Machine Learning Pipeline Research Articles

An Empirical Evaluation of Adapting Hybrid Parameters for CNN-based Sentiment Analysis

Sentiment analysis aims to understand human emotions and perceptions through various machine-learning pipelines. However, feature engineering and inherent semantic gap constraints often hinder conventional machine learning techniques and limit their accuracy. Newer neural network models have been proposed to automate the feature learning process and enrich learned features with word contextual embeddings to identify their semantic orientations to address these challenges. This article aims to analyze the influence of different factors on the accuracy of sentiment classification predictions by employing Feedforward and Convolutional Neural Networks. To assess the performance of these neural network models, we utilize four diverse real-world datasets, namely 50,000 movie reviews from IMDB, 10,662 sentences from LightSide Movie_Reviews, 300 public movie reviews, and 1,600,000 tweets extracted from Sentiment140. We experimentally investigate the impact of exploiting GloVe word embeddings on enriching feature vectors extracted from sentiment sentences. Findings indicate that using larger dimensions of GloVe word embeddings increases the sentiment classification accuracy. In particular, results demonstrate that the accuracy of the CNN with a larger feature map, a smaller filter window, and the ReLU activation function in the convolutional layer was 90.56% using the IMDB dataset. In comparison, it was 80.73% and 77.64% using the sentiment140 and the 300 sentiment sentences dataset, respectively. However, it is worth mentioning that, with large-size sentiment sentences (LightSide’s Movie Reviews) and using the same parameters, only a 64.44% level of accuracy was achieved.

Early prediction of end-stage kidney disease using electronic health record data: a machine learning approach with a 2-year horizon.

In the United States, end-stage kidney disease (ESKD) is responsible for high mortality and significant healthcare costs, with the number of cases sharply increasing in the past 2 decades. In this study, we aimed to reduce these impacts by developing an ESKD model for predicting its occurrence in a 2-year period. We developed a machine learning (ML) pipeline to test different models for the prediction of ESKD. The electronic health record was used to capture several kidney disease-related variables. Various imputation methods, feature selection, and sampling approaches were tested. We compared the performance of multiple ML models using area under the ROC curve (AUCROC), area under the Precision-Recall curve (PR-AUC), and Brier scores for discrimination, precision, and calibration, respectively. Explainability methods were applied to the final model. Our best model was a gradient-boosting machine with feature selection and imputation methods as additional components. The model exhibited an AUCROC of 0.97, a PR-AUC of 0.33, and a Brier score of 0.002 on a holdout test set. A chart review analysis by expert physicians indicated clinical utility. An ESKD prediction model can identify individuals at risk for ESKD and has been successfully deployed within our health system.

Driving behavior-guided battery health monitoring for electric vehicles using extreme learning machine

An accurate estimation of the state of health (SOH) of batteries is critical to ensuring the safe and reliable operation of electric vehicles (EVs). Feature-based machine learning methods have exhibited enormous potential for rapidly and precisely monitoring battery health status. However, simultaneously using various health indicators (HIs) may weaken estimation performance due to feature redundancy. Furthermore, ignoring real-world driving behaviors can lead to inaccurate estimation results as some features are rarely accessible in practical scenarios. To address these issues, we proposed a feature-based machine learning pipeline for reliable battery health monitoring, enabled by evaluating the acquisition probability of features under real-world driving conditions. We first summarized and analyzed various individual HIs with mechanism-related interpretations, which provide insightful guidance on how these features relate to battery degradation modes. Moreover, all features were carefully evaluated and screened based on estimation accuracy and correlation analysis on three public experimental battery degradation datasets. Finally, the scenario-based feature fusion and acquisition probability-based practicality evaluation method construct a useful tool for feature extraction with consideration of driving behaviors. This work highlights the importance of balancing the performance and practicality of HIs during the development of feature-based battery health monitoring algorithms.

