Real-world Variability Research Articles

Enhancing supply chain resilience through supervised machine learning: supplier performance analysis and risk profiling for a multi-class classification problem

PurposeThis paper explores the application of supervised machine learning (ML) classification models to address supplier performance analysis and risk profiling as a multi-class classification problem. The research highlights that current applications of machine learning in supplier selection primarily focus on binary classification problems, underscoring a significant gap in the literature.Design/methodology/approachThis research paper opts for a structured approach to solve supplier selection and risk profiling using supervised machine learning multi-class classification models and prediction probabilities. The study involved a synthetic data set of 1,600 historical data points, creating a supplier selection framework that simulates current supply chain (SC) performance. The “Supplier Analysis and Selection ML Module” guided supplier selection recommendations based on ML analysis. Real-world variability is introduced through random seeds, impacting actual delivery dates, quantity delivered and quality performance. Supervised ML models, with hyperparameter tuning, enable multi-class classification of suppliers, considering past delivery performance and risk calculations.FindingsThe study demonstrates the effectiveness of the supervised ML-based approach in ensuring consistent supplier selection across multi-class classification problems. Beyond evaluating past delivery performance, it introduces a new dimension by predicting and assessing supplier risks through ML-generated prediction probabilities. This can enhance overall SC visibility and help organizations optimize strategies associated with risk mitigation, inventory management and customer service.Research limitations/implicationsThe findings highlight the adaptability of ML-based methodologies in dynamic SC environments, providing a proactive means to identify and manage risks. These insights are vital for organizations aiming to bolster SC resilience, particularly amid uncertainties.Practical implicationsThe practical implications of this study are significant for both commercial and humanitarian supply chain management (SCM). For commercial applications, the ML-based methodology allows businesses to make more informed supplier selection decisions, reducing risks and improving operational efficiency. In disaster and humanitarian SC contexts, the use of ML can improve preparedness and resource allocation, ensuring that critical supplies reach affected areas promptly.Social implicationsThe study’s implications extend to disaster and humanitarian SCM, where timely and efficient delivery is critical for saving lives and alleviating suffering. ML tools can improve preparedness, resource allocation and coordination in these contexts, enhancing the resilience and responsiveness of humanitarian supply chains.Originality/valueUnlike conventional methods focused on quality, cost and delivery performance aspects, the current study introduces supervised ML to identify and assess supplier risks through prediction probabilities for multi-class classification problems (delivery performance as late, on-time and ahead), offering a refined understanding of supplier selection in dynamic SC environments.

The Data Heterogeneity Issue Regarding COVID-19 Lung Imaging in Federated Learning: An Experimental Study

Federated learning (FL) has emerged as a transformative framework for collaborative learning, offering robust model training across institutions while ensuring data privacy. In the context of making a COVID-19 diagnosis using lung imaging, FL enables institutions to collaboratively train a global model without sharing sensitive patient data. A central manager aggregates local model updates to compute global updates, ensuring secure and effective integration. The global model’s generalization capability is evaluated using centralized testing data before dissemination to participating nodes, where local assessments facilitate personalized adaptations tailored to diverse datasets. Addressing data heterogeneity, a critical challenge in medical imaging, is essential for improving both global performance and local personalization in FL systems. This study emphasizes the importance of recognizing real-world data variability before proposing solutions to tackle non-independent and non-identically distributed (non-IID) data. We investigate the impact of data heterogeneity on FL performance in COVID-19 lung imaging across seven distinct heterogeneity settings. By comprehensively evaluating models using generalization and personalization metrics, we highlight challenges and opportunities for optimizing FL frameworks. The findings provide valuable insights that can guide future research toward achieving a balance between global generalization and local adaptation, ultimately enhancing diagnostic accuracy and patient outcomes in COVID-19 lung imaging.

