Narrowing of the carotid arteries by plaque deposition is a leading cause of stroke and heart disease in humans. The mechanical forces exerted by blood on the artery wall, particularly wall shear stress (WSS) and pressure, play important roles in both plaque growth and artery wall rupture. This project uses computational fluid dynamics (CFD) to simulate pulsatile blood flow in simplified carotid-like tubes with three levels of stenosis (75%, 50%, and 25% reduction in cross-sectional area) to investigate how narrowing severity affects outlet pressure fluctuations and WSS distributions. Preliminary results from two cases (75% and 50% stenosis) show that increasing stenosis depth enhances both the pulsatile pressure drop and the extremes of wall shear stress at the throat, consistent with clinical observations linking severe stenoses to cardiovascular risk.
Chest radiography is the primary imaging modality for pulmonary disease screening, yet the limited interpretability of deep learning models remains a major barrier to clinical adoption. This paper presents a unified framework for evaluating and improving ex-plainability in CNN-based chest X-ray analysis. We first conduct a systematic benchmark of five convolutional neural network ar-chitectures (DenseNet121, ResNet50, EfficientNetV2, MobileNetV3, and ConvNeXtV2) combined with four widely used explainabil-ity methods (GradCAM, GradCAM++, Integrated Gradients, and SmoothGrad), evaluated with five complementary metrics span-ning localization accuracy, attribution faithfulness, and robustness on the NIH ChestX-ray14 dataset. Evaluated on all 880 bounding-box-annotated images, the benchmark shows that explanation quality depends more on the backbone than on the attribution method: ConvNeXtV2 leads on localization while DenseNet121 of-fers the highest explanation consistency at far lower parameter cost. Building on this, we propose AttrNet, an explainability harness that is backbone-agnostic by construction: it wraps a classifica-tion network with a lightweight attribution branch that predicts class-specific explanation maps natively in the forward pass and couples them back to the classifier through a cross-attention gate and differentiable localization supervision, turning explainability from a post-hoc by-product into a trained objective. Instantiated on DenseNet121 and evaluated on held-out lesions under cross-validation, AttrNet raises heatmap–lesion IoU from 0.200 to 0.345 and Pointing Game accuracy from 0.333 to 0.596 (both 𝑝 < 10−18, paired) at a cost of 0.011 classification AUC, an exchange favourable wherever spatial reliability governs clinical utility. We further in-troduce C-GradCAM, a contrastive extension of GradCAM that suppresses activations shared with competing disease classes to produce more faithful and disease-discriminative explanations, im-proving attribution quality without requiring network retraining or architectural modification. Collectively, the benchmark, Attr-Net, and C-GradCAM demonstrate complementary strategies for advancing transparent, trustworthy, and clinically meaningful ex-plainable AI for pulmonary disease diagnosis. CCS Concepts • Computing methodologies → Neural networks; Computer vision; • Applied computing → Health informatics.
The use of artificial intelligence to adjudicate health insurance claims offers opportunities to increase processing speed, consistency, and scalability. But, accuracy of AI-driven claims adjudication does not depend solely on the performance of the AI system. It depends on the accuracy of the health plan configuration against which the claims are processed. In this paper, we examine a gap in the correctness of the health plan configuration in which the AI system applies the wrong benefit and ultimately returns a false decision. We study the process through which the plan information is translated into structured configuration and discuss common configuration errors associated with regard to cost sharing, network, coverage and exclusions, authorization, plan assignment, and effective dates. We distinguish between decision correctness (or correctness) in that the AI system applied the appropriate rule and benefit correctness (ie.g., whether the rule accurately represents the health plan). We propose a governance framework based on the validation of the benefits, a model, configuration testing, effective date controls, human escalation, monitoring, and auditability. Our analysis shows that the accuracy of AI models is not enough for a reliable health insurance adjudication. Accurate benefit configuration is an important component of AI governance, claims accuracy, and responsible automation.
AI is now a major component in the health insurance claims adjudication, utilization, and coverage decision making. But, this decision should be made accurately by the AI not only by following the rules, but also by knowing whether the benefit configuration is consistent with the health plan’s document. An error in the configuration of benefits can lead to incorrect coverage or payment even when the AI applies the rules correctly. The purpose of this study is to develop a reproducible model for identifying such errors. The proposed framework proposes a domain-specific taxonomy to describe accumulator and out-of-pocket limits, network and tier mappings, cost sharing, exclusions and limitations, effective dates, and conditional benefit rules. Based on public health insurance benefit documents and structured data sources, the proposed benchmark proposes a reference configuration and a method to introduce controlled errors. The proposed evaluation protocol evaluates the performance of AI on four dimensions: error detection, error classification, error localization, and false assurance. It also specifies sample construction, model selection, prompting, and scoring in order to assure reproducibility. This paper does not report unverified experimental results, but rather offers a structured test to distinguish between configuration fidelity failures and AI rule execution failures. It argues that this need to validate benefit configurations is independent from automated judgments of trustworthiness. It provides an important foundation for future testing, auditability, and governance of AI-aided decisions in health insurance.
The shipping industry faces increasing pressure to reduce greenhouse gas emissions while maintaining operational efficiency. This study evaluates the energy-saving potential of an Organic Rankine Cycle (ORC) system for waste heat recovery from a 21,840 kW MAN 8S65ME-C8.6 main marine diesel engine during a realistic 1,610.7-nautical-mile coastal voyage from Dalian to Qinzhou, China. R245fa was utilized as the working fluid, with the main engine operating at various loading conditions (50% and 90%) and seawater temperatures (10-30°C). A steady-state thermodynamic model using MATLAB® with CoolProp™ library was developed and validated against published benchmarks. The analysis integrates exhaust gas and jacket cooling water as dual heat sources, post-boiler steam, and various sulfur fuel qualities under IMO MARPOL regulations. Results indicate that LSFO significantly outperforms HSFO for waste heat recovery, delivering 753.1 kW net power output (68% higher than HSFO) at 90% load, with thermal efficiencies of 11.31% and 10.86% respectively. The ORC system produces 121 MWh per voyage, replacing 8% of auxiliary power demands. CO₂ reduction reaches 62.1 tons per voyage for LSFO in winter conditions, aligning with IMO's 40% reduction target by 2030. Seawater temperature variation from 10°C to 30°C reduces ORC power output by approximately 7%, highlighting the importance of seasonal considerations in system design. This research provides a validated, open-source computational framework for marine waste heat recovery assessment and demonstrates that ORC technology offers a practical pathway for improving marine engine efficiency while reducing emissions.