381 Glioma Three-Dimensional Shape Predicts Underlying Genetic Mutations

INTRODUCTION: Phenotype is the detectable expression of genotype. Despite this, little is known about how the three-dimensional structure of brain tumors correlates with their genetic biomarkers. METHODS: The UCSF-PDGM dataset was filtered for histopathologically-proven gliomas. Segmentation masks of the enhancing tumor, non-enhancing/necrotic tumor, and surrounding FLAIR abnormality were obtained from each patient’s preoperative brain MRI. All tumors were evaluated for IDH mutations and 1p/19q codeletion; all grade III and IV tumors were tested for MGMT methylation status. The segmentation masks were processed to create topological features describing the tumor’s 3D shape. These features, without other clinical variables, were used in a custom machine learning pipeline to predict the presence of IDH mutations, 1p/19q codeletion, and MGMT methylation. RESULTS: After filtration, 103 of 494 gliomas tested had an IDH mutation, 29 of 494 had a 1p/19q codeletion, and 247 of 409 had methylation of MGMT. On the blinded test-subset, the machine learning model, using only features describing shape, rendered an AUROC of 0.902 (95% CI: 0.875-0.929), specificity of 90.3% (83.9%-96.6%), and sensitivity of 75.7% (68.0%-83.3%) for IDH mutation. For 1p/19q codeletion, we found an AUROC of 0.949 (0.908-0.990), specificity of 94.5% (84.3%-100%), and sensitivity of 86.6% (79.6%-93.6%). For MGMT methylation, the performance was poor with an AUROC of 0.445 (0.385-0.504). CONCLUSIONS: The three-dimensional shape of a glioma may be used to predict the presence of some key underlying genetic mutations before biopsy. Additional research is needed to validate this, improve its fidelity, and generalize it to other biomarkers.

Whole brain morphologic features improve the predictive accuracy of IDH status and VEGF expression levels in gliomas.

Glioma is a systemic disease that can induce micro and macro alternations of whole brain. Isocitrate dehydrogenase and vascular endothelial growth factor are proven prognostic markers and antiangiogenic therapy targets in glioma. The aim of this study was to determine the ability of whole brain morphologic features and radiomics to predict isocitrate dehydrogenase status and vascular endothelial growth factor expression levels. This study recruited 80 glioma patients with isocitrate dehydrogenase wildtype and high vascular endothelial growth factor expression levels, and 102 patients with isocitrate dehydrogenase mutation and low vascular endothelial growth factor expression levels. Virtual brain grafting, combined with Freesurfer, was used to compute morphologic features including cortical thickness, LGI, and subcortical volume in glioma patient. Radiomics features were extracted from multiregional tumor. Pycaret was used to construct the machine learning pipeline. Among the radiomics models, the whole tumor model achieved the best performance (accuracy 0.80, Area Under the Curve 0.86), while, after incorporating whole brain morphologic features, the model had a superior predictive performance (accuracy 0.82, Area Under the Curve 0.88). The features contributed most in predicting model including the right caudate volume, left middle temporal cortical thickness, first-order statistics, shape, and gray-level cooccurrence matrix. Pycaret, based on morphologic features, combined with radiomics, yielded highest accuracy in predicting isocitrate dehydrogenase mutation and vascular endothelial growth factor levels, indicating that morphologic abnormalities induced by glioma were associated with tumor biology.

Offshore field experimentation for novel hybrid condition monitoring approaches

This study details the development of a fully automated pipeline for the condition monitoring of wind turbine drive trains. Vibration data is collected using hardware designed and manufactured in-house and used directly to monitor the condition of the drive trains. The complex nature of wind turbine vibration signals, due to the large number of components and highly variable operating conditions, makes drive train condition monitoring a challenging task. This paper details the full data measurement and analysis flow from sensor to insights and proposes a hybrid automated pipeline with signal processing and data-driven techniques to address the complexity of dealing with wind turbine vibration data. The vibration signals are directly employed to estimate the wind turbine’s instantaneous angular speed to compensate for any rotation speed fluctuations. Pre-processing is performed on the speed-independent signals to evaluate condition indicators in both the time and spectral domain for the vibration signals and their envelopes. Machine learning is then employed to distinguish the healthy state of the machine from a faulty one using the computed condition indicators. Besides the scalar indicators, also two-dimensional vibration decompositions such as the cyclic spectral correlation maps are used as inputs to the machine learning pipeline. This comprehensive and automated approach ensures both an early and reliable fault detection. Experimental results demonstrate that the fully automated hybrid pipeline can effectively be used for fleet-based health tracking of offshore wind turbine drivetrains.