In Shift and In Variance: Assessing the Robustness of HAR Deep Learning Models Against Variability

Deep learning (DL)-based Human Activity Recognition (HAR) using wearable inertial measurement unit (IMU) sensors can revolutionize continuous health monitoring and early disease prediction. However, most DL HAR models are untested in their robustness to real-world variability, as they are trained on limited lab-controlled data. In this study, we isolated and analyzed the effects of the subject, device, position, and orientation variabilities on DL HAR models using the HARVAR and REALDISP datasets. The Maximum Mean Discrepancy (MMD) was used to quantify shifts in the data distribution caused by these variabilities, and the relationship between the distribution shifts and model performance was drawn. Our HARVAR results show that different types of variability significantly degraded the DL model performance, with an inverse relationship between the data distribution shifts and performance. The compounding effect of multiple variabilities studied using REALDISP further underscores the challenges of generalizing DL HAR models to real-world conditions. Analyzing these impacts highlights the need for more robust models that generalize effectively to real-world settings. The MMD proved valuable for explaining the performance drops, emphasizing its utility in evaluating distribution shifts in HAR data.

Evaluating Artificial Intelligence Models for Resource Allocation in Circular Economy Digital Marketplace

This study assesses the application of artificial intelligence (AI) algorithms for optimizing resource allocation, demand-supply matching, and dynamic pricing within circular economy (CE) digital marketplaces. Five AI models—autoregressive integrated moving average (ARIMA), long short-term memory (LSTM), random forest (RF), gradient boosting regressor (GBR), and neural networks (NNs)—were evaluated based on their effectiveness in predicting waste generation, economic growth, and energy prices. The GBR model outperformed the others, achieving a mean absolute error (MAE) of 23.39 and an R2 of 0.7586 in demand forecasting, demonstrating strong potential for resource flow management. In contrast, the NNs encountered limitations in supply prediction, with an MAE of 121.86 and an R2 of 0.0151, indicating challenges in adapting to market volatility. Reinforcement learning methods, specifically Q-learning and deep Q-learning (DQL), were applied for price stabilization, resulting in reduced price fluctuations and improved market stability. These findings contribute a conceptual framework for AI-driven CE marketplaces, showcasing the role of AI in enhancing resource efficiency and supporting sustainable urban development. While synthetic data enabled controlled experimentation, this study acknowledges its limitations in capturing full real-world variability, marking a direction for future research to validate findings with real-world data. Moreover, ethical considerations, such as algorithmic fairness and transparency, are critical for responsible AI integration in circular economy contexts.

Spiking Neural Networks for Real-Time Pedestrian Street-Crossing Detection Using Dynamic Vision Sensors in Simulated Adverse Weather Conditions

This study explores the integration of Spiking Neural Networks (SNNs) with Dynamic Vision Sensors (DVSs) to enhance pedestrian street-crossing detection in adverse weather conditions—a critical challenge for autonomous vehicle systems. Utilizing the high temporal resolution and low latency of DVSs, which excel in dynamic, low-light, and high-contrast environments, this research evaluates the effectiveness of SNNs compared to traditional Convolutional Neural Networks (CNNs). The experimental setup involved a custom dataset from the CARLA simulator, designed to mimic real-world variability, including rain, fog, and varying lighting conditions. Additionally, the JAAD dataset was adopted to allow for evaluations using real-world data. The SNN models were optimized using Temporally Effective Batch Normalization (TEBN) and benchmarked against well-established deep learning models, concerning their accuracy, computational efficiency, and energy efficiency in complex weather conditions. This study also conducted a comprehensive analysis of energy consumption, highlighting the significant reduction in energy usage achieved by SNNs when processing DVS data. The results indicate that SNNs, when integrated with DVSs, not only reduce computational overhead but also dramatically lower energy consumption, making them a highly efficient choice for real-time applications in autonomous vehicles (AVs).

Bounded subadditivity in management decisions

PurposeAs subjects irrationally perceive probability changes as more impactful when shifting an event from impossible to possible or from possible to certain, compared to increasing the likelihood of an already possible event, this study examines how workers process success probabilities and whether their resource allocation decisions are distorted by bounded subadditivity.Design/methodology/approachWe conduct an online randomized experiment with 3,980 employees.FindingsWe detect a certainty effect (upper subadditivity), whereby professionals are willing to devote a disproportionate number of hours to a project when their contribution transforms the success of the initiative from possible to certain rather than increasing the likelihood of success by the same percentage points. We find no evidence of the possibility effect (lower subadditivity), whereby workers would devote a disproportionate effort when their contribution turns a sure failure into a possible success rather than simply increasing the likelihood of success by the same percentage points. We observe a rational tendency to try harder for a greater increase in the probability of success, but only far from the limits of the probability spectrum and not close to the limits.Originality/valueAttempts to understand bounded subadditivity in management decisions have been incomplete. We disentangle two real-world variables and offer a more refined operationalization.

