataset of New York City yellow taxi trips from 2023, the analysis involves data cleaning, feature engineering, and model training. The data consists of over 12 million records, managed, and processed that involves configuring the Spark driver and executor memory to efficiently process the Parquet-format data stored on hadoop distributed file system (HDFS). Key features influencing fare amount, such as passenger count, trip distance, fare amount, and tip amount, were analyzed for correlation. Models were trained on an 80-20 train-test split, and their performance was evaluated using root-mean-square error (RMSE) and mean squared error (MSE). Results show that both methods provide comparable accuracy, with slight differences in coefficients and training time. Additionally, vendor performance metrics, including total trips, average trip distance, fare amount, and tip amount, were analyzed to reveal trends and inform strategic decisions for fleet management. This comprehensive analysis demonstrates the efficacy of linear regression techniques in predicting taxi fares and offers valuable insights for optimizing taxi operations.

Comparative Accuracy of dCAIS and Freehand Techniques for Immediate Implant Placement in the Maxillary Aesthetic Zone: An In Vitro Study

ObjectiveTo evaluate the accuracy of immediate implant placement in fresh extraction sockets in the maxillary aesthetic zone using a dynamic computer-assisted implant surgery system (dCAIS), with the evaluation of possible deviations versus freehand placement. MethodsA total of 18 implants were placed by an experienced surgeon in fresh extraction sockets of anterior teeth in 6 maxillary models. Nine implants were placed using the dCAIS system and 9 implants were placed using the conventional freehand technique. The following outcome parameters were measured and compared: positional deviation at entry, apex point and angular deviations between planned and placed implant position. Surgery time was measured for each procedure. Descriptive and statistical analyses were performed on all outcome parameters. ResultsGlobal entry deviations were not significantly different between the two techniques (p=0.078). dCAIS resulted in significantly more accurate implant placement in terms of global apex deviation with values of 1.28±0.36 mm and angular deviations with values of 1.29±0.64°, compared to 2.06±0.60 mm and 5.05±2.54° with freehand placement (p<0.001). The dental implant placement time was approximately three times longer when using dCAIS (10.99 ± 3.43 min) versus freehand (3.25± 0.63 min) (p < 0.001). ConclusionsdCAIS achieved more precise immediate implant placement in terms of apex deviation and angulation than freehand placement, but increased the surgery time. Clinical significancedCAIS provides greater accuracy in the placement of immediate implants in the maxillary aesthetic zone following prosthetic-driven digital planning compared to freehand surgery.

Towards transcervical ultrasound-guided transoral robotic surgery.

In the context of transoral robotic surgery (TORS) for oropharyngeal squamous carcinoma (OPSCC), preoperative imaging and intraoperative visualization plays a pivotal role in optimizing resection margins. Prior work has demonstrated the ability of transoral ultrasound (US) in identifying OPSCC margins and vascular structures. This study evaluates the effectiveness of transcervical ultrasound (TUS), as well as other preoperative imaging modalities, in evaluating OPSCC volumes and compares this to post TORS pathological OPSCC volumes. Forty-one patients undergoing TORS between 2021 and 2023 were included. TUS was performed in all 41 patients, of which 37 had preoperative CT, 16 had PET-CT and 15 had MRI. Tumor dimensions on TUS, CT, and MRI were measured in craniocaudal, anteroposterior, and mediolateral planes to compute tumor volumes. Preoperative PET-CTs were analyzed to compute the metabolic tumour volume (MTV). Pathological tumor volumes served as the gold standard for comparison. No statistically significant differences were found between pathological tumor volumes and those measured by TUS, CT, PET-CT, or MRI (p=0.57, 0.47, 0.28, 0.29). Both TUS and PET-CT showed strong correlation with pathology (R=0.92, p<0.0001), followed by CT (R=0.83, p<0.0001) and MRI (R=0.55, p=0.031). The percent difference of radiologic volumes from pathology volumes was lowest for MRI (19.37% ± 28.28), followed by TUS (26.12% ± 20.97), PET-CT (32.59% ± 21.95), and CT (39.94% ± 62.94). TUS demonstrates comparable accuracy to CT, PET-CT, and MRI in assessing primary tumor volumes in patients with oropharyngeal squamous cell carcinoma (OPSCC) undergoing TORS. The strong correlation of TUS with final pathology, combined with its relatively non-invasive transcervical (versus transoral) approach and real-time acquisition, suggests that TUS has the potential to supplement TORS with image guidance.

Noncontrast imaging for the surveillance of treated and untreated meningiomas.