Crop Recommendation Predictive Analysis using Ensembling Techniques

Crop recommendation systems play a crucial role in modern agriculture by aiding farmers in making well-informed choices to optimize crop yield and resource utilization. Ensemble learning approaches can significantly improve the effectiveness of crop recommendation systems. To achieve this, multiple forecasts are combined from various models. In this paper, a complete Machine Learning Pipeline is used to evaluate the performance of ensemble learning models in crop recommendation tasks. A diverse dataset is used to select and train four ensemble learning methods, Bagging, Voting, Stacking, and One-Vs-Rest (OVR), as separate classifiers. The dataset includes various agricultural factors such as soil characteristics, meteorological conditions, and past crop productivity. Various metrics, including accuracy, precision, recall, F1-score, and support, are utilized for each model. Bagging is considered the most effective ensemble learning technique, demonstrating excellent levels of accuracy and overall performance. The bagging algorithm achieves a high level of accuracy, reaching 99.32%. It also achieves perfect precision, recall, and F1-score metrics, with values of 0.99, 1.00, and 1.00 respectively. The support value, which represents the number of instances used for evaluation, is 141. This study provides valuable perspectives on the choice of appropriate ensemble learning models for crop recommendation tasks. Consequently, it enables farmers and other individuals involved in agriculture to make well-informed choices using data, resulting in enhanced agricultural output and sustainability.

Big Data and Java are integrated with machine learning

This in-depth investigation explores the revolutionary nexus of Java, Big Data, and Machine Learning (ML), clarifying the innovations and synergies that result from their integration. The voyage commences with a summary of the past, following the growth from Java's fundamental function in high-level development to the revolutionary influence of Big Data frameworks such as Apache Hadoop and Apache Spark. The story then delves into the fundamentals of machine learning in Java, highlighting its adaptability, rich library support, and crucial role in building strong ML pipelines. The investigation delves into the transformative power of Big Data frameworks, highlighting the distributed file system of Hadoop and the in-memory processing capabilities of Spark. We observe the significant effects of this convergence on a variety of industries through real-world case studies, from e-commerce personalized suggestions to fraud detection in banking. The insights gained from these implementations highlight how crucial it is to use ML models with ethical considerations, interdisciplinary cooperation, and ongoing learning. The following sections cover the nuances of data preprocessing, including the use of Java in ETL workflows, scalable feature engineering using Big Data frameworks, and data quality assurance via transformation and cleansing. ML model deployment is the major focus, along with an exploration of the Java runtime environment, micro services architecture, and crucial aspects of model robustness monitoring and maintenance. The investigation concludes with a focus on case studies and success stories that demonstrate the real-world effects of this convergence in sectors like e-commerce, finance, and healthcare. These real-world examples highlight the accomplishments of companies like Netflix, Uber, Airbnb, and others and provide insightful information about how well integration works to accomplish a range of business objectives.

Dynamic Digital Radiography Pulmonary Function Testing: A Machine Learning Lung Study Alternative

BackgroundCommon diagnostic tests for pulmonary disorders include chest radiography and pulmonary function tests (PFTs). Although essential, these tests only offer a static assessment. Chest dynamic digital radiography (DDR) integrates lung and diaphragm motion in one study with limited radiation exposure. DDR is relatively easy to obtain, but barriers to its clinical adoption include time consuming manual analysis and unclear correlation with PFTs. Research questionsCan a machine learning pipeline automate DDR analysis? What is the strength of the relationship between PFT-measures and automated DDR-based lung area measurements? Study design and MethodsPFT and DDR studies were obtained in fifty-five participants. We developed an analysis pipeline utilizing convolutional neural networks (CNN) capable of quantifying lung areas in DDR images to generate DDR-based PFTs (dPFTs). PFT and dPFT measures were correlated in patients with normal, obstructive and restrictive lung physiology. ResultsWe observed statistically significant (p values < 1x10-6), strong correlations between dPFT areas and PFT volumes including TLC (r=0.764), FEV1 (r=0.591), VC (r=0.763) and FRC (r=0.756). Automated DDR and lung shape tracking revealed differences between normal, restrictive and obstructive physiology using diaphragm curvature indices and strain-analysis measurements. Linear regressions allowed for derivation of PFT values from dPFT measurements. InterpretationStatistically significant correlations found between dPFTs and PFTs suggest that dPFTs can act as a surrogate to PFTs when these are not available or unable to be performed. This study contributes to the potential integration of DDR as a reliable alternative to PFTs.