Improving vehicle warm-up performance in cold conditions using phase change materials

In cold climates, improving vehicle warm-up performance is crucial for reducing emissions and fuel consumption. Traditional methods often fail to efficiently address this issue, particularly during initial cold start periods. This study aims to enhance vehicle warm-up performance in cold conditions using phase change materials (PCMs). A thermal energy storage device was developed using paraffin wax, selected for its high energy density and a melting point of 69.3 °C, integrated with an optimized heat exchanger. Real road driving tests were conducted under urban and highway conditions, with temperatures ranging from −5 to 0 °C, using a 2.2L diesel engine vehicle. The results demonstrated a significant reduction in engine warm-up time by 20–30 %, leading to a decrease in fuel consumption and CO emissions by 365–517 g annually. The thermal energy storage device supplied up to 694.63 kJ of heat energy to the coolant, further improving vehicle efficiency and reducing environmental impact. Furthermore, evaluating PCM-based systems under real-world driving conditions is crucial for validating their practical effectiveness. Real-world variables introduce challenges that help confirm the applicability of PCM-based systems in everyday use. This approach shows potential for enhancing warm-up performance in cold climates without additional fuel consumption, outperforming conventional methods.

A Model of Alcohol Consumption with Semi-Markov Variable Coefficients

This paper focuses on the in-depth study of a stochastic approximation methodology involving semi-Markov switches in an averaging scheme with a minor parameter. We present a model where perceived objects are impacted by noise variables dependent on the semi-Markov process. The emphasis is on analyzing the convergence and stability of the stochastic approximation process within this context. Theoretical results are presented alongside numeric simulations, demonstrating the practical applications in managing complex stochastic systems. This work opens promising paths for the use of stochastic approximation approaches in the wider field of semi-Markov processes. Recognizing the mutable nature of alcohol consumption and its dependency on a variety of factors, we propose encoding such dynamism within a semi-Markov parameter structure. This model handles the patterns of individual alcohol consumption as a semi-Markov process; the transition probabilities between different states of alcohol consumption are directed by sociodemographic variables that change over time. Our approach, thus, bridges the gap between the realities of ever-changing alcohol consumption trends and static traditional Markov-chain models. By integrating real-world variables into our innovative model, we offer a cutting-edge analytical tool that lays down new paths for understanding and addressing the challenges with alcohol consumption patterns. Further, our insights have the potential to significantly impact the formulation of effective strategies and public health interventions aimed at alcohol-related harm reduction

TOMATO PLANT DISEASE DETECTION AND PREDICTION USING MACHINE LEARNING AND DEEP LEARNING TECHNIQUES: A COMPREHENSIVE REVIEW

Tomatoes (Solanum lycopersicum) are critical to global agriculture, but their cultivation is threatened by diseases caused by fungi, bacteria, and viruses. Traditional disease detection methods are often manual and inefficient, resulting in delayed diagnosis and increased crop loss. Recent advances in machine learning (ML) and deep learning (DL) technologies have revolutionized disease detection by automating this process. These technologies utilize extensive image datasets and environmental data to train algorithms that are capable of identifying disease symptoms with high accuracy. This review explores various ML and DL techniques for tomato disease detection, including support vector machines, artificial neural networks, and convolution neural networks. This review also examines the integration of these technologies into practical tools, such as mobile applications for realtime diagnostics. The findings indicate that deep learning models offer superior accuracy compared to traditional methods, and that incorporating environmental data enhances prediction reliability. Challenges, such as data scarcity, real-world variability, and the need for userfriendly applications, remain. Future research should focus on integrating Internet of Things (IoT) technologies for real-time monitoring, fostering collaborative datasharing, and developing intuitive tools for farmers. By harnessing ML and DL, the agricultural sector could advance disease management, improve crop productivity, and promote sustainability.