Patients with meningiomas require serial MRI for surveillance of tumor size and growth rate. The cost and resource requirements for contrast-enhanced MRI include intravenous cannulation, the contrast agent, risk of adverse reaction, and the time needed to acquire, review, and report the additional sequences. With repeated doses, gadolinium is known to accumulate in neural tissues. The authors compared the correlation and accuracy of axial T2-weighted imaging (T2WI) sequences alone for assessing tumor growth, dimensions, and dural venous sinus invasion compared with the current clinical practice of assessing both contrast-enhanced T1-weighted imaging (CE-T1WI) and T2WI sequences. The authors retrospectively identified 136 adult patients (65 patients with treated and 71 patients with untreated meningiomas) with two MRI scans obtained at least 6 months apart. For each patient, the two CE-T1WI sequences separated by time were paired, as were the two T2WI sequences, and assessed independently. The paired scans were assessed by a neuroradiologist and advanced radiology trainee blinded to clinical data. Tumor location, dimensions, growth, and venous invasion were evaluated. Peritumoral edema was assessed on T2WI only. Agreement between assessments on both CE-T1WI and T2WI sequences compared with T2WI alone was evaluated using Cohen's kappa (κ), the intraclass correlation coefficient (ICC), and Bland-Altman plots. Growth was detected in 36 tumors on T2WI compared with 39 when both CE-T1WI and T2WI were assessed. Growth assessed on T2WI alone showed near-perfect agreement with growth assessed on CE-T1WI and T2WI together (κ = 0.945). T2WI alone had an accuracy of 97.8%, specificity of 100%, and sensitivity of 92.3%. Interrater correlation between the radiologists for tumor dimensions was good to excellent (ICC > 0.843). Intrarater agreement between T2WI and CE-T1WI measurements of anteroposterior and transverse tumor dimensions was good (ICC > 0.883 for observer 1, > 0.767 for observer 2). There was substantial agreement between venous invasion on T2WI and both CE-T1WI and T2WI (κ = 0.771). Subgroup analysis for skull base (58.1%), treated (47.8%), and large (> 20-mm diameter; 38.2%) meningiomas did not show any significant difference in agreement between T2WI only and CE-T1WI and T2WI assessments of growth, venous invasion, or tumor dimension. In patients with treated and untreated meningiomas, unenhanced T2WI can assess tumor dimensions, detect growth, and detect venous invasion with comparable reliability and accuracy to the current clinical practice of using both CE-T1WI and T2WI.

Estimating Skeletal Maturity by Segmented Linear Modeling of Key AP Wrist Radiographic Parameters.

The recently described Modified Fels wrist skeletal maturity system (mFels wrist SMS) allows for accurate skeletal maturity estimation using a single anteroposterior wrist radiograph but requires evaluation of 8 parameters. A faster method may have clinical utility in the outpatient setting. The 8 anteroposterior wrist radiographic parameters comprising the mFels wrist SMS were analyzed in 80 children. Segmented linear regression and generalized estimating equation (GEE) analyses were used to identify the subsets of parameters most important for accurate skeletal maturity estimation for different patient demographics and parameter scores. This process produced multiple abbreviated Fels wrist skeletal maturity system options (abFels wrist SMS), which require fewer parameters. The accuracy of the resulting abFels wrist SMS options was evaluated and compared with the full 8-parameter mFels wrist SMS, the Greulich and Pyle system (GP) optimized with age and sex data, and the Sanders Hand System (SHS) optimized with age and sex data. A total of 372 wrist radiographs from 42 girls (age range, 7 to 15y) and 38 boys (age range, 9 to 16y) were included. The abFels wrist SMS options required fewer parameters (range, 2 to 7 parameters, one option with 2 to 4 parameters) than the full mFels wrist SMS (8 parameters). Skeletal age estimates produced by the full mFels SMS were more accurate than GP (P<0.001), SHS (P<0.001), and similar to the all of the abFels wrist SMS options (P>0.05). The mFels wrist SMS made fewer outlier estimations (estimation >1y off from the patient's actual skeletal age) than GP (P<0.05), and made similar rates of outlier estimations compared with SHS (P>0.05) and the abFels wrist SMS options (P>0.05). An abbreviated version of the Modified Fels wrist SMS estimates skeletal maturity with comparable accuracy to the full version but only requires 2 to 4 parameters for quicker use.