Using machine learning to predict outcomes of patients with blunt traumatic aortic injuries.

The optimal management of blunt thoracic aortic injury (BTAI) remains controversial, with experienced centers offering therapy ranging from medical management to TEVAR. We investigated the utility of a machine learning (ML) algorithm to develop a prognostic model of risk factors on mortality in patients with BTAI. The Aortic Trauma Foundation registry was utilized to examine demographics, injury characteristics, management and outcomes of patients with BTAI. A STREAMLINE (A Simple, Transparent, End-To-End Automated Machine Learning Pipeline Facilitating Data Analysis and Algorithm Comparison) model as well as logistic regression (LR) analysis with imputation using chained equations was developed and compared. From a total of 1018 patients in the registry, 702 patients were included in the final analysis. Of the 258 (37%) patients who were medically managed, 44 (17%) died during admission, 14 (5.4%) of which were aortic related deaths. Four hundred forty-four (63%) patients underwent TEVAR and 343 of which underwent TEVAR within 24 hours of admission. Among TEVAR patients, 39 (8.8%) patients died and 7 (1.6%) had aortic related deaths ( Table 1 ). Comparison of the STREAMLINE and LR model showed no significant difference in ROC curves and high AUCs of 0.869 (95% confidence interval, 0.813-0.925) and 0.840 (95% confidence interval, 0.779-0.900) respectively in predicting in-hospital mortality. Unexpectedly, however, the variables prioritized in each model differed between models. The top 3 variables identified from the LR model were similar to that from existing literature. The STREAMLINE model, however, prioritized location of the injury along the lesser curve, age and aortic injury grade. Machine learning provides insight on prioritization of variables not typically identified in standard multivariable logistic regression. Further investigation and validation in other aortic injury cohorts are needed to delineate the utility of ML models. Prognostic and Epidemiological; Level III.

Terahertz Time-Domain Spectroscopy of Blood Serum for Differentiation of Glioblastoma and Traumatic Brain Injury

The possibility of the differentiation of glioblastoma from traumatic brain injury through blood serum analysis by terahertz time-domain spectroscopy and machine learning was studied using a small animal model. Samples of a culture medium and a U87 human glioblastoma cell suspension in the culture medium were injected into the subcortical brain structures of groups of mice referred to as the culture medium injection groups and glioblastoma groups, accordingly. Blood serum samples were collected in the first, second, and third weeks after the injection, and their terahertz transmission spectra were measured. The injection caused acute inflammation in the brain during the first week, so the culture medium injection group in the first week of the experiment corresponded to a traumatic brain injury state. In the third week of the experiment, acute inflammation practically disappeared in the culture medium injection groups. At the same time, the glioblastoma group subjected to a U87 human glioblastoma cell injection had the largest tumor size. The THz spectra were analyzed using two dimensionality reduction algorithms (principal component analysis and t-distributed Stochastic Neighbor Embedding) and three classification algorithms (Support Vector Machine, Random Forest, and Extreme Gradient Boosting Machine). Constructed prediction data models were verified using 10-fold cross-validation, the receiver operational characteristic curve, and a corresponding area under the curve analysis. The proposed machine learning pipeline allowed for distinguishing the traumatic brain injury group from the glioblastoma group with 95% sensitivity, 100% specificity, and 97% accuracy with the Extreme Gradient Boosting Machine. The most informative features for these groups’ differentiation were 0.37, 0.40, 0.55, 0.60, 0.70, and 0.90 THz. Thus, an analysis of mouse blood serum using terahertz time-domain spectroscopy and machine learning makes it possible to differentiate glioblastoma from traumatic brain injury.