A survey on various edge detection techniques in image processing and applied disease detection

This paper surveys various edge detection techniques in image processing, focusing on their applicability to disease detection. Many researchers encompass studies conducted in the context of various crops and fruits, shedding light on their effectiveness and adaptability. However, the more techniques are used and improved, less comparison has been made between them to look further at their challenges, such as noise sensitivity, scale variability, edge linking, and real-world variability. Also, the study will systematically survey and analyze literature on the ability of edge detection, including classical methods like Robert, Sobel, Prewitt, and Canny, as well as more advanced techniques such as gradient-based and Gaussian-based. This research aims to comprehensively understand the strengths and limitations of different edge detection techniques and can be used as a reference point for selecting and enhancing novel techniques in image processing. Overview, this paper makes a substantial contribution to the field by addressing both traditional edge detection in image processing and applied disease detection. It serves as a comprehensive guide, offering insights, practical advice, and a consolidated view of current research trends, and highlights the potential of edge detection in contributing to advancements in disease detection methodologies making it a valuable resource for researchers and practitioners.

Real-world biomarker variability and effects of biologics on severe asthma in Alberta

Rationale: Guidelines recommend biomarker testing to phenotype patients with severe asthma (SA), to guide treatment. However, biomarker levels fluctuate, and individual responses to biologics are not fully understood. Objectives: This study estimated the proportion of patients experiencing biomarker variability and the real-world effectiveness of biologics in SA. Methods: A population-based retrospective cohort study was conducted using administrative data from Alberta, Canada (April 1, 2010 to March 31, 2020) for patients with SA. Year-to-year variability in biomarker levels was assessed using clinical thresholds (blood eosinophil count [EOS] ≥ 300 cells/µL; immunoglobin E [IgE] ≥ 30 IU/mL) to explore category switching. Incidence rate ratios of exacerbations were estimated by biomarker level. Associations between follow-up time with biologics exposure (bio-experience), biomarker levels and exacerbation rates were modeled. Results: Up to 28% of SA patients displayed year-over-year biomarker threshold switching for EOS and up to 12% for IgE. Severe exacerbation rates were higher with blood EOS count ≥300cells/µL or IgE ≥30 IU/mL. Bio-experience was associated with lower blood EOS versus pre-bio-experience (relative mean EOS: 0.53 [0.42-0.68]), persisting >1 year following discontinuation. Bio-experience was associated with a nearly 50% reduction in exacerbation risk after initiating biologics (IRR = 0.54 [0.48, 0.62]), regardless of comorbid status. Conclusions: Higher biomarker levels were associated with higher exacerbation rates, but a proportion of patients demonstrated variability, crossing clinical thresholds. Treatments with upstream agents targeting multiple pathways of inflammation may circumvent this issue. Biologics had real-world effectiveness in reducing SA biomarkers. Their persistent temporal effect provides a groundwork for exploring cost-effective biologics use.

Real-world data on ALK rearrangement test in Chinese advanced non-small cell lung cancer (RATICAL): a nationwide multicenter retrospective study.

Anaplastic lymphoma kinase (ALK) test in advanced non-small cell lung cancer (NSCLC) can help physicians provide target therapies for patients harboring ALK gene rearrangement. This study aimed to investigate the real-world test patterns and positive rates of ALK gene rearrangements in advanced NSCLC. In this real-world study (ChiCTR2000030266), patients with advanced NSCLC who underwent an ALK rearrangement test in 30 medical centers in China between October 1, 2018 and December 31, 2019 were retrospectively analyzed. Interpretation training was conducted before the study was initiated. Quality controls were performed at participating centers using immunohistochemistry (IHC)-VENTANA-D5F3. The positive ALK gene rearrangement rate and consistency rate were calculated. The associated clinicopathological characteristics of ALK gene rearrangement were investigated as well. The overall ALK gene rearrangement rate was 6.7% in 23,689 patients with advanced NSCLC and 8.2% in 17,436 patients with advanced lung adenocarcinoma. The quality control analysis of IHC-VENTANA-D5F3 revealed an intra-hospital consistency rate of 98.2% (879/895) and an inter-hospital consistency rate of 99.2% (646/651). IHC-VENTANA-D5F3 was used in 53.6%, real-time polymerase chain reaction (RT-PCR) in 25.4%, next-generation sequencing (NGS) in 18.3%, and fluorescence in-situ hybridization (FISH) in 15.9% in the adenocarcinoma subgroup. For specimens tested with multiple methods, the consistency rates confirmed by IHC-VENTANA-D5F3 were 98.0% (822/839) for FISH, 98.7% (1,222/1,238) for NGS, and 91.3% (146/160) for RT-PCR. The overall ALK gene rearrangement rates were higher in females, patients of ≤ 35 years old, never smokers, tumor cellularity of > 50, and metastatic specimens used for testing in the total NSCLC population and adenocarcinoma subgroup (all P < 0.05). This study highlights the real-world variability and challenges of ALK test in advanced NSCLC, demonstrating a predominant use of IHC-VENTANA-D5F3 with high consistency and distinct clinicopathological features in ALK-positive patients. These findings underscore the need for a consensus on optimal test practices and support the development of refined ALK test strategies to enhance diagnostic accuracy and therapeutic decision-making in NSCLC.