Accuracy Comparison of Hard, Soft, and UnQuantized Decisions across Diverse MIMO Channels and M-QAM Demodulation

Accuracy Comparison of Hard, Soft, and UnQuantized Decisions across Diverse MIMO Channels and M-QAM Demodulation

Prognostic value of exercise right ventricular pulmonary arterial coupling in severe primary mitral regurgitation: new atrial fibrilation and unplanned cardiovascular hospitalization

Abstract Background Exercise-induced pulmonary hypertension, characterized by mean pulmonary artery pressure over cardiac output slope (mPAP/CO slope) &amp;gt; 3mmHg/L/min is associated with worse outcome in greater than moderate primary mitral regurgitation (PMR). However, the prognostic value of right ventricle to pulmonary artery coupling (RVPAc) is unknown. Purpose Assess the prognostic value of RVPAc; determine the additional value of exercise over rest RVPAc and compare these findings to the mPAP/CO slope. Methods The single center study included consecutive patients with greater than moderate PMR, no/discordant symptoms, left ventricular ejection fraction &amp;gt;60% and absence of concomitant valvular disease greater than moderate or permanent atrial fibrillation (AF) referred to simultaneous CPET and exercise echocardiography (CPET-echo). A thorough echocardiographic assessment of right ventricle (RV) systolic function and RVPAc (TAPSE/sPAP, ratio of tricuspid annular plane systolic excursion over systolic pulmonary artery pressure) was performed using a dedicated RV window. mPAP and CO were obtained by Doppler echocardiography. Primary outcome was the composite of cardiovascular mortality, unplanned cardiovascular hospitalization and new AF episodes. Results A total of 159 consecutive patients (64±11 years, 59% men) were included. The event-free survival rate was 84% at 1 year and 78% at 2 years. Patients who fulfilled the primary combined endpoint had significantly larger left atrium indexed volumes (LAVi), lower left atrial strain and strain rate at rest and strain at intermediate exercise, lower absolute and normalized peak oxygen uptake (VO2peak), and a significantly higher mPAP/CO slope. They had significantly lower TAPSE, RV free wall S’ and TAPSE/sPAP. Sequentially adding intermediate or high exercise TAPSE/sPAP and percent-predicted VO2peak to the baseline predictive model (age, LAVi, mitral regurgitation grade and TAPSE/sPAP at rest) significantly improved the area under the curve (AUC) of the baseline logistic regression model (AUC: 0.71 vs. 0.80 and 0.71 vs. 0.81, p&amp;lt;0.05, respectively), with LAVi and TAPSE/sPAP at intermediate or high exercise remaining as significant independent variables (although coupling assessment at high exercise technically less feasible). Replacing exercise TAPSE/sPAP with mPAP/CO yields models with comparable accuracy (Figure 1). Exercise TAPSE/sPAP &amp;lt;0,6 was related to a higher event rate (Figure 2) Conclusion Decreased rest or exercise TAPSE/sPAP are single point measures of RVPAc, associated with adverse outcome in patients with greater than moderate PMR and no or discordant symptoms. Exercise TAPSE/sPAP has independent additional value over rest TAPSE/sPAP in predicting adverse events, with a similar accuracy as mPAP/CO slope. Exercise TAPSE/sPAP represents a potential alternative to mPAP/CO slope in this population, being readily available and simpler to adopt in clinical practice.

P-2229. Comparative Evaluation of Multimodal Large Language Models with Vision in Infectious Diseases Cases with Clinical Images

Abstract Background Real-world patient scenarios often involve a composite of text and various clinical images. We explored the performance of advanced multimodal Large Language Models (LLMs) equipped with image-to-text models (vision) in diagnosing complex infectious disease cases containing both text and clinical images.Figure 1:Case #19003; Right foot cutaneous lesion [Photography]. Partners’ Infectious Disease Images; Accessed on April 4, 2024 from https://www.idimages.org/images/detail/?imageid=2356 A sample image from IDImages.org showing a physical exam finding and the descriptions from the AI chatbots. Methods We assessed the performance of two publicly accessible LLMs with vision—GPT-4 and Claude Opus—on a series of 25 complex infectious disease scenarios with 67 clinical images sourced from www.IDImages.org. Prompts were standardized for each case, including instructions to identify the diagnosis, generate differential diagnoses, and describe each image. The models' accuracy in identifying the diagnosis and including it within the differential diagnoses was evaluated using a binary scoring system (0 or 1). Image description quality was assessed using a Likert scale ranging from 0 to 2 comparing the models' outputs against provided descriptions (Figure 1). Scoring was conducted by two independent clinicians, with the final scores being computed as averages. Figure 2 Mean scores of GPT-4 and Opus (1 denoting maximum accuracy) in correctly identifying the diagnosis (left) and in including the main diagnosis in the list of differential diagnoses (right). Results Opus achieved a higher mean accuracy (0.74) compared to GPT-4 (0.56) in determining the correct diagnosis, with the observed difference being statistically significant (p = 0.043). Nevertheless, both models demonstrated comparably high accuracy in incorporating the primary diagnosis within the differential diagnoses, with scores of 0.90 for GPT-4 and 0.88 for Opus (Figure 2). Additionally, both models (GPT-4 average score 1.79 ± 0.43, Opus 1.75 ± 0.50) performed well in providing accurate descriptions of different clinical images. In terms of image type, they performed best in interpreting images with physical findings (GPT-4 1.92 ± 0.23, Opus 1.83 ± 0.46), followed by radiographic images (GPT-4 1.80 ± 0.41, Opus 1.83 ± 0.36), and worst in pathological specimens (GPT-4 1.61 ± 0.58, Opus 1.57 ±0.58) (Figure 3). No statistical differences were found between the two models in their overall ability to describe images or within each image category. Figure 3 Mean accuracy scores of GPT-4 and Opus (up to 2) in interpreting clinical images grouped according to image type. Conclusion The two models demonstrated comparable accuracy in handling complex infectious disease scenarios that integrated both texts and images, suggesting potential utility in enhancing clinical support. Disclosures All Authors: No reported disclosures