Correlative Assessment of Machine Learning-Based Cobb Angle Measurements and Human-Based Measurements in Adolescent Idiopathic and Congenital Scoliosis.

Purpose: Scoliosis is a complex spine deformity with direct functional and cosmetic impacts on the individual. The reference standard for assessing scoliosis severity is the Cobb angle which is measured on radiographs by human specialists, carrying interobserver variability and inaccuracy of measurements. These limitations may result in lack of timely referral for management at a time the scoliotic deformity progression can be saved from surgery. We aimed to create a machine learning (ML) model for automatic calculation of Cobb angles on 3-foot standing spine radiographs of children and adolescents with clinical suspicion of scoliosis across 2 clinical scenarios (idiopathic, group 1 and congenital scoliosis, group 2). Methods: We retrospectively measured Cobb angles of 130 patients who had a 3-foot spine radiograph for scoliosis within a 10-year period for either idiopathic or congenital anomaly scoliosis. Cobb angles were measured both manually by radiologists and by an ML pipeline (segmentation-based approach-Augmented U-Net model with non-square kernels). Results: Our Augmented U-Net architecture achieved a Symmetric Mean Absolute Percentage Error (SMAPE) of 11.82% amongst a combined idiopathic and congenital scoliosis cohort. When stratifying for idiopathic and congenital scoliosis individually a SMAPE of 13.02% and 11.90% were achieved, respectively. Conclusion: The ML model used in this study is promising at providing automated Cobb angle measurement in both idiopathic scoliosis and congenital scoliosis. Nevertheless, larger studies are needed in the future to confirm the results of this study prior to translation of this ML algorithm into clinical practice.

Discrimination of Etiologically Different Cholestasis by Modeling Proteomics Datasets.

Cholestasis is characterized by disrupted bile flow from the liver to the small intestine. Although etiologically different cholestasis displays similar symptoms, diverse factors can contribute to the progression of the disease and determine the appropriate therapeutic option. Therefore, stratifying cholestatic patients is essential for the development of tailor-made treatment strategies. Here, we have analyzed the liver proteome from cholestatic patients of different etiology. In total, 7161 proteins were identified and quantified, of which 263 were differentially expressed between control and cholestasis groups. These differential proteins point to deregulated cellular processes that explain part of the molecular framework of cholestasis progression. However, the clustering of different cholestasis types was limited. Therefore, a machine learning pipeline was designed to identify a panel of 20 differential proteins that segregate different cholestasis groups with high accuracy and sensitivity. In summary, proteomics combined with machine learning algorithms provides valuable insights into the molecular mechanisms of cholestasis progression and a panel of proteins to discriminate across different types of cholestasis. This strategy may prove useful in developing precision medicine approaches for patient care.

Fixed Point Label Attribution for Real-Time Bidding

Problem definition: Most of the display advertising inventory is sold through real-time auctions. The participants of these auctions are typically bidders (Google, Criteo, RTB House, and Trade Desk for instance) that participate on behalf of advertisers. In order to estimate the value of each display opportunity, they usually train advanced machine learning algorithms using historical data. In the labeled training set, the inputs are vectors of features representing each display opportunity, and the labels are the generated rewards. In practice, the rewards are given by the advertiser and are tied to whether a particular user converts. Consequently, the rewards are aggregated at the user level and never observed at the display level. A fundamental task that has, to the best of our knowledge, been overlooked is to account for this mismatch and split, or attribute, the rewards at the right granularity level before training a learning algorithm. We call this the label attribution problem. Methodology/results: In this paper, we develop an approach to the label attribution problem, which is both theoretically justified and practical. In particular, we develop a fixed point algorithm that allows for large-scale implementation and showcase our solution using a large-scale publicly available data set from Criteo, a large demand-side platform. We dub our approach the fixed point label attribution algorithm. Managerial implications: There is often a hidden leap of faith when transforming the advertiser’s signal into display labeling. Demand Side Platforms providers should be careful when building their machine learning pipeline and carefully solve the label attribution step.

