The answer, the review finds, is both promising and sobering. The researchers identified three core dimensions along which LLM-driven health systems personalize their behavior, eight distinct technical architectures that developers use to adapt these models to individual users, and four categories of implementation challenges that consistently undermine effectiveness. That structure matters because personalization in health is not a single feature you can bolt onto a chatbot. It is a design problem that spans what the system knows about you, how it reasons about that knowledge, and how it changes its outputs over time as your needs evolve.

To understand why this review lands at such a pivotal moment, consider what makes large language models different from the health apps that came before them. Earlier generations of digital health tools relied on rigid rule engines and decision trees: if a user logged three workouts in a week, the app sent a congratulatory notification; if a diabetic patient’s glucose reading crossed a threshold, an alert fired. These systems were transparent but brittle. They could not hold a conversation, could not interpret the messy, ambiguous way people actually describe their lives, and could not adjust their tone or content to a user’s literacy level, culture, or emotional state. LLMs, by contrast, are trained on vast corpora of text and can generate fluent, contextually sensitive responses to essentially any prompt, which makes them natural candidates for conversational health coaching, mental wellness support, medication guidance, and lifestyle intervention.

But fluency is not the same as personalization. A model that answers every question beautifully but identically for a teenager in Nairobi, a retiree in Halifax, and a shift nurse in Mumbai is not adaptive; it is merely articulate. The Dalhousie team’s synthesis of the literature shows that researchers have attacked this problem from multiple technical directions. Among the eight adaptation architectures catalogued in the review are approaches such as fine-tuning, where a base model is retrained on domain-specific health data to sharpen its medical reasoning; retrieval-augmented generation, where the model consults external knowledge sources or a user’s personal records before responding; prompt engineering and in-context learning, where user profiles, preferences, and history are packed into the prompt itself so the model conditions its answers on them; and agent-based designs, where multiple specialized model components collaborate, with one handling medical accuracy, another managing conversational tone, and a third tracking long-term user goals.

Each architecture carries distinct trade-offs that the review helps clarify for the field. Fine-tuning can produce deep domain competence but requires curated datasets, computational resources, and careful validation to avoid degrading a model’s general capabilities or introducing subtle biases. Retrieval-augmented approaches keep personal data outside the model weights, which eases privacy concerns and allows information to be updated in real time, but they depend on the quality and currency of the underlying data stores. Prompt-based personalization is cheap and flexible, yet it consumes context window space and can be fragile when a user’s profile grows complex. Agentic pipelines offer modularity and auditability but add engineering complexity and latency. The review’s contribution is to map these options against the health domains where they have been deployed, giving designers a conceptual framework for choosing the right tool for the right intervention.

The health domains covered in the synthesized literature are strikingly broad. LLM-driven systems have been explored for mental health support, where conversational agents can offer around-the-clock availability and nonjudgmental interaction that some users find easier than talking to a human; for chronic disease management, where models can interpret self-reported symptoms and medication logs; for physical activity and nutrition coaching, where adaptive feedback loops respond to a user’s progress; and for health information access, where models translate clinical jargon into plain language tailored to a reader’s background. Across these domains, the review highlights evidence that LLM-driven applications can be effective, while emphasizing that their success hinges on how well they meet individual needs and preferences rather than on raw model capability alone.

Here is where the review delivers its most uncomfortable finding: inclusivity and contextual sensitivity remain underexplored across the entire field. Most systems studied to date are built and evaluated with narrow populations in mind, often English-speaking, digitally literate, and from high-income settings. A personalization engine that models a user’s preferences but ignores their cultural context, language, disability status, or socioeconomic reality risks producing interventions that work beautifully in a pilot study and fail, or even cause harm, in the real world. The authors frame this as a central gap for the human-computer interaction community, arguing that future systems must be designed to be adaptive, personalized, inclusive, effective, and responsive to evolving individual characteristics and needs, a five-part standard that raises the bar well beyond current practice.

The four categories of implementation challenges identified in the review compound that concern. Technical hurdles include the tendency of LLMs to generate plausible but incorrect health information, a failure mode that is dangerous when the subject is medication dosing or symptom triage. Data challenges revolve around the scarcity of high-quality, representative datasets for training and evaluation, particularly for underrepresented groups. Ethical and privacy challenges loom largest for deployment: health conversations are among the most sensitive data a person can share, and questions about consent, data retention, and secondary use remain unsettled in most jurisdictions. Finally, evaluation challenges persist because there is no consensus on how to measure personalization quality itself; a system can score well on generic benchmarks while failing catastrophically at adapting to any single real user.