Advancing Iron Ore Grade Estimation: A Comparative Study of Machine Learning and Ordinary Kriging

Mineral grade estimation is a vital phase in mine planning and design, as well as in the mining project’s economic assessment. In mining, commonly accepted methods of ore grade estimation include geometrical approaches and geostatistical techniques such as kriging, which effectively capture the spatial grade variation within a deposit. The application of machine-learning (ML) techniques has been explored in the estimation of mineral resources, where complex correlations need to be captured. In this paper, the authors developed four machine-learning regression models, i.e., support vector regression (SVR), random forest regression (RFR), k-nearest neighbour (KNN) regression, and extreme gradient boost (XGBoost) regression, using a geological database to predict the grade in an Indian iron ore deposit. When compared with ordinary kriging (R2 = 0.74; RMSE = 2.09), the RFR (R2 = 0.74; RMSE = 2.06), XGBoost (R2 = 0.73; RMSE = 2.12), and KNN (R2 = 0.73; RMSE = 2.11) regression models produced similar results. The block model predictions generated using the RFR, XGBoost, and KNN models show comparable accuracy and spatial trends to those of ordinary kriging, whereas SVR was identified as less effective. When integrated with geological methods, these models demonstrate significant potential for enhancing and optimizing mine planning and design processes in similar iron ore deposits.

Graph Neural Network Learning on the Pediatric Structural Connectome

Purpose: Sex classification is a major benchmark of previous work in learning on the structural connectome, a naturally occurring brain graph that has proven useful for studying cognitive function and impairment. While graph neural networks (GNNs), specifically graph convolutional networks (GCNs), have gained popularity lately for their effectiveness in learning on graph data, achieving strong performance in adult sex classification tasks, their application to pediatric populations remains unexplored. We seek to characterize the capacity for GNN models to learn connectomic patterns on pediatric data through an exploration of training techniques and architectural design choices. Methods: Two datasets comprising an adult BRIGHT dataset (N = 147 Hodgkin’s lymphoma survivors and N = 162 age similar controls) and a pediatric Human Connectome Project in Development (HCP-D) dataset (N = 135 healthy subjects) were utilized. Two GNN models (GCN simple and GCN residual), a deep neural network (multi-layer perceptron), and two standard machine learning models (random forest and support vector machine) were trained. Architecture exploration experiments were conducted to evaluate the impact of network depth, pooling techniques, and skip connections on the ability of GNN models to capture connectomic patterns. Models were assessed across a range of metrics including accuracy, AUC score, and adversarial robustness. Results: GNNs outperformed other models across both populations. Notably, adult GNN models achieved 85.1% accuracy in sex classification on unseen adult participants, consistent with prior studies. The extension of the adult models to the pediatric dataset and training on the smaller pediatric dataset were sub-optimal in their performance. Using adult data to augment pediatric models, the best GNN achieved comparable accuracy across unseen pediatric (83.0%) and adult (81.3%) participants. Adversarial sensitivity experiments showed that the simple GCN remained the most robust to perturbations, followed by the multi-layer perceptron and the residual GCN. Conclusions: These findings underscore the potential of GNNs in advancing our understanding of sex-specific neurological development and disorders and highlight the importance of data augmentation in overcoming challenges associated with small pediatric datasets. Further, they highlight relevant tradeoffs in the design landscape of connectomic GNNs. For example, while the simpler GNN model tested exhibits marginally worse accuracy and AUC scores in comparison to the more complex residual GNN, it demonstrates a higher degree of adversarial robustness.