Reinforcement Learning as a Parsimonious Alternative to Prediction Cascades: A Case Study on Image Segmentation

Deep learning architectures have achieved state-of-the-art (SOTA) performance on computer vision tasks such as object detection and image segmentation. This may be attributed to the use of over-parameterized, monolithic deep learning architectures executed on large datasets. Although such large architectures lead to increased accuracy, this is usually accompanied by a larger increase in computation and memory requirements during inference. While this is a non-issue in traditional machine learning (ML) pipelines, the recent confluence of machine learning and fields like the Internet of Things (IoT) has rendered such large architectures infeasible for execution in low-resource settings. For some datasets, large monolithic pipelines may be overkill for simpler inputs. To address this problem, previous efforts have proposed decision cascades where inputs are passed through models of increasing complexity until desired performance is achieved. However, we argue that cascaded prediction leads to sub-optimal throughput and increased computational cost due to wasteful intermediate computations. To address this, we propose PaSeR (Parsimonious Segmentation with Reinforcement Learning) a non-cascading, cost-aware learning pipeline as an efficient alternative to cascaded decision architectures. Through experimental evaluation on both real-world and standard datasets, we demonstrate that PaSeR achieves better accuracy while minimizing computational cost relative to cascaded models. Further, we introduce a new metric IoU/GigaFlop to evaluate the balance between cost and performance. On the real-world task of battery material phase segmentation, PaSeR yields a minimum performance improvement of 174% on the IoU/GigaFlop metric with respect to baselines. We also demonstrate PaSeR's adaptability to complementary models trained on a noisy MNIST dataset, where it achieved a minimum performance improvement on IoU/GigaFlop of 13.4% over SOTA models. Code and data are available at github.com/scailab/paser.

A Machine Learning Pipeline for Cervical Cancer Prediction

Cervical cancer ranks as the fourth leading cause of cancer-related deaths among women and often remains asymptomatic during its initial stages. Therefore, it’s very important to predict it. The study's objective is to explore cervical cancer risk variables and develop an ensemble prediction model to forecast cervical cancer more precisely. The cervical cancer data were represented by 11 risk factors and a dependent variable consisted of Hinselmann, Schiller, Cytology, and Biopsy. SVM, decision tree, and random forest methods are compared for prediction, with models trained on 75% of the data and evaluated using confusion matrices. Precision and recall metrics assess model performance, with recall being vital in critical situations like identifying cervical cancer patients. The comparison was made among these three methods and compared the ranking result of risk factors with the ground truth, random forest method offers high recall (0.8821) and accuracy (0.8744). Recognizing risk factors can enhance cure rates and female patient survival. And the cervical cancer prediction system enhances medical care and reduces the cost. Data mining techniques in the medical sector facilitate diagnostic and prognostic apps for early treatment of life-threatening diseases.

Quilt: Robust Data Segment Selection against Concept Drifts

Continuous machine learning pipelines are common in industrial settings where models are periodically trained on data streams. Unfortunately, concept drifts may occur in data streams where the joint distribution of the data X and label y, P(X, y), changes over time and possibly degrade model accuracy. Existing concept drift adaptation approaches mostly focus on updating the model to the new data possibly using ensemble techniques of previous models and tend to discard the drifted historical data. However, we contend that explicitly utilizing the drifted data together leads to much better model accuracy and propose Quilt, a data-centric framework for identifying and selecting data segments that maximize model accuracy. To address the potential downside of efficiency, Quilt extends existing data subset selection techniques, which can be used to reduce the training data without compromising model accuracy. These techniques cannot be used as is because they only assume virtual drifts where the posterior probabilities P(y|X) are assumed not to change. In contrast, a key challenge in our setup is to also discard undesirable data segments with concept drifts. Quilt thus discards drifted data segments and selects data segment subsets holistically for accurate and efficient model training. The two operations use gradient-based scores, which have little computation overhead. In our experiments, we show that Quilt outperforms state-of-the-art drift adaptation and data selection baselines on synthetic and real datasets.