The ethical dimension receives particular attention in the review, which situates technical choices inside questions of fairness, autonomy, and global impact. If LLM health assistants become a mainstream interface to care, the populations excluded from their training data could face systematically worse automated advice, deepening existing health disparities rather than closing them. Conversely, the authors point to genuine opportunities for global reach: conversational agents that speak low-resource languages, that operate on inexpensive hardware, and that deliver evidence-informed guidance to communities with limited access to clinicians could extend the frontier of wellness support to billions of people. Whether that promise materializes depends on the design decisions being made in laboratories right now, which is precisely why a structured map of the field arrives at such a useful moment.

The Dalhousie team’s conceptual framework, built from the three personalization dimensions, eight adaptation architectures, and four challenge categories, is offered as a shared vocabulary for the next generation of research. Rather than each lab reinventing its own taxonomy of what personalization means, future studies can locate their contributions within a common structure, compare results across domains, and identify which combinations of architecture and dimension remain untested. The review’s dataset, including all coding records from the 149 articles, is available in supplementary material, and the article is open access, lowering the barrier for researchers in resource-constrained settings to build on it. The work was published on 30 September 2026 in Artificial Intelligence Review, with no conflicts of interest declared by the authors.

What emerges from this synthesis is a field in transition. The first wave of LLM health applications proved that the technology could hold convincing, helpful conversations about well-being. The second wave, which this review both documents and aims to accelerate, must prove something harder: that these systems can know who they are talking to, respect who that person is, and change what they say accordingly, safely, and fairly. The 149 studies analyzed here suggest the ingredients are in place, from mature adaptation architectures to growing awareness of ethical pitfalls. What remains is the disciplined, inclusive engineering work of turning a chatty generalist into a genuinely personal health companion, one carefully validated system at a time.

Subject of Research: Systematic review of large language model-driven personalized and adaptive systems for health and wellness

Article Title: LLM-driven personalized and adaptive systems for health and wellness: a systematic review

Article References: Kimeu, J. M., Chan, G., Orji, R., & Oyebode, O. (2026). LLM-driven personalized and adaptive systems for health and wellness: a systematic review. Artificial Intelligence Review. https://doi.org/10.1007/s10462-026-11703-6

Image Credits: AI Generated

DOI: 10.1007/s10462-026-11703-6

Keywords: large language models, personalization, adaptive systems, digital health, wellness, human-computer interaction, systematic review, health informatics, artificial intelligence, machine learning, e-health, conceptual framework

The framework, developed by researchers at the Technical University of Košice in the Slovak Republic, integrates six distinct concerns that the literature has typically handled separately: heterogeneous industrial data, a discrete-event simulation-based digital twin, MQTT and OPC UA communication protocols, external predictive analytics, operational ESG-oriented evaluation, and planner decision support. Rather than treating each as a standalone project, the authors arrange them into a six-layer, event-driven architecture in which information flows between layers as discrete events, triggering analysis, prediction, and ultimately human-approved decisions on the factory floor.

At the heart of the proposal is the idea that events, not periodic reports, should drive production planning and control. In an event-driven design, every meaningful occurrence on the shop floor, such as a machine fault, a material shortage, or a completed batch, becomes a message that propagates through the architecture. Lightweight publish-subscribe protocols like MQTT can carry high-frequency telemetry from sensors to cloud services, while OPC UA provides the standardized, semantically rich interface that industrial equipment uses to expose its state. Prior performance studies of these protocols, cited in the paper, show they are well suited to exchanging data between industrial plants and cloud servers, which makes them natural candidates for the communication backbone of a digital twin.

The digital twin itself is built on discrete-event simulation, a modeling technique that represents a production system as a chronological sequence of events and state changes. This choice matters because discrete-event simulation can answer the questions production planners actually ask: what happens to throughput if a machine goes down for two hours, how a rush order reshapes the schedule, or where buffers will overflow. By coupling the simulation model to live industrial data streams, the twin becomes a predictive instrument rather than a static mirror, allowing planners to test candidate schedules and interventions in silico before committing resources on the floor.