Distinct populations suppress or escalate intake of cocaine paired with aversive quinine.

Only a subset of individuals who encounter drugs of abuse become habitual users. Aversive subjective effects like coughing and unpleasant taste are predictors for continued use. While several preclinical studies have explored self-administration involving aversive cues, none have simultaneously introduced aversion with the initial drug self-administration. We aimed to develop a clinically relevant model by pairing intravenous cocaine with intraoral quinine self-administration from the outset and investigating whether repeated exposure to an aversive stimulus would alter its hedonic value under laboratory conditions. Twenty-seven male and female Sprague Dawley rats self-administered intravenous/intraoral (cocaine/quinine) for 2 hr/day over 14 days. This was followed by a 1-day quinine-only extinction session, a 3-day return to self-administration, and an intraoral infusion session to assess quinine taste reactivity (TR). We identified three distinct groups. The first self-administered very little cocaine, while the second sharply escalated cocaine intake. Both groups had similar aversive TR to quinine, suggesting that the escalating group did not habituate to the aversive cue but pursued drug despite it. We also identified a third group with high initial intake that decreased over time. This decrease predicted high aversive TR, and we argue this group may represent individuals who "overindulge" on their first use and subsequently find self-administration to be aversive. Our novel model mimics real-world variability in initial interactions with drugs of abuse and yields three distinct groups that differ in self-administration patterns and aversive cue valuation.

A data-driven multi-objective optimization approach for enhanced methanol yield and exergy loss minimization in direct hydrogenation of CO[formula omitted]

The stability and control of process industries have historically faced obstacles due to inherent uncertainties in their operations. This work focuses on leveraging machine learning models as a surrogate to facilitate real-time optimization of operating conditions of the direct hydrogenation of CO2 to enhance methanol production while minimizing exergy losses. Numerous studies have explored steady-state exergy analysis and the impact of various process conditions on methanol production. However, there is a notable absence of research investigating the influence of exergy destruction on methanol production under dynamic conditions. Initially, a commercial software Aspen HYSYS was utilized to design, model, and simulate the methanol synthesis plant. Exergy analysis was conducted to quantify the process exergy losses, exergy efficiency, and potential areas for improvement. The process model transitioned to a dynamic mode by incorporating ±5% uncertainty into critical operating conditions, i.e., temperature, pressure, and molar flow rate, simulating real-world variability and resulting in a dataset of 370 samples. Two machine learning models; the Gaussian process regression (GPR) and artificial neural network (ANN) model were developed using the data samples to predict the process exergy losses and molar flow rate of methanol produced. To enhance the predictive capabilities of these deployed models, Bayesian optimization was employed for hyperparameter tuning. The developed models were employed as surrogates in multi-objective genetic algorithm (MOGA) environments with the aim of maximizing the production of methanol with minimum exergy losses under uncertainty. The optimized process conditions, derived from MOGA-based methods, underwent cross-validation using the Aspen model. This analysis revealed that the methanol synthesis plant exhibited an exergy loss of 2425 kW, an exergy efficiency of 97.78%, and an improvement potential of 53.74 kW. The GPR and ANN models demonstrated high correlation coefficients (R2) of 0.9928 and 0.9375, and root mean square error (RMSE) values of 38.93 and 1.122, respectively. Incorporating these machine learning models as surrogates within the optimization frameworks notably outperformed the base case, achieving maximum methanol production while minimizing exergy losses in the process.