Should Fairness be a Metric or a Model? A Model-based Framework for Assessing Bias in Machine Learning Pipelines

Fairness measurement is crucial for assessing algorithmic bias in various types of machine learning (ML) models, including ones used for search relevance, recommendation, personalization, talent analytics, and natural language processing. However, the fairness measurement paradigm is currently dominated by fairness metrics that examine disparities in allocation and/or prediction error as univariate key performance indicators (KPIs) for a protected attribute or group. Although important and effective in assessing ML bias in certain contexts such as recidivism, existing metrics don’t work well in many real-world applications of ML characterized by imperfect models applied to an array of instances encompassing a multivariate mixture of protected attributes, that are part of a broader process pipeline. Consequently, the upstream representational harm quantified by existing metrics based on how the model represents protected groups doesn’t necessarily relate to allocational harm in the application of such models in downstream policy/decision contexts. We propose FAIR-Frame, a model-based framework for parsimoniously modeling fairness across multiple protected attributes in regard to the representational and allocational harm associated with the upstream design/development and downstream usage of ML models. We evaluate the efficacy of our proposed framework on two testbeds pertaining to text classification using pretrained language models. The upstream testbeds encompass over fifty thousand documents associated with twenty-eight thousand users, seven protected attributes and five different classification tasks. The downstream testbeds span three policy outcomes and over 5.41 million total observations. Results in comparison with several existing metrics show that the upstream representational harm measures produced by FAIR-Frame and other metrics are significantly different from one another, and that FAIR-Frame’s representational fairness measures have the highest percentage alignment and lowest error with allocational harm observed in downstream applications. Our findings have important implications for various ML contexts, including information retrieval, user modeling, digital platforms, and text classification, where responsible and trustworthy AI is becoming an imperative.

Abstract 4915: Identification and validation of novel prognostic genetic markers in HER2-negative advanced gastric cancer (AGC) by artificial intelligence (AI) deep learning and machine learning algorithm

Abstract Introduction: Despite advances in anticancer treatments, there has been urgent need to identify novel prognostic markers for HER2-negative AGC. We propose AI based deep learning and machine learning pipeline to explore genetic markers that predict patient response to chemotherapy in HER2-negative AGC. Methods: We retrospectively assessed 179 stage IV HER2-negative AGC patients who have been treated with first-line chemotherapy in Yonsei Cancer Center, Korea between 2015 and 2021. In-house targeted sequencing panel data (CancerSCANTM and CancerMaster) were used to identify candidate prognostic markers. DeepSurv was mainly used to analyze the progression-free survival (PFS), which is a deep learning algorithm that can investigate prognostic roles. The SHapley Additive exPlanations (SHAP) method was applied to open the deep learning models’ black-box and calculate the importance ranking of gene features. Machine learning models-random survival forest, elastic-net, and lasso-were also investigated to identify potentially meaningful gene variants. To rank the gene features, we calculated variable importance (VIMP) for random survival forest, and penalized regression coefficients for elastic-net and lasso. All deep learning and machine learning models were trained using 5-folds cross-validation in 1,000 iterations of re-sampling bootstrapping with hyperparameter tuning by grid search. To classify responder or non-responder related genetic markers to chemotherapy, we used the signs of averaged penalize coefficients obtained from elastic-net and lasso. Based on the median linear predictor with candidate gene features, patients were classified into high-risk and low-risk group. Results: A total of 7,824 common variants were analyzed and 20 prognostic genetic markers were identified based on the top average values of SHAP, VIMP, and penalized regression coefficients. The top 10 non-responder genetic markers are included as follows: PARP4(p.S873N), CHUK(p.V268I), BARD1(359_365del), CREBBP(p.Y1125F), BARD1(p.P24S), PHLPP2(p.R1312Q), FAT3(p.Q3375R), CASP5(p.T106A), PHLPP2(p.I544V), and PTCH2(p.T988M). Otherwise, the top 10 responder genetic markers are defined as follows: IL7R(p.I66T), NOTCH4(L16delinsLL), FGFR3, BCL2A1(p.G82D), ZNF217(p.T548I), BRCA1(p.K1183R), FANCA(p.A412V), ADGRA2, CDKN2B, and FANCM(p.S175F). The high-risk group had worse PFS compared with low-risk group (median PFS 4.1 vs. 9.1 months; P-value &amp;lt;0.0001). Conclusion: This study showed the potential of AI deep learning and machine learning pipeline that employs an innovate prognostic genetic markers for HER2-negative AGC. Current ongoing analyses of larger cohort with more comprehensive clinical data will be presented. Citation Format: Sejung Park, Seok-Jae Heo, Choong-Kun Lee, Yaeji Lee, Woo Sun Kwon, Jingmin Che, Minseok Kim, Hyun Cheol Chung, Sun Young Rha. Identification and validation of novel prognostic genetic markers in HER2-negative advanced gastric cancer (AGC) by artificial intelligence (AI) deep learning and machine learning algorithm [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2024; Part 1 (Regular Abstracts); 2024 Apr 5-10; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2024;84(6_Suppl):Abstract nr 4915.

