Artificial Intelligence is transforming medicine from episodic, symptom-driven care to continuous, predictive, and preventive medicine, with direct implications for insurance risk selection. This shift from treatment to prevention is challenging how insurers assess risk. Cardiovascular (CV) disease, the leading cause of global mortality, provides a clear example, with AI already identifying risk before symptoms appear, with significant implications for underwriting, pricing, and prevention strategies.
In this Quick Read, Dr Emoke Posan, PartnerRe’s Chief Medical Officer for North America, shares key insights from her recent presentations and research, highlighting how advances in AI‑driven diagnostics, prediction, and prevention are reshaping medicine and raising important considerations for insurance.
AI is transforming medicine from episodic, symptom-driven care to continuous, predictive, and preventive medicine.
AI is transforming routine CV tests into predictive tools. ECGs, CT scans and other imaging diagnostics are no longer limited to detecting existing diseases. Instead, they are becoming predictive engines that estimate future risk based on patterns not visible through conventional interpretation.
The principle is consistent across applications. AI-augmented ECGs can detect or predict heart pump dysfunction from conventionally normal tracings. AI-enhanced coronary CT angiography can prognosticate outcomes based on plaque burden.[1] ECG models estimating biological age show that an AI-derived age gap of eight years or more independently predicts all-cause mortality.[2]
Opportunistic screening further expands these capabilities, allowing AI to extract CV risk signals from tests performed for unrelated indications, such as head CT or abdominal scans. In each case, AI identifies meaningful patterns that traditional analysis cannot detect.
Together, these advances point to a new clinical concept: the AI-detected disease stage, or “previvor” – an individual with subclinical disease detectable by AI but not yet visible through standard diagnostics. In effect, AI can identify risk years before symptoms emerge.
For insurers, this challenges the traditional binary distinction between “healthy” and “diseased.” Risk may increasingly need to be understood as a continuum, raising fundamental questions about how applications are assessed when elevated risk is identified in the absence of clinical disease.
Importantly, these examples all involve predictive (analytical) AI systems, which are narrow, task-specific models validated against clinical ground truth. To date, more than 1000 FDA-cleared AI medical devices (algorithms) exist, predominantly of this type.
AI can identify risk years before symptoms emerge.
The distinction is clear when considering ECG interpretation, as an example. Predictive AI models have been specifically developed and clinically validated to analyze ECG signals, often demonstrating high diagnostic performance and, in certain tasks, performance comparable to or exceeding that of clinicians.[3]
Generative AI models are fundamentally different. When applied to ECG interpretation, their performance remains inconsistent, as outputs depend on prompt and context rather than fixed, validated pathways.[4], [5] This variability makes them unsuitable for diagnostic use.
This reflects a broader divide: predictive AI is task-specific, trained on labelled datasets, and validated against defined benchmarks, while generative AI excels at language-based tasks such as summarizing documents, drafting correspondence, and streamlining workflows, but is not suited to interpreting diagnostic data.
A promising development is the rise of domain-specific language models trained on curated medical datasets rather than broad internet data. Models such as vision-language systems trained on ECG data may outperform general-purpose LLMs in clinical tasks and should be monitored closely, as they may combine the flexibility of generative AI with the reliability of predictive AI.
As AI rapidly evolves, the boundaries between predictive and generative approaches will increasingly blur through multimodal platforms and foundation models (massive systems pre-trained on vast unlabeled datasets through unsupervised and self-supervised learning).
AI-enabled wearables extend data collection from episodic testing to continuous monitoring. They enable real-time, scalable observation of physiological signals, making advanced diagnostics accessible beyond specialized centers and into everyday life.
This continuous, multimodal data stream underpins precision medicine. It enables earlier detection of silent pathology before symptoms emerge and supports a shift in healthcare from reactive treatment to proactive prevention.
Continuous monitoring, detection, and prediction mark a transition from snapshot views of health to dynamic, longitudinal insights. In the future, autonomous systems may continuously monitor data, identify issues, propose interventions, and support AI health coaches that translate findings into actionable recommendations. Together, these developments point toward continuous care and even self-care, reshaping how medicine is delivered.
AI has evolved rapidly from single-purpose models to platforms capable of integrating multiple data types, including ECGs, imaging, genomics, and clinical records into unified risk assessments.
Increasingly powerful systems are trained on vast unlabeled datasets through unsupervised and self-supervised learning. Foundation models represent the most advanced expression of this approach, as they can adapt to multiple clinical tasks without requiring retraining from scratch.
While these advances significantly enhance predictive capability, they also introduce new considerations, including substantial computational intensity and associated economic and environmental impacts.[6]
Risk may increasingly need to be understood as a continuum rather than a binary distinction between ‘healthy’ and ‘diseased’.
The potential of AI in medicine is substantial.[7] It enables more precise diagnostics, earlier detection (and intervention) during reversible disease stages, and personalized risk stratification, alongside a shift toward continuous and preventive care. Over time, this could contribute to meaningful improvements in mortality, extended lifespan, and health span. Most notably, AI-generated health insights may encourage behavioral change, advancing both primordial and primary prevention, potentially the single most important lever for global mortality. At the same time, these developments introduce a set of interrelated challenges for insurers, which may include the following:
Among these developments, one stands out: the potential for large-scale prevention. If AI-driven early detection and continuous monitoring shift behavior toward proactive prevention, the impact on global mortality and longevity could be transformative, exceeding that of individual high-cost intervention.
Ultimately, however, the full potential of AI must be confirmed through real-world outcomes. While early evidence is promising, prospective validation will be critical to determine whether these advances translate into sustained improvements in health outcomes and risk assessment.[8]
PartnerRe closely monitors the evolution of AI and its clinical applications. If you would like to discuss these developments, we invite you to contact us.
Dr. Emoke Posan, Chief Medical Officer North America
[1] https://www.jacc.org/doi/10.1016/j.jcmg.2023.05.020
[2] https://www.jacc.org/doi/10.1016/j.jacep.2024.02.011
[3] Validation of an artificial intelligence model for 12-lead ECG interpretation | European Heart Journal | Oxford Academic
[4] Bocz B, Janosi K F, Ferencz A B et al: Evaluation of AI-based ECG analysis accuracy using Chat GPT-4 and Gemini AI models, EP Europace, 2025. 27: euaf085.032 https://doi.org/10.1093/europace/euaf085.032
[5] Soubh N, Rasenack E, Haarmann H et al. Performance of Vision-Enabled Large Language Models in Image-Based Electrocardiogram Interpretation: Exploratory Evaluation. J Med Internet Res. 2026. 28:e86692. doi: 10.2196/86692.
[6] https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0343056