Ultrahigh frequency path loss prediction based on K-nearest neighbors

Abstract Path loss prediction (PLP) is an important feature of wireless communications because it allows a receiver to anticipate the signal strength that will be received from a transmitter at a given distance. The PLP is done by using machine learning models that take into account numerous aspects such as the frequency of the signal, the surroundings, and the type of antenna. Various machine learning methods are used to anticipate path loss propagation but it is difficult to predict path loss in unknown propagation conditions. In existing models rely on incomplete or outdated data, which can affect the accuracy and reliability of predictions and they do not take into account the effects of environmental factors, such as terrain, foliage, and weather conditions, on path loss. Furthermore, existing models are not robust enough to handle the real-world variability and uncertainty, leading to significant errors in predictions. To tackle this issue, a novel ultrahigh frequency (UHF) PLP based on K-nearest neighbors (KNNs) is developed for predicting and optimizing the path loss for UHF. In this proposed model, a KNN-based PLP has been used to predict the path loss in the UHF. This technique is used for high-accuracy PLP through KNN forecast route loss by determining the K-nearest data points to a particular test point based on a distance metric. Moreover, the existing models were not able to optimize path loss due to complex and large-scale machine learning models. Therefore, the stochastic gradient descent technique has been used to minimize the objective function, which is often a measure of the difference between the model’s predictions and the actual output that will fine-tune the parameters of the KNN model, by measuring the similarity between data points. This model is implemented using Python to make it a lot more convenient.

Abstract PO3-03-12: Factors associated with response to neoadjuvant chemoimmunotherapy in triple-negative breast cancer

Abstract Background: The addition of pembrolizumab (pembro) to neoadjuvant chemotherapy (NAC) became a standard-of-care for the treatment of early-stage triple negative breast cancer (TNBC) after the phase III KEYNOTE-522 trial demonstrated improved pathologic complete response (pCR) rates and event free survival with the combination. To date, clinical predictors of response to pembro with NAC are lacking. We sought to identify real-world clinical characteristics and treatment variables associated with response to NAC with pembro. Methods: We performed a single institution, retrospective study of early-stage TNBC patients diagnosed between February 1st, 2020 and December 1st, 2022 who were treated with NAC and pembro. Patients who had not undergone definitive surgery were excluded. Demographic information, clinical and pathologic characteristics, and treatment data were collected. pCR was defined as ypT0/Tis and ypN0. Univariate and multivariate analysis was performed using logistic regression to identify factors associated with pCR. This study was approved by an Institutional Review Board. Results: Of the 94 patients analyzed, 93 were female and 1 was male. The median age was 55 years (IQR 47 – 61.8) and median body mass index (BMI) was 30 (24.1-33.6). Self-reported racial/ethnic groups included 37 non-Hispanic White (39.4%), 33 Hispanic (35.1%), 13 Asian (13.8%), 7 Black (7.4%), and 4 other/unknown (4.3%). 8 (8.5%) and 2 (2.1%) of patients had deleterious germline BRCA1+ and BRCA2+ mutations, respectively. The majority of patients had invasive ductal histology (90.4%), clinical T2 stage (68.1%), and node negative (55.3%) disease. 61 (64.9%) patients completed the planned 8 cycles of NAC, while 42 (44.7%) completed the planned 8 cycles of neoadjuvant pembro. A pCR (ypT0/Tis ypN0) was achieved in 60 (63.8%) of the 94 patients. Among those who achieved a pCR, 32 (53.3%) completed the prescribed course of NAC and pembro in combination, whereas 9 (47.1%) of the 34 patients with residual disease did not. In univariate analyses, patients under 55 years at time of diagnosis (vs. age &amp;gt;55 years), higher ki-67, those completing the prescribed course of NAC, those completing the prescribed course of pembro, the months from start of NAC to surgery, and the months from start of pembro to surgery were associated with pCR. The duration from start of NAC or pembro to surgery overlapped considerably with the response explained by completion of NAC or pembro, respectively. Although pCR estimates were lower for Hispanic (60.6%) and non-Hispanic White patients (Black: 57.1%, Asian: 53.8%, and Other: 50%) when compared to non-Hispanic White patients (73.0%), the results were not statistically significant. In multivariate analyses, patients under 55 years at time of diagnosis (OR 3.06, 95% CI:1.18-8.0, p=0.02) and those completing the full course of pembro (OR 2.86, 95%CI: 1.08-7.58, p=0.034) had improved rates of pCR. BMI, race, BRCA1/2 status, histology, clinical T and N stage, grade, time between start/end of NAC and IO to surgery were not significantly associated with pCR on multivariate analysis. Conclusion: In our experience, younger TNBC patients were more likely to achieve a pCR with NAC with pembro. Moreover, the completion of the planned course of pembro prior to surgery was also associated with improved pCR rates. These findings suggest that adherence to neoadjuvant pembro prior to surgery may be more important than the completion of the NAC. Further research to identify the optimal exposure of NAC with pembro is needed to refine treatment recommendations in this high-risk population. Citation Format: Alexis LeVee, Megan Wong, Sarah Flores, Nora Ruel, Heather McArthur, James Waisman, Joanne Mortimer. Factors associated with response to neoadjuvant chemoimmunotherapy in triple-negative breast cancer [abstract]. In: Proceedings of the 2023 San Antonio Breast Cancer Symposium; 2023 Dec 5-9; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2024;84(9 Suppl):Abstract nr PO3-03-12.