Abstract 6416: Early stage endometrioid endometrial cancer recurrence risk stratification using a machine learning RNA-seq gene expression signature

Abstract Endometrioid endometrial cancer (EEC) adjuvant therapy decisions rely on risk stratification using histology, grade, stage, and lymphovascular space invasion (LVSI). Recently, molecular classification systems originating from TCGA, evaluated in GOG-210 and PORTEC-3 defined four prognostic subtypes based on POLE, MSI-H/MMR-D, and p53 alterations. Although valuable, this molecular approach still has significant limitations such as applicability to the majority of EEC patients categorized as no-specific molecular profile (NSMP) and the potential need to resolve pathogenic and prognostic heterogeneity within MMR-D, and TP53 subtypes. These unmet needs were key motivators for Tempus to develop a molecular classifier predicting distant recurrence risk in early-stage EEC with a focus on high intermediate-risk (HIR) patients. The RNA-seq-based gene expression profiler (GEP) was trained using a machine-learning pipeline with TCGA data which resulted in a 24 gene signature that classifies EEC patients as molecular risk (MR) high or low (MR-high, MR-low). The GEP MR test was then evaluated on a de-identified cohort of EEC patients (N=1037) from Tempus to test associations with known pathologic or molecular prognostic features. The GEP-MR risk predictor showed significant enrichment of MR high-risk in G3 versus G1/2 histology (p-value &amp;lt; 5e-8). A high correlation was found between the MR score and copy number alteration score (t-test p-value &amp;lt; 1e-5). Next, a clinical evaluation was performed in an early-stage EEC case-control cohort of patients with documented recurrence or no recurrence event at four years (N=109), from Stanford. In the entire cohort, the MR-high group had a significantly higher rate of distant recurrence in comparison to the MR-low group (HR=4.8, N=109). Next, we performed a subgroup analysis in the clinically important HIR patients. In this subgroup, the MR-high group had a significantly higher rate of distant recurrence in comparison to the MR-low group (HR=8.0, N=56). Lastly, given the significance of genomic biomarkers in the evolution of EEC FIGO staging, we stratified outcomes by the established TCGA subtypes as a reference standard and performed a subgroup analysis in patients classified as having no specific molecular profile (NSMP). Among patients who were NSMP, the MR-high group showed a significantly higher rate of distant recurrence in comparison to the MR-low group (HR=7.92, N=67). These evaluation studies demonstrate the performance of the GEP MR test in early-stage EEC distant recurrence risk stratification, specifically HIR patients, necessitating further studies to validate the test for clinical use in informing adjuvant clinical management. Citation Format: Rafi Pelossof, Talal Ahmed, Seung Won Hyun, Chithra Sangali, Ben Terdich, Calvin Chao, Michael Toboni, Casey Cosgrove, Elizabeth Kidd, Brook Howitt, Timothy Taxter. Early stage endometrioid endometrial cancer recurrence risk stratification using a machine learning RNA-seq gene expression signature [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2024; Part 1 (Regular Abstracts); 2024 Apr 5-10; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2024;84(6_Suppl):Abstract nr 6416.