Healthcare concerns the regulation of biological states, the reduction of uncertainty about disease processes, and the continual selection of diagnostic and therapeutic actions under incomplete evidence — a set of problems that maps directly onto active inference's core constructs of generative models, prediction error, expected free energy, and policy selection. Over the past decade a literature has grown around computational psychiatry, interoception and pain, digital twins, medical robotics, and — most recently — the safety of large language models used as clinical decision-support tools. Most of this work remains theoretical or simulation-based, but reviews in computational psychiatry and interoceptive psychopathology have begun to consolidate the field, and conceptual roadmaps for "Active Inference AI" in medicine sketch how the framework could reach digital twins, precision diagnostics, and clinical decision support.
Why the domain fits
Clinical practice is the continual inference of hidden causes — disease processes — from noisy, partial, and delayed observations (symptoms, labs, imaging), followed by action selection (tests, treatments) that must balance learning about the patient's condition against improving outcomes. Active inference formalizes this as minimizing expected free energy, which decomposes into epistemic value (resolving uncertainty) and pragmatic value (achieving preferred outcomes), giving a single variational account of diagnosis and treatment choice. The framework's emphasis on maintaining physiological variables within viable bounds also maps onto homeostasis and allostasis: interoceptive signals are treated as observations, regulatory set points as priors, and disorders such as anxiety, chronic pain, or eating disorders as consequences of inaccurate generative models or maladaptive priors.
State of the literature
The most mature work sits in psychology and psychiatry: Smith et al. survey active inference as a candidate unifying paradigm across learning, attention, and psychopathology, while Kirchner et al. argue it can link symptom profiles to generative-model parameters (priors, precision) to guide personalized treatment selection, for example modeling depression via overly pessimistic priors and reduced precision on positive prediction errors. Interoception and pain form a second cluster: Smith et al.'s account of interoceptive psychopathology and Petzschner et al.'s review of interoception and allostasis treat anxiety, panic, and functional somatic disorders as failures of interoceptive inference, and a specific active inference theory of placebo analgesia explains how the brain either updates priors from conflicting nociceptive signals or discounts them as noise, producing divergent placebo trajectories. Palacios et al. extend the framework to morphogenesis, proposing that cellular and tissue-level processes minimize free energy in ways that could eventually reframe cancer, fibrosis, and regenerative medicine, though this remains speculative. A decision-making study (novelty/variability, published via eLife) offers the field's clearest empirical head-to-head, showing active inference models outperforming standard reinforcement learning at explaining both behavior and neural correlates of exploration.
Key projects and tools
The report finds no widely adopted, healthcare-specific active inference software; researchers currently adapt general-purpose active inference libraries and custom probabilistic-programming code to clinical constructs, mostly for simulation and parameter fitting rather than deployed systems. Three named application patterns stand out as the closest things to concrete projects: a ScienceDirect architecture for "active digital twins" that use active inference to let patient or organ models actively seek information rather than passively simulate; a Cancer Journal conceptual report on "Active Inference AI and the Spatial Web for Medicine" describing virtual tumor boards where digital twins are explored under different treatment policies; and, in robotics, the Royal Society Interface case study and Da Costa et al.'s review showing active inference driving real robot motor control, with clear extensions sketched for prosthetics and rehabilitation. A 2025 npj Digital Medicine paper proposes treating clinical LLM prompts themselves as active-inference policies chosen to minimize expected free energy over model outputs, aimed at LLM-based medical devices.
Open problems
The central gap is validation: active inference models of depression, interoception, or placebo response have rarely been tested for predictive power against clinical outcomes or compared rigorously to simpler models, and the field needs standardized pipelines for model specification, cross-validation, and prospective trials. Building the hierarchical, multi-scale, multi-modal generative models healthcare requires — linking molecular, organ, and behavioral data — remains technically hard, with open questions about state-space design, missing data, and inference algorithms at scale. Ethical and regulatory questions are unresolved as well: how preferences (survival, symptom relief, quality of life) should be encoded as priors, how to prevent generative models from inheriting biased priors from historical data, and how to bound exploratory (epistemic) actions so they don't introduce clinical risk. The report treats digital twins, Active Inference AI, and LLM prompting strategies as conceptual and largely undeployed, calling for pilot decision-support roles with human oversight before any move toward more autonomous clinical use.
Reference Backbone
Karl J. Friston (2010). The free-energy principle: a unified brain theory? Nature Reviews Neuroscience. DOI: 10.1038/nrn2787. Christopher L. Buckley, Chang Sub Kim, Simon McGregor, Anil K. Seth (2017). The free energy principle for action and perception: A mathematical review. Journal of Mathematical Psychology. DOI: 10.1016/j.jmp.2017.09.004. Thomas Parr, Giovanni Pezzulo, Karl J. Friston (2022). Active Inference: The Free Energy Principle in Mind, Brain, and Behavior. MIT Press. Giovanni Pezzulo, Francesco Rigoli, Karl J. Friston (2017). Active Inference, homeostatic regulation and adaptive behavioural control. Progress in Neurobiology. DOI: 10.1016/j.pneurobio.2017.08.001.