Douglas Talbert

Building a national workforce development model for artificial intelligence.

Rapidly understanding the rationale behind a model’s recommendation is vital in time-sensitive clinical situations such as trauma triage. Such environments may benefit from an automated translation of explanations from a tool such as LIME to a more clinical-friendly explanation that reduces the total amount of information presented to the paramedic and uses a more natural format and language. Generating clinician-friendly explanations requires an iterative process that reflects the goals, needs, knowledge, and values of the human decision-makers. In this paper, we apply concepts from Human-Centered eXplanable AI to assess an initial iteration of translating LIME-based explanations into a clinician-friendly language and format, and we successfully identify several high-priority tasks that need to be addressed to improve the explanation generation and evaluation process.

The rapid deployment of Artificial Intelligence (AI) in high-stakes domains necessitates robust approaches to transparency, fairness, and trustworthiness. Current advancements in AI performance are outpacing our understanding and ability to govern these systems. This special issue presents research addressing explainability, fairness, and trust as interconnected socio-technical challenges. Accepted papers demonstrate novel techniques for revealing hidden model dependencies, aligning explanations with domain expertise, rigorously benchmarking model classes for explanation robustness, and refining methods for measuring interpretability. We synthesize these contributions, situate them within current policy and standardization (EU AI Act; ¹ NIST AI RMF; ^{2 , 3} ISO/IEC 23894 and 42001 ^{4 , 5} ), and connect them to emerging evaluation science in XAI (e.g., BEExAI, ⁹ Saliency-Bench, ¹⁰ F-Fidelity ¹¹ ). Finally, we outline a forward-looking agenda emphasizing multi-aspect evaluation, context-sensitive trust, and the development of governance-ready AI systems.

Machine learning (ML) is capable of aiding and improving medical diagnostics for a wide variety of pathologies. When used to process data from smart implantable medical devices, ML can offer timely and automated diagnostic tools to improve patient care. However, diagnostic ML models that influence healthcare decisions must have outputs that can be understood and trusted. Currently, ML-based medical diagnostic research focuses mainly on accuracy, lacking investigation of interpretability, explainability, and trust in the models — fundamental principles of diagnostic ML. To address this gap, this study seeks to improve explainability in ML models trained to diagnose aseptic tibial loosening in smart piezoelectric total knee replacements (TKRs). Specifically, a pathway to explainability is presented by applying the local interpretable model-agnostic explanations (LIME) method to interpret previously trained [ ]-nearest neighbor (KNN), support vector machine (SVM), and discriminant analysis (DA) ML models that classify cement damage from electromechanical impedance (EMI) signatures of piezoelectric-instrumented simulated TKRs. Two simple yet novel feature engineering techniques are proposed to align the models with domain knowledge of piezoelectric impedance-based structural health monitoring (SHM), thus improving explainability. These feature engineering techniques are broadly applicable to data types where adjacent features have inherent relations (i.e., time series, spectra, images, etc.). The original KNN, SVM, and DA models demonstrated explainability scores of 52%, 64.4%, and 42.1%, respectively. The first feature engineering technique improved the KNN and SVM scores to 84.6% and 87.6%, respectively, with DA falling to 29.5%. The second feature engineering technique improved the KNN and SVM scores to 74.5% and 81.1%, respectively, with DA falling to 33.5%. The pathway to explainability and the feature engineering techniques presented in this study yield models with improved explainability that are better aligned with domain knowledge, while either improving (SVM) or maintaining (KNN) their original accuracy.

In real-world applications, computational constraints often require transforming large models into smaller, more efficient versions through model compression. While these techniques aim to reduce size and computational cost without sacrificing performance, their evaluations have traditionally focused on the trade-off between size and accuracy, overlooking the aspect of model faithfulness. This limited view is insufficient for high-stakes domains like healthcare, finance, and criminal justice, where compressed models must remain faithful to the behavior of their original counterparts. This paper presents a novel approach to evaluating faithfulness in compressed models, moving beyond standard metrics. We introduce and demonstrate a set of faithfulness metrics that capture how model behavior changes post-compression. Our contributions include introducing techniques to assess predictive consistency between the original and compressed models using model agreement, and applying chi-squared tests to detect statistically significant changes in predictive patterns across both the overall dataset and demographic subgroups, thereby exposing shifts that aggregate fairness metrics may obscure. We demonstrate our approaches by applying quantization and pruning to artificial neural networks (ANNs) trained on three diverse and socially meaningful datasets. Our findings show that high accuracy does not guarantee faithfulness, and our statistical tests detect subtle yet significant shifts that are missed by standard metrics, such as Accuracy and Equalized Odds. The proposed metrics provide a practical and more direct method for ensuring that efficiency gains through compression do not compromise the fairness or faithfulness essential for trustworthy AI.

Mental health challenges among college students are a growing concern, necessitating innovative tools for early detection and intervention. This study hypothesizes that predictive modeling of mood states from multivariate time-series data collected via mobile sensors can be enhanced by leveraging sequence-aware models over non-sequential alternatives. Specifically, we compare the abilities of Long Short-Term Memory networks to those of Support Vector Machines and traditional Artificial Neural Networks for modeling temporal dependencies and predicting mood state. We further hypothesize that variations in data sampling rates will influence model performance and hypothesize that a combination of high data resolution and a focus on more recent movement will support peak performance. These hypotheses guide our exploration of data preparation techniques and model selection strategies for addressing technical challenges related to Just-In-Time Adaptive Interventions aimed at mental health monitoring. Future research directions include real-time deployment and personalized model refinement to improve predictive capabilities.

Autonomous vehicles require intelligent computer vision (CV) to perform critical navigational perception tasks. To achieve this, sensors such as camera, LiDAR and radar are utilized to provide data to artificial intelligence (AI) systems. Continuous monitoring of these intelligent CV systems is required to achieve a trustworthy AI system in a zero-trust (ZT) environment. This paper introduces a novel two-stage framework that provides a mechanism towards achieving this monitoring for a ZT environment. We made use of Monte Carlo (MC) dropout with One-Class Classification techniques to propose a framework towards trustworthy AI systems for AVs. Through extensive experimentation with varying noise levels and number of MC samples, we demonstrate that our framework achieves promising results in anomaly detection. In particular, our framework explores the trade-off between detection accuracy and computational overhead, where we achieved a high FPS of 46.4 with MC size of 5 while the accuracy is as low as 61.5%. This study would provide us with valuable insights for real-world AV applications.

The ability to incorporate original and adapted data into query-based storage structures to provide dynamic and timely service to sequence recommendation systems is a continuous goal of learning management systems. This can be a challenging goal when data integrity and student privacy are paramount. We are developing a hybrid machine learning-assisted system (CyberTaliesin) for cybersecurity educational support. In this poster, we present the early building blocks of the system involving the use of federated knowledge graphs as a trusted knowledge source capable of learning from “less restricted” models such as large language models. How can integrating these tools yield a flexible system that improves sequence recommendations, facilitates concepts such as adaptive and personalized learning, and achieves improved competency-based educational outcomes?