This position paper argues that current insurance markets are fundamentally misaligned with AI risk, creating significant coverage, control, and measurement gaps that threaten both organizations and insurers. Through analysis of insurance policy coverages, risk controls, and measurement approaches across 15 AI risk categories, we demonstrate that conventional insurance structures are inadequately addressing the unique challenges presented by AI systems. This misalignment stems from AI's autonomous nature, probabilistic operations, opacity, and rapid development cycles, which conflict with insurance assumptions about human control, causality, deterministic failures, and stable risk environments. While some argue that existing policies sufficiently cover AI risks, our evidence shows that even the most relevant cyber and technology liability insurance products leave organizations exposed to significant AI-specific harms. This qualitative analysis establishes a foundational framework for future quantitative studies and proposes measurement approaches that could enable more rigorous empirical analysis as AI incident data matures. Without deliberate evolution in AI risk transfer mechanisms, organizations face a protection gap while insurers confront potentially catastrophic unpriced exposure, creating an urgent need for market-based risk transfer solutions that can drive adoption of technical AI security and safety standards.
AI has shown remarkable potential in healthcare, but faces accessibility challenges due to high computational and expertise demands, especially in medical image analysis. Vector embeddings, compact representations of medical images achieved from foundation models in zero-shot inference, offer a potential solution. Recently, an equivalent vector embeddings dataset of existing large publicly available medical images has been released, for which training an AI model requires significantly lower computing infrastructure and storage needs. Such data sets provide greater accessibility to AI in medical imaging for those who do not have access to large computing resources. The burning question remains: What is the gain or loss in using vector embedding to replace medical images, particularly from a fairness and utility point of view? In this work, we compare AI models trained in vector embeddings (Emb) with raw chest radiograph images for disease diagnosis, focusing on both performance and fairness. Our results show that Emb-based models match or exceed image-based models in diagnostic performance while improving fairness. Crucially, Emb achieve this with far less computational cost. These findings position Emb as a powerful, scalable alternative to image-based AI, especially valuable for low-resource settings where access to GPUs and expert infrastructure is limited.
Time series (TS) analysis is an active application area for Artificial Intelligence (AI) methods where the objective is to analyze numeric quantities indexed by time for tasks like classification, forecasting, and abnormality detection. In health, TS manifests as biosignals like the electroencephalogram (EEG), where electrical signals from the brain are analyzed. AI and health communities can tremendously benefit each other in TS, with the former offering advanced analytical methods while the latter provides complex data sets and trust-sensitive use cases. But the communities also need to overcome confusing terminologies, hidden assumptions, and a lack of necessary domain contexts for result evaluation and interpretation. In this paper, we attempt to bridge the gap using the problem of channel selection in EEG. We outline challenges in working with EEG data, demonstrate via two experiments how simple explainable AI (XAI) methods can be quite effective for channel selection irrespective of EEG tasks/paradigms, and argue that recent TS trends in AI, like LLMs and XAI methods, can benefit health as well. We hope that this work will bring researchers working on TS problems at the intersection of AI and health closer to work in AI trustworthiness so that they can better leverage results from their respective areas to overcome common challenges. All code and resources are released on GitHub to help others replicate.
The generation of structured medical reports using large language models (LLMs) presents unique challenges, particularly in maintaining clinical relevance and adhering to strict formatting requirements. In this work, we investigate the effectiveness of fine-tuning LLMs for structured report generation using DeepSeek R1 models. We conduct experiments with two model variants: DeepSeek R1 8B and DeepSeek R1 14B. For both models, we apply Group Relative Policy Optimization (GRPO) using the Medical Information Mart for Intensive Care (MIMIC-IV) dataset, leveraging Low-Rank Adaptation (LoRA) for parameter-efficient fine-tuning. Our results show that the GRPO fine-tuned DeepSeek-R1 8B and 14B models outperformed all baseline models, including the larger 32B DeepSeek-R1 model, demonstrating the effectiveness of parameter-efficient tuning. These findings underscore the potential of reinforcement learning-based fine-tuning of LLMs for generating structured reports in the medical domain.
Deep learning models have demonstrated impressive accuracy in predicting acute kidney injury (AKI), a condition affecting up to 20% of ICU patients, yet their black-box nature prevents clinical adoption in high-stakes critical care settings. While existing interpretability methods like SHAP, LIME, and attention mechanisms can identify important features, they fail to capture the temporal dynamics essential for clinical decision-making, and are unable to communicate when specific risk factors become critical in a patient's trajectory. This limitation is particularly problematic in the ICU, where the timing of interventions can significantly impact patient outcomes. We present a novel interpretable framework that brings temporal awareness to deep learning predictions for AKI. Our approach introduces three key innovations: (1) a latent convolutional concept bottleneck that learns clinically meaningful patterns from ICU time-series without requiring manual concept annotation, leveraging Conv1D layers to capture localized temporal patterns like sudden physiological changes; (2) Temporal Concept Tracing (TCT), a gradient-based method that identifies not only which risk factors matter but precisely when they become critical addressing the fundamental question of temporal relevance missing from current XAI techniques; and (3) integration with MedAlpaca to generate structured, time-aware clinical explanations that translate model insights into actionable bedside guidance. We evaluate our framework on MIMIC-IV data, demonstrating that our approach performs better than existing explainability frameworks, Occlusion and LIME, in terms of the comprehensiveness score, sufficiency score, and processing time. The proposed method also better captures risk factors inflection points for patients timelines compared to conventional concept bottleneck methods, including dense layer and attention mechanism. This work represents the first comprehensive solution for interpretable temporal deep learning in critical care that addresses both the what and when of clinical risk factors. By making AKI predictions transparent and temporally contextualized, our framework bridges the gap between model accuracy and clinical utility, offering a path toward trustworthy AI deployment in time-sensitive healthcare settings.