The second major contribution is what the authors call a four-layer transformation logic, and it is here that the framework makes its most distinctive claim. Digital technologies, the authors argue, do not directly improve environmental, social, or governance outcomes. Instead, they influence ESG-relevant results only through changes in production planning and control processes and in managerial evaluation. A sensor network by itself reduces nothing; it is the rescheduled batch, the avoided machine failure, or the reweighted performance indicator, approved by a human planner, that translates digital capability into measurable sustainability impact. This mediating logic is intended to correct what the authors see as a common weakness in the literature, where digitalization and sustainability are linked loosely without specifying the causal pathway.

To make the framework concrete, the paper illustrates it with a hypothetical manufacturing scenario, a complete event trace showing how messages move through the six layers, and formal definitions of measurable indicators. The indicator definitions are designed to be operational, meaning they specify exactly what is measured, from what data, and how the resulting values feed into ESG-oriented evaluation. This level of specification is what separates a genuine decision-support architecture from a conceptual diagram: planners can, in principle, implement the indicators directly and audit how each recommendation was derived from underlying events.

Human oversight is built into the architecture rather than bolted on. The framework explicitly positions planner decision support as the final layer, so that predictive outputs and ESG evaluations arrive as recommendations that a human decision-maker reviews and approves before any change to production planning takes effect. This design reflects the broader Industry 5.0 movement, which, as the European policy literature cited in the paper emphasizes, seeks a sustainable, human-centric, and resilient industry rather than full autonomy. The authors position their work in this tradition, drawing on recent scholarship that frames Industry 5.0 as a corrective to the technology-first ethos of Industry 4.0.

The intellectual lineage of the framework is broad. It builds on established digital twin reference models and six-layer architectural patterns from the manufacturing literature, on systematic reviews of machine learning applications in production lines and predictive quality, and on a growing body of work connecting intelligent manufacturing to ESG performance, including empirical studies of Chinese manufacturing firms and recent analyses asking whether smarter production is also greener. By synthesizing these strands, the authors aim to provide what they describe as an explicit integration of technical event flows with human-approved, ESG-oriented production planning and control decisions, something they argue no single existing framework delivers.

The authors are candid about the limits of the current study. The framework is derived through a structured conceptual synthesis, and the paper does not report implemented or empirically validated performance results. No datasets were generated or analyzed, and the hypothetical scenario exists to illustrate the architecture, not to prove it. The authors list as future work a set of concrete validation steps: discrete-event simulation experiments, construction of an MQTT and OPC UA testbed, expert assessment, and industrial validation in real manufacturing settings. Until those studies are complete, the framework should be read as a rigorous design proposal and a research agenda rather than a demonstrated solution.

Even so, the timing of the proposal is significant. Manufacturers worldwide face simultaneous pressure to digitize operations and to report credible environmental, social, and governance performance, and regulators and investors increasingly demand that sustainability claims rest on verifiable operational data. A framework that traces a straight line from a sensor event, through a predictive simulation, to a planner-approved scheduling decision and a defined ESG indicator offers exactly the kind of auditable causal chain that both engineers and sustainability officers need. If the planned testbed and industrial validation bear out the design, event-driven digital twins could become the connective tissue that finally unifies the smart factory’s many brains, turning streams of industrial telemetry into decisions that are simultaneously faster, more predictive, and demonstrably more sustainable.

Subject of Research: An event-driven digital twin framework for predictive production planning and ESG-oriented decision support in smart manufacturing

Article Title: A Conceptual Framework for Event-Driven Digital Twin-Based Predictive Production Planning and ESG-Oriented Decision Support in Smart Manufacturing

Article References: Karakai, M., & Bobko, D. (2026). A Conceptual Framework for Event-Driven Digital Twin-Based Predictive Production Planning and ESG-Oriented Decision Support in Smart Manufacturing. Mobile Networks and Applications. https://doi.org/10.1007/s11036-026-02535-3

Image Credits: AI Generated

DOI: 10.1007/s11036-026-02535-3

Keywords: digital twin, smart manufacturing, Industry 5.0, discrete-event simulation, MQTT, OPC UA, predictive production planning, ESG, decision support, industrial IoT, sustainability, conceptual framework