Impact of neoadjuvant pembrolizumab adherence on pathologic complete response in triple-negative breast cancer: a real-world analysis.

The addition of pembrolizumab (pembro) to neoadjuvant chemotherapy (NAC) is standard of care for the treatment of early triple-negative breast cancer (TNBC) after KEYNOTE-522 trial demonstrated improved pathologic complete response (pCR) rates with the combination. However, the optimal treatment strategy for TNBC remains uncertain as questions persist about which patients benefit from pembro and the best treatment schedule and regimen. We identified real-world clinical characteristics and treatment variables associated with response to NAC plus pembro. Patients with early TNBC treated with NAC plus pembro between February 2020 and September 2023 were identified. Univariate and multivariate analysis was performed using logistic regression to identify factors associated with pCR. Cox proportional hazard prediction models were used to identify predictors of invasive disease-free survival and overall survival in this cohort. A pCR was achieved in 75 (63.6%) of 118 patients. Age at diagnosis (P = .04), Ki-67 (P = .004), duration from start of pembro to surgery (P = .006) and NAC to surgery (P = .01), number of cycles of pembro (P = .04) and NAC (P = .02), and completion of at least 8 cycles of pembro (P = .015) and NAC (P = .015) were each significantly associated with pCR in univariate analysis. In multivariate analysis, patients younger than 55 years at time of diagnosis (vs age > 55 years) and those completing at least 8 cycles of pembro remained predictive of pCR (OR's 2.50, 2.49, P = .035 and .037, respectively). In this real-world analysis of patients with TNBC treated with NAC plus pembro, younger age and the completion of at least 8 cycles of pembrolizumab were associated with pCR.

Impact of Nature of Medical Data on Machine and Deep Learning for Imbalanced Datasets: Clinical Validity of SMOTE Is Questionable

Dataset imbalances pose a significant challenge to predictive modeling in both medical and financial domains, where conventional strategies, including resampling and algorithmic modifications, often fail to adequately address minority class underrepresentation. This study theoretically and practically investigates how the inherent nature of medical data affects the classification of minority classes. It employs ten machine and deep learning classifiers, ranging from ensemble learners to cost-sensitive algorithms, across comparably sized medical and financial datasets. Despite these efforts, none of the classifiers achieved effective classification of the minority class in the medical dataset, with sensitivity below 5.0% and area under the curve (AUC) below 57.0%. In contrast, the similar classifiers applied to the financial dataset demonstrated strong discriminative power, with overall accuracy exceeding 95.0%, sensitivity over 73.0%, and AUC above 96.0%. This disparity underscores the unpredictable variability inherent in the nature of medical data, as exemplified by the dispersed and homogeneous distribution of the minority class among other classes in principal component analysis (PCA) graphs. The application of the synthetic minority oversampling technique (SMOTE) introduced 62 synthetic patients based on merely 20 original cases, casting doubt on its clinical validity and the representation of real-world patient variability. Furthermore, post-SMOTE feature importance analysis, utilizing SHapley Additive exPlanations (SHAP) and tree-based methods, contradicted established cerebral stroke parameters, further questioning the clinical coherence of synthetic dataset augmentation. These findings call into question the clinical validity of the SMOTE technique and underscore the urgent need for advanced modeling techniques and algorithmic innovations for predicting minority-class outcomes in medical datasets without depending on resampling strategies. This approach underscores the importance of developing methods that are not only theoretically robust but also clinically relevant and applicable to real-world clinical scenarios. Consequently, this study underscores the importance of future research efforts to bridge the gap between theoretical advancements and the practical, clinical applications of models like SMOTE in healthcare.

Advancing Crack Detection Using Deep Learning Solutions for Automated Inspection of Metallic Surfaces

The deep Convolutional Neural Network (CNN) architecture used in this research study provides a proof of concept for crack detection on the metallic surface of a hex nut. The goal is to create an automated receiving inspection process to supplement human inspections conducted on-site. Conventional image processing techniques (IPTs) have been extensively used for mechanical infrastructure fault detection. These techniques focus on image modification to extract typical features, such as surface fractures in materials like steel and concrete. However, obstacles presented by a variety of real-world variables, such as changes in lighting and shadows, make it difficult to use IPTs. Our suggested vision-based method employs a deep learning CNN to overcome these difficulties, eliminating the need to explicitly compare fault features. CNNs are more resilient to shifting real-world situations than IPTs since they are naturally trained to identify characteristics in images. Following training on a dataset of 1081 images with dimensions of 256 x 256 pixels, the VGG16 CNN architecture achieved an impressive accuracy of around 94.17%. Additional CNN architectures, including ResNet, MobileNet, AlexNet, and LeNet-5, are employed to assess and compare fault detection accuracies in order to select the most appropriate architecture for the model. To evaluate the robustness and flexibility of the suggested method in various situations, we conducted tests with 206 images from an alternative structure that was not part of the training dataset. These images depicted a range of circumstances, such as intense light patches and tiny fissures. The outcomes showed that our proposed method outperforms current approaches, highlighting its usefulness in practical situations involving the identification of metallic defects.

Abstract A007: Picturing progression: A new role of whole slide imaging in comprehensive clinical genomic datasets

Abstract Introduction: Real world data is needed to understand cancer evolution based on changing standards of care. While Next Generation Sequencing (NGS) is the preferred standard for confirmatory testing of mutational status, its use varies tremendously across settings and countries, with the most influential variable being cost. In this research, we examined the utility of Whole Slide Images (WSI) as a discrete data type to predict mutational and other statuses to inform assessments of clinical progression. This could include initial mutational status or the evolution of resistance to a specific treatment and the emergence of a new driver mutation. Experimental Procedures: We have curated datasets from 4000 NSCLC and 2000 CRC patients with clinically tested WES, WTS and WSI tumors with deeply curated longitudinal clinical electronic medical records were compiled through a partnership between ConcertAI and Caris Life Science. This de-identified clinical dataset is drawn from a diverse range of U.S. community and academic oncology practices. The practice patterns reflect the real-world variability of treatment and utilization of molecular testing. Additional stratification was performed based on lines of therapy and evidence of clinical progression for patients with information available. Results: Features from whole slide H&amp;E images were extracted and compared between subcohorts of interest. We demonstrated the correlation of the WSI features with TMB, MSI-H, molecular variants, differential gene expression and PD-L1 IHC of these cohorts within the context human interpretable features (HIFs) of whole slide images. Conclusion: The availability of whole slide images (WSI) is a dimensionality requirement for real world longitudinal comprehensive clinical datasets. Digital images should be considered as a key deliverable of clinical genomic datasets and have the potential to inform treatment selection and outcomes. In settings where NGS is not the standard of care, WSI could be used to predict the value of a test for treatment decisions or determination of prospective clinical trial eligibility. Where the cost of NGS is fundamentally prohibitive, such as less developed economies, WSI is a much lower cost modality that could have very high utility in guiding treatment decisions with better predicted outcomes. In all cases WSI imaging could be used to assess potential status or acquired resistance and the evolution of a new driver mutation. All critical in the decision to continue or discontinue a specific treatment. Citation Format: Elizabeth G. Sweeney, Yawei J. Wang, Jacob Picard, Jim Abraham, Eghbal Amidi, Michael R. Rossi, Sergey Klimov, Claudio D'Ambrosio, Ming Chen, Jeff Elton, George Sledge. Picturing progression: A new role of whole slide imaging in comprehensive clinical genomic datasets [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Translating Cancer Evolution and Data Science: The Next Frontier; 2023 Dec 3-6; Boston, Massachusetts. Philadelphia (PA): AACR; Cancer Res 2024;84(3 Suppl_2):Abstract nr A007.